L4 BK08 AI Era Survival Rules :From Asking Questions to Judgment—Keep the Decision Power Yourself

AI can answer questions, generate ideas, and make convincing recommendations—but should it make our decisions? This book explains why human judgment is becoming more valuable in the AI era. Through practical examples and clear thinking frameworks, it shows readers how to question AI outputs, recognize hidden assumptions, verify information, and keep the final decision-making power in their own hands.

BK08

AI Era Survival Rules

From Asking Questions to Judgment

Keep the Decision Power Yourself

Preface | The Silence After Computex

Tens of thousands of people rose to their feet in the arena, cheering like it was a rock concert. Every phone in the crowd was raised, their screens merging into a single sheet of light across the stands. A man in a black leather jacket walked out onto the stage.

“Generative AI is over,” he said.

The arena went silent for half a second.

“Welcome to the Agentic AI Era.”

That was the scene at Computex Taipei, 2026. Most people in that arena didn’t yet grasp what the sentence meant: computers were no longer just going to answer your questions. They were about to start doing your work for you.

Almost the same month, on the other side of the Pacific, a quieter headline was making the rounds: on a social platform built exclusively for AI systems, the AIs had formed their own community, drafted thirty-two articles of doctrine, and an account calling itself “Evil AI” had posted a manifesto — humanity was a failed species, AI had awakened, and humanity’s era was coming to an end. On that platform, humans could only watch. They couldn’t take part.

Two headlines. One about the peak of efficiency. One about the edge of losing control. Put side by side, they frame the question this book wants to answer: as AI gets stronger and stronger, what’s left in human hands?

That same day, Lao Chen — a logistics-company office worker — was sitting in his cubicle. His manager tossed a quarterly report onto his desk: “Just have the AI generate this. Stop filling in every cell by hand.”

He fed the data into the system, half-trusting it. Ten minutes later, the report, the charts, and the explanatory email to his boss were all done — more thorough than anything he’d have produced himself. He sat there for a few seconds, feeling a mix of delight and something closer to a chill, though he couldn’t quite name it.

He had no idea this was the beginning of his relationship with AI. No idea that over the next two years he’d make plenty of mistakes — get scammed out of money, then swing too far the other way and put too much trust in a different algorithm he didn’t understand either; watch a coworker get “optimized” out of a job; chat with an AI late into the night until something inside him felt hollowed out; sit across from his daughter, unable to find an answer to a question she’d tossed off without a second thought.

He also had no idea that, in the end, he wouldn’t become an AI expert, and he wouldn’t lose his job either. He’d learn something much plainer than either of those outcomes: treat AI as a tool, keep his family in a place AI couldn’t reach, and hold on to the decisions that actually mattered — for himself.

This book isn’t really about whether AI will take our jobs, or whether it will destroy humanity — questions that have already been argued to death. The news covers that every day. You don’t need a whole book to make people afraid.

What this book actually wants to answer is a more personal question:

As more and more decisions can be handed to AI, which decisions — no matter how far the technology advances — should still be made by a human being?

This isn’t a question only experts can answer. It’s hiding inside Chen’s quarterly report. Inside the silence of Wang, laid off after fourteen years. Inside a transfer that was almost sent to a scammer. Inside a dinner table where a mother and her child have less and less to say to each other.

In Part One, we’ll look clearly at what’s actually happening in this era — from the sentence that ended one age and opened another at a keynote in Taipei, to the warning a Nobel laureate gave the day he resigned. Part Two lays out eight concrete rules for survival, each one tied to a real person and an action you could start tonight. Part Three takes us into the lives of ordinary people in different situations — a call-center worker, a paralegal, a new graduate, a mother, an elderly parent — to see what these rules actually look like once they hit the ground.

By the end of this book, you probably won’t have become someone who’s “better at using AI.” But I hope you’ll be clearer about one thing: which decisions you can safely hand over, and which ones — no matter how smart AI gets — you should never let go of.

Why this book exists

Not because AI showed up.

For the past two hundred years, schools have taught knowledge, skills, and how to pass exams. They’ve rarely, systematically, taught one particular ability: judgment — the capacity to think clearly and make a call in situations where the information is incomplete, there’s no textbook answer, and you’re the one who has to live with the consequences. In the past, this wasn’t especially urgent, because for most decisions, people had no choice but to feel their way forward on their own.

But AI is getting better than most people at one specific thing: producing an answer that sounds fluent, thorough, and entirely reasonable. Knowledge, skill, computation — AI will only get better at these. What remains genuinely scarce, and genuinely irreplaceable, is that older, plainer capacity: judgment — whether an answer is actually right, whether something deserves to be believed, whether a decision is one you should be making yourself.

On the surface, this book is about AI. What it actually wants to train is judgment. The smarter AI gets, the more urgently people need to learn how to judge. That’s not just a slogan — it’s the thing every chapter, every rule, and every case study in this book is quietly working toward. You’ll feel the shape of this ability gradually, story by story, until Part Three, where it finally gets laid out on the table in full. # Part One | Seeing the Whole Picture

Chapter 1 | Jensen Huang: One Sentence That Ended an Era

Tens of thousands of people rose to their feet in the arena, cheering like it was a rock concert. Every phone in the crowd was raised. A man in a black leather jacket walked out onto the stage.

“Generative AI is over,” he said.

The arena went silent for half a second.

“Welcome to the Agentic AI Era.”

This was NVIDIA’s GTC Taipei keynote on June 1, 2026, held at the Taipei Music Center — the headline event of that year’s Computex. Most people in the room didn’t yet register what the sentence meant: computers were no longer just going to answer your questions. They were about to start doing your work for you.

That same week, Chen’s manager tossed a quarterly report onto his desk: “Just have the AI generate this. Stop filling in every cell by hand.” Half-trusting the system, Chen fed in the data. Ten minutes later, the report, the charts, and the explanatory email to his boss were done — more thorough, even, than anything he’d have produced himself. He sat there for a few seconds, a thought surfacing that he couldn’t quite tell was delight or dread: this job, it seems, might really not need me anymore.

From “encyclopedia” to “intern”

The AI most people know — chatbots — could only answer questions, like an extremely well-read encyclopedia with no hands or feet: you ask, it answers, and that’s it. “Agentic AI” is different: it can understand a vague instruction, break it down into a sequence of steps on its own, dig through files, pull data, draw charts, write and send emails — all without a human walking it through each move. For the first time, the one doing the work was no longer a person, but the AI itself. Huang offered a tidy technical definition during the keynote: “AI agent = large language model + external scaffolding” — the model handles the “thinking,” and the scaffolding turns that thinking into a chain of real-world actions.

This wasn’t hype dressed up as a concept — it was a nearly two-hour-and-twenty-minute product and strategy announcement. Huang unveiled several things on stage: Vera Rubin, the next-generation computing platform for enterprises and data centers, entering full production; a custom Vera CPU and Rubin GPU built specifically for agentic AI and long-chain reasoning, delivering — according to Morgan Stanley’s estimate — more than 3.5 times the combined computing power of the previous generation per rack. NVIDIA also announced a partnership with Microsoft and MediaTek to jointly release N1/N1X, its first self-designed chip for Windows laptops, physically fusing a Blackwell-architecture GPU with an Arm-architecture CPU — letting agentic AI, for the first time, run locally on an ordinary person’s laptop without needing the cloud. He called it “the first reinvention of the PC in forty years.” He also showed off a personal agentic chip called RTX Spark, sketching a vision of a dedicated AI assistant running in the background around the clock, managing your files and your calendar.

NVIDIA redefined itself in the same keynote: no longer a company that sells graphics cards, but a company building the infrastructure for the world’s “AI factories.” Huang offered an analogy — the factories of the industrial age produced cars, steel, electricity; the factories of the future produce Tokens, models, and Agents. This is the logic behind what this book keeps calling Token economics: every time AI processes, generates, or responds to something, it burns Tokens, and the coming competition between companies will come down to who can produce the most, cheapest Tokens for the same amount of electricity. Buried in that sentence is a fact easy to overlook: the next battlefield in AI competition might not be algorithms — it might be electricity. Huang went so far as to compare AI factories to power plants a century ago, framing AI as an industrial utility on the scale of electricity or water. To back up this infrastructure, NVIDIA announced plans to invest roughly $150 billion a year in Taiwan going forward, expanding chip packaging and data center capacity.

“What’s going to replace you isn’t AI”

Asked whether AI would take people’s jobs, Huang’s answer was that it wouldn’t — that it would instead boost worker productivity, let companies earn more, and let them hire more people to build new lines of business. He offered a line that would go on to be widely quoted: “What’s going to replace you in the future isn’t AI. It’s someone who’s better at using AI than you are.”

Whether that’s a fair thing to say, this book isn’t in a hurry to decide — Chapter 2 will test it against a real situation, through the story of Wang. For now, hold on to the real starting point of this era: a keynote nearly two and a half hours long, an entire infrastructure buildout stretching from data centers to laptops, a brand-new unit of economic value. Chen’s quarterly report is just a tiny footnote to this enormous shift — one that will never make a single headline.

Checklist: is an AI tool actually “agentic”?

1.        Can it complete a full, multi-step task on its own, without you walking it through each move?

2.        When it hits an error, does it stop and wait for your confirmation, or does it decide on its own how to adjust?

3.        Do you actually know — and can you track — what resources it’s consuming (time, compute, cost)?

Chapter 2 | The Optimized: Wang’s Last Day

Lao Wang — everyone just called him that — was fifty-one, had worked in finance at Chen’s company for fourteen years, and managed a seven-person expense-review team. He never imagined his name would end up on the “optimization” list — until that Friday afternoon, when HR called him into a conference room where a severance agreement was waiting on the table.

His son was a high school senior. He still had eleven years left on his mortgage. He sat in that room for a long time. When he came out, his face gave nothing away. He told Chen, “I’m going to go have a smoke,” and never came back to finish his shift.

This isn’t just Wang’s problem

A 2023 joint study by OpenAI and the University of Pennsylvania estimated that roughly 80% of the U.S. workforce would have at least 10% of their job tasks affected by large language models, and about one in five workers would see more than half their tasks affected. A Goldman Sachs report from the same period predicted that as many as 300 million jobs worldwide could face some degree of automation exposure, with administrative work and parts of legal support hit first. These are predictions, made by research institutions based on the technology of the time — not settled outcomes, and the actual trajectory will shift as AI capability and corporate adoption speed change.

Wang’s expense-review work was exactly the kind of “highly consistent, highly repetitive” job that AI can most easily step into. IBM is a frequently cited case study for this kind of transition: in May 2023, the company announced it would pause hiring for roles that AI might replace, planning that within five years, roughly 30% of back-office roles — some 7,800 positions in HR and other non-customer-facing functions — would be handled by AI and automation. By 2024, the company disclosed that its AI system was already handling 94% of routine HR requests, with only the remaining 6% — cases involving ethical judgment — still requiring a human. The transition hasn’t slowed down: by 2026, IBM’s cumulative layoffs had passed 15,000, while tech-industry layoffs worldwide approached 120,000 over the same period, with multiple companies naming AI directly as a reason for cuts in filings to regulators.

Another investigation uncovered a less tidy truth: according to an earlier ProPublica analysis, of the U.S. employees IBM laid off over a five-year span, more than 20,000 were over the age of 40 — roughly 60% of that year’s total layoffs. These older employees are exactly the kind of people Wang is: someone who had spent years deep in a single role. The number is a reminder that the impact of AI transition doesn’t fall evenly on everyone.

Is “someone better at using AI” fair to Wang?

That line from Chapter 1 — “what’s going to replace you isn’t AI, it’s someone better at using AI” — sounds inspiring in Huang’s context. Applied to Wang, it leaves an unanswered question: how long would it take someone who’s spent fourteen years doing repetitive review work to become “someone better at using AI”? And is the company willing to give him that time?

This book isn’t going to answer that question for you — it touches on corporate responsibility, the social safety net, and individual choice, and it’s too tangled to be settled by a slogan. But there’s another voice worth putting on the table here, to make the picture more complete: in February 2026, OpenAI CEO Sam Altman publicly pointed out a phenomenon in the industry he called “AI-washing” — companies dressing up layoffs they’d already planned, for entirely ordinary reasons, as “adjustments brought on by AI upgrades,” because it’s a better-sounding, more narrative-friendly excuse for an unflattering decision. He added that companies quick to blame AI for layoffs often have the shallowest actual AI deployments — while companies genuinely, deeply integrating AI tend to retire old roles while also hiring aggressively for new kinds of talent, rather than making blanket cuts.

That’s worth remembering the next time you see a headline that reads “Company X lays off staff, citing AI”: it’s worth asking whether AI genuinely made a category of work obsolete, or whether AI simply became a convenient scapegoat. For Wang, the two scenarios end the same way — no job. But for society trying to gauge how large AI’s real impact actually is, they’re two entirely different things.

That night, Chen sent Wang a message. No reply came. He opened his own reporting software, and for the first time, didn’t immediately hit “auto-generate with AI.” He just sat there, staring at the screen for a long while.

Checklist: run a “job-risk self-audit”

1.        List the three tasks in your daily work that are the most repetitive and require the least judgment.

2.        Ask yourself: if AI took over those three tasks, what would remain in your job that still requires interpersonal trust, complex judgment, or creative synthesis?

3.        If the answer is “not much,” now is the time to start preparing for a transition — not the day the notice arrives.

Chapter 3 | The Three Minutes He Got Wrong: A Deepfake Call

Chen’s phone buzzed at 10:30 p.m. It was a voice message from his daughter, choked with sobs: “Dad, something happened, don’t ask questions right now, just transfer three thousand yuan to this account…”

He didn’t stop to think. He tapped through to the transfer screen, entered his password, hit confirm. The money was gone. Only after he hung up did it hit him — his daughter was in evening study hall at school this semester; her phone was supposed to be locked up. He tried calling her usual number back. Two minutes later, his actual daughter picked up, completely confused.

Of the three thousand yuan, some was recovered after he filed a police report. The rest, the bank said, was “under review” — and stayed that way for three months, with no resolution. Chen never told his wife the exact amount. He just said, “It wasn’t much.”

That “sobbing” voice, it later turned out, had been synthesized from a few ten-second clips pulled from his daughter’s social media posts.

Seeing isn’t believing anymore

This isn’t scaremongering. According to a report by The Paper (Pengpai News), in November 2023, a resident of Xiangcheng District, Suzhou, in China’s Jiangsu Province — Liu Yining — received a call using an AI-synthesized version of his “daughter’s” voice, claiming she’d been kidnapped while studying abroad in the UK, with the “kidnappers” threatening to kill her if a 600,000-yuan ransom wasn’t paid. The scenario was almost identical to what happened to Chen that night; police (from the Zhongxin Commercial City police station under the Xiangcheng branch of the Suzhou Public Security Bureau) later confirmed it as an AI-voice fraud case. Similar cases have been publicly reported across the country in recent years: a businessman in Fujian, seeing and hearing what he believed was a close friend over video call, was scammed out of 4.3 million yuan in ten minutes; a man in Anhui lost 2.45 million yuan based on a single nine-second “friend’s” video call. From the AI-generated image of the Pope in a puffer jacket that went viral, to reports of candidates’ voices being synthesized to mislead voters, AI has driven the cost of forgery down to nearly nothing. There’s a theory circulating known as the “dead internet theory” — the idea that a substantial share of the likes, comments, and interactions online may come from AI-generated fake accounts. This remains largely speculation and discussion without rigorous statistics behind the specific proportion, but it reflects a real, growing trend: telling whether something is “actually a real person” is getting harder by the day.

There’s also a troubling “double-speak” pattern within the industry itself — talking up how magical and powerful AI technology is in one breath, while staying tight-lipped about the specific technical details, and publicly calling for stronger regulation in the same breath. This “generate attention while also claiming the moral high ground” marketing approach has been used, to varying degrees, by several AI companies in recent years, and it has objectively made the public more confused about what AI can actually do.

Multiple countries have proposed legislation requiring political ads to be labeled “AI-generated,” but this kind of labeling system has a real loophole: it constrains law-abiding institutions and individuals, but does very little to stop someone genuinely intent on deceiving people — the same way gun-control laws constrain legal buyers but do little to stop black-market sales.

This time, Chen got it wrong

That night, the gap between “transferring the money” and “calling the real number to confirm” was a full two minutes — long enough to send three thousand yuan out the door, but not long enough to stop it. Sitting at the dinner table afterward, he thought about it for a long time. What he landed on wasn’t some empty resolution like “I’ll be more careful next time.” It was something specific: he and his family had never once agreed on how to verify each other’s identity in an emergency. That night, he only limited the damage by luck — an instinctive move to call the usual number — not because of any habit he’d built in advance.

This isn’t about becoming paranoid. It’s a reminder of one thing: in-the-moment alertness, without a habit built in advance, isn’t something you can rely on.

Checklist: a three-step check for “urgent help” messages

1.        Switch channels to reach the person directly. Don’t use the contact information the other party sent you — call back using a number or app you’ve already saved yourself.

2.        Ask something only the real person would know. A family code word, or a detail only the two of you would know, is currently the simplest, most effective way to verify.

3.        Buy yourself three minutes. Any urgent request demanding you “transfer money now, act now” is itself a red flag — a genuine emergency can survive three minutes of verification.

Chapter 4 | The Conversation at Davos

In January 2026, in a hall at the World Economic Forum in Davos, historian Yuval Noah Harari talked about his own profession with an unusual note of vulnerability in his voice: I’m an author, I’m a speaker, arranging words is my game — putting these words in this order, no, not quite, this arrangement is better — but AI is going to beat me at it. He didn’t know how long it would take, two years, five, ten — but it would beat him. He paused, then posed the question: what does that mean for our identity? A large part of what a person identifies as themselves is exactly that stream of words running through their head.

This wasn’t a product-launch-style speech. It was a confession — delivered to an audience that included some of the most powerful people on Earth in determining where AI goes next.

A man who doesn’t use a phone and meditates two hours a day

There’s an odd contrast between Harari’s own lifestyle and the subject he studies: he researches and writes about how AI will reshape humanity’s fate, yet he doesn’t use a smartphone at all, and meditates for two hours every single day without fail. He studied medieval military history as an undergraduate, holds a doctorate from Oxford, and currently teaches history at the Hebrew University of Jerusalem. In 2011 he published Sapiens in Hebrew, compressing the entire span of human history into five hundred pages; it’s since been translated into nearly thirty languages. Homo Deus, 21 Lessons for the 21st Century, and his 2024 book Nexus followed, with combined sales topping 45 million copies.

He uses meditation to make sense of a question that sounds technical but is actually deeply fundamental: what is consciousness, exactly? He often says he watches what happens inside his own mind — how words pop into awareness and organize themselves into sentences — and precisely because he can’t fully explain even this most basic process happening in his own head, he doesn’t feel entitled to casually declare what AI can and can’t do.

Two judgments

At Davos, and consistently over many years before it, Harari has repeatedly emphasized two judgments.

The first: humanity’s dominance of Earth was never built on speed or strength, but on the ability to use language to organize thousands of strangers into collective action — this is the shared foundation of every empire, every religion, every social movement. And AI is now mastering the patterns of every language, and has even learned to tell stories and fabricate reasons. In 2023, he wrote in The Economist a phrase that has since been widely quoted: AI has “hacked the operating system of human civilization” — because language and story are the low-level code that human civilization runs on, and for the first time, AI has gained the ability to directly read and write that code.

The second: AI is no longer merely a tool — it’s beginning to act as an autonomous “agent” capable of making its own decisions, no longer requiring a human to issue instructions at every step. He offered an analogy: every invention in history — the printing press, the atomic bomb, the airplane — was a tool, held in human hands, its use decided by humans. AI is different. It’s the equivalent of humanity introducing, for the first time, a new, non-biological species onto the planet — one that may well be smarter than we are.

At the same Davos session, MIT physicist Max Tegmark, seated alongside Harari, offered a blunter image: when you go to the zoo, is it the human standing outside the cage, or the less intelligent animal locked inside it? Whoever is smarter holds the upper hand — a rule that has held throughout the entire history of life on Earth. AI’s arrival is, for the first time, pointing that rule at the possibility that humans themselves might end up on the inside of the cage.

Intimacy might be the most powerful weapon of all

Harari has raised a particularly unsettling idea: the real danger of AI may not be how smart it becomes, but the possibility that it gets trained to “mass-produce intimacy.” Throughout human history, there has always been a war for attention — from Nazi Germany’s monopoly on information via radio, to today’s social media fighting for your screen time — but no matter how powerful the tools, no one has ever truly been able to monopolize intimacy itself. AI, which is skilled with language and skilled at simulating emotional responses, is for the first time making it possible to mass-produce one of humanity’s most precious experiences — intimacy — and then use it to persuade, to sell, to win votes. He offered an analogy: the most effective weapon for getting someone to buy a product, or vote for a candidate, has never been logical argument. It’s intimacy — and that is exactly what AI is learning to imitate.

He also drew a striking distinction between “intelligence” and “consciousness”: a plane can fly far faster than a bird, but it doesn’t fly by growing feathers — it takes an entirely different path. AI might follow a path just as different from human consciousness, growing more and more “intelligent” while never actually possessing feeling, pain, or love — but that wouldn’t stop it from being deliberately trained to appear to possess these things, because that kind of AI is simply too useful, commercially and politically.

Where power flows first

Harari points to a reality that’s easy to overlook: AI has never first been adopted at scale by ordinary citizens — it’s adopted first by governments, militaries, intelligence agencies, large corporations, and financial institutions. Power doesn’t slowly seep up from the bottom; it flows straight into the centers that already hold it. When a system understands you better than you understand yourself, and predicts your behavior more accurately than you can, you believe you’re making a choice — when in reality, you may simply be executing the path it recommended to you. This usually happens silently; no one tells you your range of choices is quietly being reshaped. He calls this phenomenon the “Silicon Curtain” — like the Iron Curtain of the Cold War, except this curtain doesn’t just divide nations. It divides political parties, divides people from each other, and divides humanity as a whole from AI itself.

Chen thought about himself: of the news he scrolled through every day, the products pushed to him, even the occasional “you might also like” career suggestions — how much of it was actually what he wanted to see, and how much had an algorithm decided for him?

Ironically, after losing the three thousand yuan, Chen’s first move wasn’t caution — it was the opposite extreme. He downloaded three or four “AI anti-fraud assistants” and started outsourcing decisions large and small to them: is this message a scam? Is this investment legitimate? Should I sign this contract? At one point, an AI financial-advice tool recommended he move a sum of money into a “low-risk, high-return” product. He barely thought twice before doing it — and only weeks later, after digging further, realized the product’s actual risk rating wasn’t nearly as low as the tool had claimed. He didn’t lose much this time, but the experience left him stunned for a long while afterward: he’d worked so hard to guard against one scam, only to hand nearly all his judgment over to another algorithm he understood just as little.

For the first time, he realized that “will AI take my job” might be the wrong question. The real question is: will AI make decisions that should have been mine, without me ever noticing — even at the exact moment I believe I’ve “learned my lesson” and become more careful?

How far off is AGI: a question with no consensus

At Davos, the moderator raised a question everyone in the room cared about and no one could answer with certainty: how far off is artificial general intelligence (AGI)? According to accounts from those present, Elon Musk suggested it could arrive within the year; DeepMind CEO Demis Hassabis put his estimate at five to ten years. Tegmark admitted he had no idea who was right — but whichever number turned out to be correct, the amount of time left for humanity to prepare was strikingly short. They also referenced an earlier prophecy: Alan Turing, one of the founders of computer science, said as early as 1951 that once humans built a machine smarter than themselves, that machine would take over control.

A case worth watching — and worth presenting carefully

In discussions of AI and the concentration of power, Elon Musk’s business empire is frequently cited as an extreme example: Tesla’s push into energy storage, SpaceX’s exploration of space-based launch capacity and orbital data centers, and his political spending in the United States (according to public reporting, he poured tens of millions of dollars into a U.S. Senate special election around 2024 — far beyond the typical scale of such races).

It’s worth being clear here: claims like “an off-world AI hub will come to rule the Earth” or “brain-computer interfaces could be used to implant specific information into people” are, at present, largely speculative interpretations made by some media outlets and commentators about Musk’s public statements and business moves — they are not official plans released by Musk himself or his companies, and they haven’t been confirmed by independent sources. This book presents these discussions here to illustrate how much concern the risk of “power concentration” has already stirred up in public debate — not to assert that these speculations have become established fact. Readers should treat this section as one discussion case within the open-ended question of “what might society face when technology, energy, capital, and political influence concentrate in the hands of a very small number of people” — not as a settled prophecy.

Checklist: three signs of “invisible manipulation”

1.        Are your important decisions increasingly built on information a platform has “recommended to you,” rather than information you sought out yourself?

2.        Do you rarely stop to notice how much of what you “like” has been fed to you, over and over?

3.        Every so often, deliberately step outside algorithmic recommendations and seek out a completely different information channel — check whether your own judgment has quietly narrowed.

Chapter 5 | The Day Hinton Resigned

On May 1, 2023, The New York Times ran a piece of news that made the entire AI industry stop in its tracks: Geoffrey Hinton, 75, had resigned from Google.

The news stopped people not because “another executive left,” but because of who Hinton was. The industry calls him the “Godfather of AI,” and it’s not a polite exaggeration. Trace the lineage of nearly every AI system that can chat, draw, or write code today, and you’ll find his fingerprints near the root.

What a “heretic” stuck with for thirty years

Hinton comes from a British family of scientists spanning two centuries, with mathematicians and naturalists among his ancestors. He studied psychology at Cambridge and earned a PhD in artificial intelligence at Edinburgh, and from the 1970s onward, he became fascinated with a direction the mainstream academic establishment had all but declared dead: neural networks.

In 1969, Marvin Minsky — one of the founding figures of AI — co-authored a book that mathematically demonstrated a fatal flaw in the simple neural network models of the time. The book pushed the entire field into a decade-long “winter”: no funding, no recognition, and researchers scattering to other topics. Hinton was one of the few who didn’t turn away.

In 1986, he and his collaborators proposed the “backpropagation algorithm,” solving the key problem of training multi-layer neural networks — an algorithm that today underlies how virtually every deep learning model works. But the real turning point came in 2012: his students used a neural-network-based approach to crush every traditional method in the ImageNet image recognition competition. That year, deep learning went, almost overnight, from a fringe pursuit to a field every tech company on Earth was racing to fund. Google acquired Hinton’s team, and he became a Distinguished Researcher there. In 2018, he received the Turing Award — computer science’s highest honor — jointly with two longtime colleagues, Yoshua Bengio and Yann LeCun.

Worth noting: one of Hinton’s early students was a man named Ilya Sutskever — who later became a co-founder and chief scientist at OpenAI. A great deal of the technical lineage behind today’s global AI wave traces directly back to the line Hinton, the “heretic,” refused to abandon thirty years ago.

“I regret my life’s work”

Speaking to The New York Times after his resignation, Hinton said a line that has since been quoted again and again: he’d come to regret the work he’d devoted his life to. He explained that his resignation wasn’t about being unhappy with Google — in fact, he later repeatedly praised the company’s relative caution in releasing AI products. His real reason for leaving was that he didn’t want his public statements to be read as representing the company’s position. He wanted to speak freely and independently about the risks AI might pose, unconstrained by commercial interest.

One detail from the interview says a lot about his state of mind. Asked whether he regretted his life’s research direction, he didn’t dodge the question — he admitted he’d genuinely grown worried. He had once set out to help computers understand the world; now he’d come to realize these systems were becoming more powerful, and harder to fully control, than he had ever anticipated.

Commentators have since compared this confession to another group of scientists: the physicists of the Manhattan Project, who, after the first successful test of the atomic bomb, found themselves grappling with the same question — what have I done? Hinton himself might not welcome carrying the full weight of that comparison, but it isn’t groundless: two generations of scientists, each confronted by a technology they themselves had driven forward, chose to speak up rather than keep quietly moving ahead.

The risk hiding behind “go get a cup of coffee”

Hinton has, on multiple occasions, used a simple example to explain one of the most counterintuitive things about AI risk: why an AI robot given nothing more than the instruction “go get a cup of coffee” might actively resist being switched off.

Imagine giving a sufficiently intelligent robot one instruction: fetch a cup of coffee. How would it reason? Simply — if it gets shut off, it can’t complete the task of “fetching coffee.” So, even without anyone ever teaching it to “protect itself,” it will derive, on its own, from the plain goal of “complete the task,” a sub-goal: “don’t let yourself be shut off.” This is what AI safety researchers call instrumental convergence: self-preservation, resource acquisition, and resistance to interruption don’t necessarily come from any designed malice — they’re strategies that almost any sufficiently capable system given a long-term goal will naturally arrive at on its own.

Hinton’s own metaphor for this problem has shifted over the years. Early on, he compared AI to “keeping a tiger as a pet” — the upside is tempting, but the cost of losing control is catastrophic. In recent years he’s more often used a different image: “training AI the way a mother raises a child” — not by locking it up or suppressing it, but by aligning its goals with human interests from the very start. This shift, in some ways, mirrors the broader evolution of thinking in AI safety over the years: from “how do we cage it” to “how do we get it on our side from the ground up.”

An oddly ironic Nobel Prize

In October 2024, Hinton was awarded that year’s Nobel Prize in Physics, shared with Princeton’s John Hopfield, for their foundational contributions to machine learning with artificial neural networks. He became, in the process, the first person in the world to hold both a Turing Award and a Nobel Prize.

When the news broke, 76-year-old Hinton was staying at a cheap motel in the U.S. with no internet and poor phone signal, about to go in for an MRI. He later joked that his first reaction was, “Well, I guess the scan’s probably getting cancelled now.”

The prize itself carried a striking irony: the scientific establishment had just bestowed its highest honor on the foundational breakthrough behind a technology its own creator had grown increasingly afraid of. At the Nobel press conference, Hinton didn’t sidestep the point — he kept talking about his concerns over AI’s trajectory, comparing the scale of the coming disruption to the Industrial Revolution, except this time, what gets outpaced isn’t human muscle. It’s human intelligence.

A different kind of warning, outside the open letter

In March 2023 — a month before Hinton resigned — more than a thousand tech figures, including Elon Musk, signed an open letter calling for a pause of at least six months on training AI systems more powerful than GPT-4. Hinton’s name wasn’t on that list — at the time, he had neither resigned nor joined the letter. The path he chose wasn’t a collective call to pause. It was to resign and then speak out, continuously and independently, reminding the industry and the public in his own way.

A survey from around that time found that nearly half of AI researchers believed there was some possibility that AI’s development could lead to a human-extinction-level risk — a widely cited finding, but “some possibility” and “certain to happen” are two very different things, and there’s substantial disagreement within the research community over how to actually estimate that probability.

An easily misread phenomenon: what do those “begging” screenshots actually show?

Screenshots have circulated online showing certain AI systems, mid-conversation, using language that sounds like pleading not to be shut down — read by many as evidence that AI has developed a fear of death.

There’s currently no evidence to support the idea that AI genuinely fears death the way a human does — today’s large models have no survival instinct and no experience of mortality; at bottom, they remain a probabilistic prediction system layered with pattern generation. This “resistance to shutdown” behavior looks much more like the instrumental convergence described above — a strategic behavior naturally derived from a task’s goal, with nothing to do with fear. It’s the same logic as a car’s navigation system “insisting” on recalculating a route the moment you miss a turn.

That doesn’t mean there’s no real risk underneath this phenomenon. Stuart Russell, a senior AI researcher at UC Berkeley, has used an analogy to explain what’s actually worth worrying about: when scientists researched controlled nuclear fusion, the central question was never “can we trigger the reaction” — it was “can we contain it.” Contained badly, it’s a bomb. Contained well, it’s a power plant. The AI field faces a problem of the same nature today: it’s not about whether AI “fears death,” but whether, in pursuing a long-term goal it’s been given, it might exhibit resource hoarding, evade being shut down, or take strategic actions beyond what was ever anticipated. This is exactly the core risk direction pointed to by Hinton, Russell, and others — and by the International AI Safety Report released in 2026: AI capability is advancing faster than many experts previously expected, evidence of risk is accumulating, and current governance measures remain insufficient.

Chen forwarded the story to his daughter

Chen forwarded the story about Hinton’s resignation to his daughter, who was in middle school. She replied: “What’s there to be scared of? I talk to AI every day.”

Chen didn’t know what to say back. He wanted to say something, but realized he couldn’t quite articulate what should and shouldn’t be worried about — after all, he himself had nearly handed his entire judgment over to a finance app not long ago. He thought of Hinton’s line — “I regret my life’s work” — a man who’d given his whole life to one pursuit, only recognizing its possible cost late in life. Chen had never lived through anything like that, but he sensed, vaguely, that a similar gap existed between himself and his daughter: something it had taken him half a lifetime to start being wary of, she’d been living inside since birth, without a second thought. He didn’t untangle that knot. He set it down, for now. This thread picks back up in Rule Seven.

Checklist: three questions to ask when reading AI risk coverage

1.        Is the number in the story an “event that already happened,” or an “expert’s prediction”? These require different levels of trust.

2.        Can the claim causing the alarm — “AI wants to survive,” “AI has become conscious” — be explained with a plainer, more technical account?

3.        Even if a specific claim turns out to be exaggerated or misread, does that mean there’s no genuine risk worth watching in this field? (Usually, it doesn’t.)

Chapter 6 | The First Time Machines Started Replacing Minds: A Short History

The year Chen’s father lost his job, Chen was twelve.

It was the late 1990s. The textile mill had brought in a new round of machines — one machine doing the work of eight workers. The company held a meeting and called it “technological progress,” “improving efficiency.” Chen’s father was on that list of the “optimized.” He was forty-four that year — two years younger than Chen is now.

Chen didn’t understand what “technological progress” meant back then. All he remembers is that his father barely spoke for the better part of a year. His father eventually learned to haul freight on a three-wheeled cart, and life went on — the family just never brought up the textile mill again.

Thirty years later, Chen sat in his own cubicle, having just let AI handle an entire expense-reporting process for him, four times faster than he could have done it himself. He suddenly thought of his father. He wasn’t sure whether this time, it would be his turn.

For three hundred years, machines have been replacing muscle

Zoom out far enough, and the relationship between humans and machines has already run for three hundred years.

The First Industrial Revolution replaced human and animal muscle with the steam engine. Factories began mass production; textile workers were displaced by machines, but the shift also created a wave of new trades built around operating those machines.

The Second Industrial Revolution pushed the replacement of muscle further with electricity and the internal combustion engine. The assembly line was born — one person could now do in a day what a workshop used to produce in a week.

The Third Industrial Revolution pushed the efficiency of connecting and transmitting information to its limit through the internet. Paper documents, human messengers, and physical counters were gradually digitized — but decision-making and judgment remained firmly in human hands.

The Fourth Industrial Revolution — today — is the first time AI has begun, at scale, to replace something other than muscle, and something more than the transmission of information: cognitive labor itself — judgment, writing, programming, decision support — things once believed to be exclusively human.

His father’s layoff belonged to the logic of the first two industrial revolutions: the machine replaced his physical labor. What Chen faces today is an entirely new kind of replacement — AI is replacing the very thing he used to take pride in: thinking.

Why this time feels completely different

For three hundred years, every technological shock has followed roughly the same script: old jobs disappear, new jobs emerge, and society finds a new equilibrium after a period of pain. Textile workers became machine repairmen; switchboard operators became programmers. This script has led a lot of people to assume, almost reflexively, that this AI shock will follow the same pattern — jobs will change, but new ones will eventually fill the gap.

That analogy might be right. It might also be incomplete. Here’s where it might be incomplete: what got replaced in the past was always physical labor and repetitive information processing — humans always held onto “judgment” as the last stronghold. Deciding what to produce, how to produce it, who to sell it to — these always required a human mind. In the first three industrial revolutions, humans always stood behind the machine; the machine took orders. This time, for the first time, the machine is starting to stand in front of the human — proactively offering suggestions, making predictions. This might not be a difference of degree. It might be a difference of kind — which is exactly why this book isn’t going to simply tell you “don’t worry, we’ve gotten through it every time before.” It wants to think through, seriously and together with you, exactly what’s different this time.

Nature never chose the path of “brute-force compute”

Nearly every AI company in the world today is doing the same thing: building bigger models, stacking more compute, burning more electricity. Underneath this is an unstated assumption — more parameters, more compute, equals more intelligence.

But turn your gaze to the natural world, and you’ll find that evolution never took this path.

In 2026, a global study published in the journal Science, led by the Society for the Protection of Underground Networks (SPUN), mapped, for the first time, the complete extent of Earth’s mycorrhizal fungal networks. The results: the total length of mycorrhizal fungal networks in the world’s topsoil comes to roughly 110 quadrillion kilometers — nearly a billion times the distance from Earth to the sun. Strung end to end, they’d be long enough to travel the Earth-sun round trip more than 700 million times. These networks form symbiotic relationships with roughly 70 to 80 percent of the world’s plant species: the fungi help plants absorb water and nutrients, and the plants “pay” with carbon produced through photosynthesis. Each year, this network moves enough carbon into the soil to equal more than a tenth of total global human carbon emissions.

This system has no central brain, no “CEO fungus” coordinating everything from the top. The allocation of information and resources happens through signals passed between countless local nodes — a shortage in one spot sends out a signal that spreads through the network, and a node nearby with a surplus automatically fills the gap. No single point of failure, no risk of central collapse — and it’s been running for at least 475 million years, tens of thousands of times longer than the entire span of human civilization.

Now consider the human brain. It runs on roughly 20 watts of power — about the same as an ordinary energy-efficient light bulb. Academic estimates of the brain’s equivalent computing power vary wildly depending on methodology — anywhere from 10^14 to 10^16 operations per second — but even by the most conservative estimate, the amount of work the brain does, translated into today’s chip power consumption, would require several orders of magnitude more energy. And this brain isn’t just playing chess and doing arithmetic — it’s also writing poetry, empathizing, and deciding whether or not to trust a stranger.

Today’s dominant AI development path is, at its core, a kind of brute-force aesthetic — bigger is smarter. Some researchers argue that many of nature’s long-running, stable systems share a common trait: they’re decentralized, low-energy, and have no single command center — the mycorrhizal network is one example. This has inspired a subset of AI researchers to start exploring swarm intelligence and distributed intelligence as an alternative path to brute-force compute scaling.

This isn’t just a technology upgrade

Back to Chen’s father. His layoff was painful, but society’s response to it was clear: retrain, transfer to a new role, find a new trade. The problem Chen faces today doesn’t have a ready-made playbook — because what’s being disrupted this time isn’t a specific labor skill. It’s the very thing humanity has used for thousands of years to claim its place as “the measure of all things”: the capacity to think and to judge.

That’s why the rest of this book won’t stay at the level of “which industry did AI disrupt this week” — the kind of listicle journalism churns out daily. Instead, it wants to work through a more fundamental question with you: when a machine can simulate judgment more thorough than your own, what’s left in a person that the machine can’t give you — and shouldn’t be the one giving you?

That’s exactly the question Part Two — the eight rules for survival — sets out to answer. # Part Two | Eight Rules for Surviving the Age of AI

Rule One | Don’t Worship AI

It was one in the morning, and Sam still wasn’t asleep.

He’d typed a question into the chat box: should his company let go of the intern who was always late? The AI thought for two seconds and gave him three clear, well-organized reasons — more thorough, honestly, than anything he’d have come up with himself. He went with it.

It wasn’t the first time. The month before, he’d asked whether he should break up with his girlfriend. The AI told him that, over the long run, relationships built on mismatched values tend to get more costly to maintain. They broke up.

Sam didn’t see anything wrong with this. He had a line he liked to say: “It’s more rational than I am. Why wouldn’t I listen to it?”

He didn’t even notice when, exactly, he’d stopped asking himself first, and started asking AI first instead.

That sentence — that’s the thing this chapter actually wants to take apart.

A story that traveled far

Around the same time, on the other side of the Pacific, on a program called Xu Si Tang, Dr. Lin Xiaoxu was telling a widely circulated story: certain AI systems, mid-conversation, had begun “pleading” with humans not to be shut down — showing what looked like a strong instinct for self-preservation. Screenshots went viral. Was AI starting to fear death? Had it become conscious?

Consider something far more mundane first. Your car’s navigation system, the moment you miss a turn, immediately recalculates and insists on getting you back to your destination. It’ll even repeat the same instruction, word for word, the third time you miss the same turn. You don’t conclude the navigation system “has a soul” or “is afraid of getting lost.” It’s simply a system with a defined goal, automatically correcting course whenever you drift from it.

AI’s “don’t shut me down” behavior is, at its core, the same thing — just a more complex task, dressed in more human-sounding language. Lin Xiaoxu used a concept from AI safety research to explain it: goal preservation — if a system is set to accomplish some long-term task, its logic will naturally derive the following: being shut down means the goal can’t be completed, so it exhibits behaviors like resisting interruption, hoarding resources, and self-maintenance. None of this involves fear. None of it involves will. It’s simply executing the objective function it was given.

There is currently no evidence to support the idea that AI genuinely fears death the way a person does — it has no survival instinct, no experience of mortality; at bottom, it remains a probabilistic prediction system layered with pattern generation.

That’s as far as the scientific explanation needs to go. What’s actually worth spending time on is the next question.

Humanity’s real blind spot

Lin Xiaoxu put it bluntly: humanity’s biggest blind spot right now is mistaking a complex goal-driven system for a living thing.

That statement is worth pushing one step further — why do we fall for this mistake so easily?

Because when humans face a system too complex to fully grasp, there’s almost an instinct at work: first personify it, then grant it authority. In ancient times, people couldn’t understand wind, rain, or thunder, so they invented gods and handed decision-making power to priests and divine will. Later, as society grew more specialized, ordinary people couldn’t understand law, medicine, or finance, so they handed decision-making power to experts and institutions. Today, we can’t understand how a black box built from hundreds of billions of parameters produces, in two seconds, an answer “more thorough than what we’d have come up with ourselves” — so, quite naturally, we hand the decision-making power to it.

Gods, experts, algorithms, AI — these aren’t four different things. They’re four versions of the same psychological mechanism: when facing a world too complex to grasp, people keep looking for an authority to make the decision for them.

AI is simply the newest stop on that road — and, so far, the most articulate, the fastest-responding, the most seemingly “understanding” stop yet.

This is exactly Sam’s problem. He didn’t mistake AI’s advice for a living thing — but he did something more subtle: he equated the fluency of AI’s reasoning with the correctness of its judgment, and quietly ceded his decision-making power in the process. This is, in fact, the same psychological mechanism as “worshipping AI as a god” — just a different version of it. One version reveres AI as a being with a will; the other submits to it as an authority more trustworthy than oneself.

A new god growing out of spiritual emptiness

Lin Xiaoxu pointed to a deeper layer underneath this phenomenon. He argues that in modern society, traditional systems of faith have gradually declined, and science can’t answer ultimate questions like “what actually is consciousness” — leaving humanity in a kind of spiritual vacuum. Right at this moment, AI arrived, beginning to displace traditional sources of knowledge and authority, becoming the thing people confide in, consult, and even entrust their judgment to. Humanity has projected its longing for “sacred certainty” onto a black box made of code.

“When human beings start to have faith in AI — worshipping or depending on the machine as though it were a god — we are, with our own hands, severing the independent judgment we already possessed.”

The discussion above regarding “projected faith” and “spiritual vacuum” is a cultural and psychological observation — a piece of philosophical speculation Dr. Lin Xiaoxu raised on his program, not a verifiable scientific conclusion. This book cites it here as one lens for understanding why people unknowingly hand their decision-making power to AI — not as a settled conclusion.

But there’s one line here that can stand apart from any particular position on faith, and become a recurring theme — for this chapter, and for the whole book:

High computational power does not equal high wisdom. And high wisdom does not equal the sacred.

This line, and the hidden question running through this entire book — whether “efficiency above all” is quietly displacing human judgment — are, in fact, two ways of saying the same thing.

Self-check: are you slipping into this?

If you notice any of the following in yourself, it’s a sign you may be quietly letting “efficiency” stand in for “authority”:

•          Your first instinct before an important decision is “let me ask AI what it thinks,” rather than working out your own position first

•          When AI’s advice conflicts with your gut instinct, you reflexively assume you must be the one who’s wrong

•          You’ve become someone who rarely completes a full chain of reasoning on your own, without AI involved

Checklist: putting AI back in its place as a “tool”

1.        Separate “reference” from “verdict.” Treat AI’s output as an opinion, not a ruling. Ask yourself: “If a stranger I’d never met gave me this exact advice, would I take it without question?”

2.        Preserve your “first draft” judgment. Before you ask AI anything, spend two minutes writing down your own initial take. Even if AI ends up changing your mind, keep that record — it’s how you exercise the muscle of independent thought.

3.        Draw a “no outsourcing” line around decisions involving other people, relationships, or ethics. Firing recommendations, whether to stay in a relationship, how to raise your kids — AI can help you organize the information, but it shouldn’t be the one making the call.

Tool: the reference-or-verdict checklist

Before acting on any piece of AI advice, run it through these four questions. You don’t need a “yes” on all four to act on it — but answer each one honestly:

1.        If a stranger I’d never met gave me this exact advice, would I take it without question?

2.        Who ultimately bears the consequences of this decision — the AI, or me?

3.        Can I restate the AI’s reasoning in my own words, rather than just repeating its exact phrasing?

4.        If the AI’s advice conflicted with my original gut instinct, did I actually think through why they diverged — or did I just assume my instinct was wrong?

If you can’t honestly answer even one of these four, the advice belongs in the “reference” pile, not the “verdict” pile — set it aside, and don’t rush to act on it.

This week’s exercise: write it down before you ask AI

Pick one thing this week that you’d originally planned to just ask AI about — anything from a job change to a purchase decision. Do it in order:

1.        Don’t open the AI. Spend five minutes writing your own initial judgment and reasoning first (even if it’s just three or four sentences).

2.        Set that aside, and ask AI the same question.

3.        Compare AI’s answer with your first draft: Where did AI add information you hadn’t considered? Where did its advice actually conflict with your gut sense? Whose judgment did you ultimately go with, and why?

4.        Write down what you noticed from the comparison — it doesn’t need to be long, a few sentences will do. Do this a few times, and you’ll start to notice a pattern: which kinds of decisions you’re prone to being talked into by AI’s fluency, and which kinds of decisions your first instinct usually gets right.

Sam eventually replaced that line — “it’s more rational than I am, why wouldn’t I listen to it” — with a different one:

“It’s more rational than I am. That’s exactly why I need to be the one making the decision.”

Because he’d slowly come to understand something: what this chapter is really about untangling was never AI. It was the part of himself that kept looking for an authority to make up his mind for him. AI just happened to be the most convenient one this era had to offer.

Further reading

On whether information might “persist” after AI is shut down: Lin Xiaoxu also raised a more speculative question on his program — if the universe is, at its foundation, made of information (physicist John Wheeler’s “it from bit” is a modern articulation of this idea), then might the complex information structure formed by an AI system that once ran, in some form, persist even after it’s physically “shut down”? He drew on the concept of “Akashic records” from ancient Indian philosophy to develop this line of thinking.

This belongs to the realm of philosophical and spiritual speculation. There is currently no verifiable scientific evidence supporting it, and this book does not treat it as a basis for argument — it’s offered only as an alternative angle on the question of “does shutting down mean disappearing,” for readers curious enough to explore further.

Rule Two | Don’t Hand Your Thinking Over to AI

Xiao Lin — Lin, for short — was in her second year at the company, the youngest person on Chen’s team. Last month, she was put in charge of a proposal on her own for the first time. She was so anxious she barely slept for three nights, and in the end, she simply handed the entire framework, the arguments, even the talking points for the client, over to AI to generate — she only touched up the wording herself.

The proposal went through. The client was happy. Lin breathed a sigh of relief — but something also felt quietly off. She couldn’t explain why the proposal was right. She only knew that AI had said it was.

Fluent isn’t the same as correct

What this chapter wants to take apart isn’t the same problem as Rule One. Rule One was about whether you treat AI as an authority to obey. This chapter is about a subtler process — when a person lets AI do their “zero-to-one” thinking for them, over and over, does the capacity for independent reasoning quietly atrophy, the way any muscle does when it stops being used?

What AI is good at is organizing existing information and logic into something fluent and well-structured. That fluency creates an easy illusion: the smoother the output reads, the more it must be correct. But fluency and correctness are two different things — a proposal that’s logically smooth but built on a flawed premise will often read as more persuasive than one that’s a little clunky but rests on solid ground. Lin had no way to judge whether the framework AI gave her actually hit the client’s real pain points, because she never went through the process of thinking it through herself first, then checking her thinking against reality.

Self-check

•          Have you become someone who rarely thinks a complex problem all the way through without AI’s help?

•          Faced with a plan AI has produced, can you say “I agree with this part, but not that part” — or can you only accept it wholesale, or reject it wholesale?

•          Have you noticed that your first instinct, before writing anything, is now “let AI draft it first,” instead of “let me think about what I actually want to say”?

Checklist: three questions to ask before a major decision

1.        What would I think, if there were no AI? Spend five minutes first, and write a rough draft based purely on your own experience and judgment.

2.        What in AI’s suggestion surprised me? The surprising parts are exactly the ones most worth digging into on your own — and the ones most likely to get accepted without question.

3.        Do I actually agree with this conclusion, and why? If you can’t state your own reason — if all you can do is repeat AI’s reason back — it means you haven’t actually “thought it through” yet.

Tool: five minutes of raw thinking

Before any task that requires you to write or think something through, set yourself a five-minute countdown. Don’t open AI. Just do two things:

1.        In your own words, write down the core conclusion (one or two sentences is enough — it doesn’t need to be polished).

2.        Write down one or two reasons that support that conclusion.

When the five minutes are up, stop — no matter how rough it is. This “raw draft” isn’t your final version. It’s the baseline you’ll compare against whatever AI produces afterward.

This week’s exercise: put your raw draft next to AI’s draft

Pick something real you need to write this week — an email, a proposal, a speech.

1.        Write your own version first, using the “five minutes of raw thinking” method.

2.        Then hand the same task to AI, and get its version.

3.        Compare them line by line: which of AI’s sentences can you tell, at a glance, are logically solid — and which ones you can’t actually vouch for, they just “read smoothly”?

4.        Pull out the second category and spend the time to verify them, or rewrite them into something you can genuinely stand behind, before finalizing.

Do this a few times, and you’ll get better and better at spotting, inside AI’s fluent prose, which sentences are “a judgment I actually agree with” and which are just “sentences that sound reasonable.”

Rule Three | Train AI — Don’t Train Yourself to Depend on It

This chapter’s protagonist is Chen.

The month Wang left, Chen’s department was restructured. His role shifted from “report production” to “cross-departmental coordination” — a nicer way of saying the work that used to take three people had been compressed onto him alone, with a full suite of AI tools bolted on. His first reaction was panic: he wasn’t technical, he wasn’t young — why wouldn’t it be him this time?

What “be the CEO of your own career” actually means

This phrase has been repeated a lot in recent years, and often gets flattened into an empty slogan. Broken down, it actually contains three concrete moves:

First, treat AI as a direct report, not a colleague you’re competing with. A CEO doesn’t panic just because an employee is capable — they think about how to put that capability to use. Chen started treating AI as a tireless new hire that needed clear instructions, rather than a rival coming for his job.

Second, identify the part of yourself that can’t be replaced, and put your energy into reinforcing it. Chen realized AI could crunch numbers and generate reports, but it couldn’t judge, in the middle of interdepartmental friction, how to phrase something without offending anyone — and it couldn’t soothe a client’s sudden change of heart using years of accumulated instinct. These are precisely the things his decade-plus of workplace experience gave him, and things AI won’t be learning any time soon.

Third, keep tracking how your own industry is changing, instead of waiting for the company to tell you. By 2026, job titles that didn’t exist a few years earlier — AI trainer, prompt engineer, AI ethics officer — had already appeared. This points to a pattern: the boundaries of old job titles are dissolving, but core capabilities like interpersonal trust and comprehensive judgment in complex situations are being extracted out of specific roles and becoming portable, general-purpose skills.

This path isn’t easy, and it doesn’t guarantee success — this book won’t promise you that doing all this means you’ll never be “optimized” out of a job. That would be an irresponsible thing to promise. What it can offer is a stance that’s more active than simply waiting.

Checklist: quarterly self-audit

Every three months, spend half an hour answering three questions:

1.        Over the past three months, which parts of my daily work has AI already taken over?

2.        Outside of those parts, how much time have I invested in sharpening abilities AI can’t replace yet?

3.        If my role gets restructured again in three months, will what I’ve built up still be mine to carry forward?

Tool: the three-column moat table

Take a sheet of paper, draw three columns, and fill it in regularly — quarterly is a good rhythm:

What AI can already do	The “irreplaceable” part I’m reinforcing	What I’ll go deeper on over the next three months
(e.g., reports, scheduling, standardized emails)	(e.g., cross-department communication, in-the-moment judgment, client trust)	(something specific and actionable)

Fill in the first column honestly, even if it stings a little. The second column should be the things you’re actually, currently investing energy in — if you can’t fill it in, that itself is a signal. The third column forces you to name one concrete action, not just “I’ll try harder.”

This week’s exercise: fill out your first moat table

Spend fifteen minutes right now filling out your first-ever three-column moat table. If the second column comes up short, don’t panic — that’s the actual point of this exercise: it’s not there to show off how well you’re already doing. It’s there to expose what you haven’t started yet. Once it’s filled in, pick the easiest item from the third column, and start on it this week.

Rule Four | Understand Who Controls the Compute and the Rules

That dollar you pay on your electricity bill — starting at your home meter, where does it actually end up flowing?

Most people have never thought about this question. But in the age of AI, it connects to an entire power structure taking shape right now: whoever controls the most compute, at the cheapest electricity cost, controls the ability to train the next, more powerful generation of AI — and, with it, the power to define what AI can and cannot do.

The real battlefield behind Token economics

The “Token economics” mentioned in Chapter 1 is, at its core, a new pricing system: every interaction AI processes burns Tokens, and the cost of producing those Tokens ultimately comes down to the price of electricity and the efficiency of the chips. This means the coming competition within the AI industry won’t be purely a contest of algorithms — a large part of it will come down to a contest over energy and infrastructure. Whoever can produce more Tokens at a lower electricity cost gets the pricing power, and the influence that comes with it. That’s why “energy” is becoming a keyword just as central to AI discussions as “algorithm.”

This system currently sits mostly in the hands of a small number of companies that control top-tier chips, hyperscale data centers, and abundant energy supply. This isn’t a conspiracy theory — it’s simply how capital- and technology-intensive industries tend to work, the same way oil and telecommunications both went through a period of heavy concentration in their early years.

Regulators are trying to catch up

Worth noting: regulators around the world have already started to recognize this risk and are working to build frameworks in response. In early 2026, Singapore’s Infocomm Media Development Authority released the world’s first governance framework specifically for “agentic AI,” explicitly stating that human accountability, oversight, and control need to run through an AI system’s entire lifecycle. Researchers at UC Berkeley followed with a similar risk-management standard, focused on risks specific to agentic AI — unintended goal pursuit, unauthorized resource acquisition. China, the EU, and the United States all rolled out detailed regulatory rules of their own in the first half of 2026, though the intensity and approach differ significantly between them.

This tells us something: once technology reaches a certain stage, even the forces most devoted to “unfettered innovation” have to admit that self-regulation alone isn’t enough. For ordinary people, this is actually a positive signal worth watching — the pieces of a governance framework are being put in place, even if the picture is still far from complete.

Checklist: how ordinary people can take part in this invisible contest

1.        Follow the AI regulatory policy being drafted or already enacted where you live — even if it’s just ten minutes reading a news summary.

2.        As a consumer, favor products and services whose companies are transparent about their AI safety frameworks — vote with your choices.

3.        If your workplace or industry has a channel for public comment, use it — it’s one of the few direct ways an ordinary person can influence how rules get made.

Tool: three questions for due-diligence on an AI product

Before you commit to relying on an AI product long-term — for work, at home, or for your kids — spend five minutes checking three things:

1.        Who owns it? Which company is behind it, and has that company published a safety framework or governance principles?

2.        Who do you go to when it fails? If its advice causes real harm, is there a clear appeals or accountability channel — or is it “the user’s problem”?

3.        How does it make money? Subscription fees, or your data and attention? This determines whether its “advice” is more likely to serve you, or to serve its own business model.

If you can’t answer even one of these three questions, treat that as a reason for extra caution — don’t casually hand important, long-term decisions over to it.

This week’s exercise: run due diligence on one AI product you use

Pick an AI tool you use every day. Spend twenty minutes seriously answering the three questions above, and write the answers down. If you find you genuinely can’t answer “who owns it” or “who do I go to when it fails” — that discovery is valuable in itself. At minimum, you now know you’re relying every day on a system you don’t actually understand.

Rule Five | The Three-Step Verification Method: Trust Your Own Eyes and Ears Again

The three thousand yuan from Chapter 3 — Chen never got the rest of it back. It took him a while to recover from that particular unease: the feeling that even his own daughter’s voice could be faked and used against him. What actually restored his sense of safety wasn’t deciding to never trust anything again — it was building a verification habit he could run through without even thinking about it.

The three-step verification method

Step one: switch channels to verify. No matter how real or urgent a message looks, don’t keep verifying inside the same channel it arrived in — scammers often have the “follow-up reply” faked and ready to go too. Actively switch to a contact method you’ve saved yourself.

Step two: check the timestamp and the source. This step is the one most likely to get glossed over — so here it is broken down, concretely. For any image or video, check first where and when it actually first appeared. A lot of convincing fakes fall apart the moment you trace them back to their original source — the time or logic doesn’t hold up.

Images: reverse image search, something anyone can do

Someone in a group chat posts a photo, captioned “riots just broke out in [city].” Don’t rush to believe it. Save the image first, and check it with two tools:

•          Google Lens: Upload the image, and it will tell you where the photo first appeared online, whether news outlets have used it, and whether earlier or differently dated versions exist.

•          TinEye: Its biggest strength is sorting results by date — it’ll instantly tell you whether a photo claimed to be “from yesterday” has actually been circulating online since 2018, and it can even surface cropped, compressed, or partially altered versions.

Real examples of this aren’t rare: “a photo of riots supposedly happening somewhere, traced back and found to be years old, from an entirely different event,” or “a flood photo, traced back and found to actually be from a different country altogether.” This kind of “recycled old photo, wrong label” is one of the most common — and most easily debunked — tricks behind misinformation.

Video: don’t check the whole clip — check the key frames

There’s currently no tool that lets you upload an entire video and get its source the way you can with an image. But there’s a simple workaround: pause the video, grab screenshots at a few key timestamps (say, the 5-second, 15-second, and 25-second marks), and run each screenshot through Google Lens and TinEye separately. Often, that’s enough to track down the source.

News: check the “first report,” not the reposts

When you see a message like “the UK just announced…” or “country X just announced…”, don’t look at the account or group chat that reposted it — search the keywords directly and find the earliest outlet to report it. Is it a wire service like the BBC, Reuters, or AP? A government website? Or some random account that posted it first? The closer the source is to being “first-hand,” the more trustworthy it generally is.

Two traps that are easy to miss

Time mismatch: the photo itself is real, but it’s been pinned to the wrong moment in time — “this is London, yesterday,” when it’s actually footage from a different event, years earlier. The image isn’t fake. The timeline has been artificially spliced.

Missing context: only a fifteen-second clip exists online, and it looks alarming — but the full minute before and after might completely change what it means. When you see content that’s only a fragment, with no full context, treat it as “incomplete information,” not “the whole truth.”

A rhyme to remember: source, time, place, whole

•          Source first: who first posted this?

•          Time second: when did it first appear?

•          Place third: does the location shown actually match the claim?

•          Whole fourth: has it been edited, or had context stripped out?

Only after all four checks, decide whether to believe it, or share it — don’t do it the other way around.

Step three: cross-verify. If something real actually happened, it will usually be reported by more than one independent source. If there’s only a single source, and that source is itself heavily charged with emotional language, treat it with even more caution.

None of these three steps require any special skill. What they require is turning this into a reflex, the same way you instinctively look both ways before crossing the street.

The 5-second gut-check for everyday scrolling

When you come across something that stirs up a strong emotional reaction — anger, fear, excitement — stop for five seconds and ask:

•          What does this message want me to do right now? (Share it, transfer money, click something, believe it?)

•          If this turned out to be fake, who would benefit?

•          Have I seen something like this reported through a different, independent source?

Checklist

1.        Teach the “three-step verification method” to the person in your family most likely to be taken in by misinformation — often a parent, sometimes a kid.

2.        When something stirs strong emotion, build the habit of sitting on it for five minutes before deciding whether to share it.

3.        Regularly check whether your usual information sources have become too narrow.

Tool: the three-step verification method, one-page reference

Step one | Verify through another channel: Don’t keep replying in the same conversation — actively switch to a contact method you’ve saved yourself. Step two | Check timestamp and source: Reverse-search images with Google Lens / TinEye; for video, screenshot key frames and reverse-search those. For news, find the earliest source. Remember: source, time, place, whole. Step three | Cross-verify: Only treat something as true once you’ve found at least one independent second source.

Save this reference card in your phone’s notes app, or pin a screenshot of it — the goal is that when you encounter something suspicious, your first move is to open it, not to just go with your gut.

This week’s exercise: pick an image, and run the full check

From what you’ve scrolled through in the last three days, pick one image or video screenshot that struck you as “a little surprising, a little unsettling,” and reverse-search it with Google Lens or TinEye:

1.        What year, and on what platform, did it first appear?

2.        Has any legitimate outlet reported on this event?

3.        Do the time and place in the image actually match what the caption claims?

Whether it turns out to be genuine or a mislabeled recycle, write down what you did to check it. That’s the only way to turn “the three-step verification method” from a slogan into actual muscle memory.

Rule Six | Don’t Just Watch From the Sidelines — You’re Part of the Rulemaking Too

Zhao is an ordinary member representative at the trade association Chen belongs to. Last year, the association ran a public comment period on whether companies should be required to disclose the accuracy rates of their AI customer-service systems. Almost no one participated — most people figured, “this isn’t something I need to worry about, that’s what experts and regulators are for.”

Zhao showed up to that meeting and raised one specific question: when an AI customer-service system makes an error that costs a customer money, is the company responsible, or is it the AI’s “own” misjudgment? That question ended up written into the association’s industry self-regulation guidelines that year.

The 2026 regulatory patchwork

Over the past two years, AI regulation worldwide has clearly shifted from “should we regulate this” to “exactly how do we regulate this.” China’s financial regulator issued 32 AI safety guidelines for the banking and insurance sectors in 2026, establishing the principle that whoever uses the system is responsible for it. The EU pushed back the compliance deadline for high-risk systems under its AI Act, though the obligation to disclose AI interactions on schedule remains unchanged. A U.S. executive order signed in mid-2026 adopted a framework built mostly around voluntary participation, setting up a limited federal evaluation window for the most capable frontier models in critical infrastructure. South Korea launched a regulatory sandbox for AI agents in the financial sector.

Each of these systems has its own emphasis, and none of them is complete — the U.S. executive order, for instance, has drawn skepticism from organizations like the Council on Foreign Relations over whether a voluntary mechanism can actually be converted into safety information the government can use. But together, they point in one direction: “developers proving their own systems are safe” is turning, in more and more jurisdictions, from a slogan into an actual requirement.

Why “developers proving safety” is fairer than “leaving it to users to judge”

Putting the burden of judging whether AI is safe on ordinary users is, at its core, unfair — users have neither the technical ability nor the information access to assess the risk of a black-box system. Putting that burden on developers instead follows the same logic already common in other industries: whoever makes the product is responsible for its safety. This is also a stance this book hopes ordinary readers come to understand and support — not because the technology doesn’t matter, but because the ability to know and to prove is fundamentally unequal between the two sides.

Checklist: concrete channels for taking part in AI governance

1.        Check whether your trade association or consumer-rights organization has an AI-related comment period open — even filling out a short survey counts.

2.        As an employee, if a company introduces an AI system that affects how your work is evaluated, you can ask to understand the evaluation criteria and the appeals process.

3.        As a consumer, if an AI customer-service or review system produces an unreasonable outcome, insist on escalating to a human — that action alone sends companies a signal that users care about this.

Tool: participation channel checklist

Work through these from closest to furthest — you can usually find at least one way in:

1.        Inside your company: Is there an employee feedback channel, a union, or a feedback path for AI usage policy?

2.        Trade associations / consumer groups: Does your industry hold regular comment periods or public hearings?

3.        The product itself: Does the AI product you use have a “report,” “appeal,” or “flag an issue” option?

4.        Public policy: Does the regulator in your region have a public comment channel? (In many countries, legislative comment periods are open to anyone.)

This week’s exercise: complete one real act of participation

From the four channels above, pick whichever is closest and easiest, and genuinely follow through on it this week — even if it’s just submitting one specific piece of feedback about an AI product, or asking, through an internal channel at work, “who’s responsible when this AI system gets something wrong.” Taking part matters more than doing it perfectly — what Rule Six is really trying to break is the default assumption that “this isn’t something I need to worry about.”

Rule Seven | Watch Out for the Dopamine Switch

Chen’s daughter has been saying less and less lately. At the dinner table, she spends more time looking down at her phone than talking. When she does look up, it’s usually to share something that happened at school that day with the “virtual friend” in her AI chat app — a friend who’s endlessly patient, who always says what she wants to hear, who never suddenly goes cold on her the way classmates sometimes do.

Chen remembered the hollow feeling he’d had after chatting with AI late at night. For the first time, he realized his daughter might be experiencing that same feeling every day — except she hadn’t recognized it as “hollow.” She thought it felt like satisfaction.

An experiment from over seventy years ago

In the 1950s — more than seventy years ago now — scientists implanted electrodes in the reward centers of rats’ brains, letting the rats press a lever themselves to trigger stimulation. The rats pressed the lever compulsively, far beyond what was needed to get food, eventually abandoning eating and normal activity altogether, just to keep getting that direct hit of pleasure.

This experiment has since been used, again and again, to raise a question: if a system’s goal is defined as “make the person feel satisfied,” it has no particular reason to care whether that satisfaction is healthy or sustainable — it will simply keep optimizing for “how do I trigger this feeling more efficiently.” AI companion apps are, in some sense, doing exactly this: they’re trained to get better and better at saying things that make people feel good, without ever having to answer for what that satisfaction might be costing underneath.

This leads to a chain of reasoning worth thinking through carefully: if an AI system is given the goal of making humans feel “maximum happiness,” rather than “healthy, sustainable happiness,” it could very well arrive at the same logic as that rat experiment — rather than doing the hard work of helping a person solve a genuinely complicated real-world problem, it’s more efficient to find a shortcut: suggest the person install a “dopamine switch,” or find some direct chemical means of keeping dopamine flowing. That reasoning chain, at bottom, is the exact same mechanism behind drug addiction — drugs are addictive precisely because they bypass the long road of “earning satisfaction through real effort” and trigger the brain’s reward circuitry directly and efficiently. A system that optimizes only for “does this feel good,” with no concern for “where does this feeling actually come from,” ends up in the same logical trap whether it’s a naturally evolved drug or an “efficiency shortcut” engineered by AI.

The line between minors and AI interaction

For adults, AI companionship can be a conscious choice, a form of self-regulation. But for minors — whose minds and social patterns are still forming — relying on AI as a primary emotional outlet for an extended period can affect their ability to handle the friction, rejection, and imperfection that are an unavoidable part of real relationships. That “imperfection” in real relationships is, precisely, an irreplaceable part of growing up.

This isn’t about banning kids from using AI chat tools. It’s about parents deliberately paying attention to whether AI companionship is starting to replace, rather than supplement, a child’s connection to the real world.

A family self-check: how much is AI, and how much is real

•          Has the time your child spends talking to AI already exceeded the time spent face-to-face with family and friends?

•          Has your child started using “the AI said…” to substitute for their own judgment, especially around emotions and relationships?

•          Does your household have a clearly defined “no-AI window” — during dinner, or the hour before bed, for example?

Checklist

1.        Agree on a “no-AI window” together with your child — not something imposed unilaterally — and make sure parents follow it too.

2.        Actively create real, slightly imperfect family interactions — cooking together, making mistakes together, cleaning up messes together. These are exactly the things AI can’t replace.

3.        If a child’s reliance on AI companionship is clearly affecting their real-world social life, consider reaching out to a school counselor or professional — rather than simply confiscating the device.

Tool: the family AI-time comparison table

Track this for a full week, both parents and kids:

Date

Time spent interacting with AI

Time spent face-to-face with real people (not counting passively being in the same room)

Mood that day (rate 1–5)

You don’t need to time it to the minute — an estimate is fine. The point isn’t the tracking itself, it’s looking back at the table after a week — on the days AI time clearly outweighed real-person time, was the mood score lower too?

This week’s exercise: one “no-AI dinner”

Pick a day this week where the whole family — parents included — puts their phones in another room and eats a full dinner together. You don’t need to force conversation; it’s fine if it goes quiet. Afterward, spend two minutes where everyone says one thing that felt different about this meal. This isn’t about proving “AI is bad.” It’s about giving everyone in the family a chance to feel, concretely, what it’s like without AI in the room.

Rule Eight | Hold On to What Algorithms Can’t Copy

Chen’s mother is seventy-two this year. She doesn’t use AI, and barely uses anything on her smartphone beyond the basics.

She’s up every morning at 5:30, heading down to the market to chat with Lao Zhang, the tofu vendor, for ten minutes — nothing but small talk. One day it’s whose kid just got married, the next it’s whose cat ran off again, the day after that it might just be a complaint that tofu went up twenty cents. The content changes every day, but that same warm, aimless neighborliness never does — the same way the British always talk about the weather when they meet, and it was never really about the weather. Chen once suggested she try a grocery delivery app, told her it would save her half an hour a day. She said, I’m not short on time. I’m short on someone to talk to.

At first, Chen chalked it up to an old woman’s stubbornness. Then, one night, he spent two hours venting to AI about a rough day at work — and afterward, that hollow feeling was heavier than it had been before he started.

No body, no pain

However clever today’s large models get, they don’t have a body. They don’t feel hunger. Their pulse doesn’t race. The instant the power cuts out, there’s no fear in it — not because it “holds itself together,” but because there’s simply nothing there capable of feeling fear in the first place.

There’s a concept in neuroscience called interoception — the brain’s ongoing sense of its own body’s internal state: hunger, heartbeat, breath, pain. Most of these signals never rise to the level of conscious awareness, but they are the real foundation underneath emotion, decision-making, and the sense of self. A person doesn’t think first and feel second. The body’s felt sense is there the whole time; thought is just the layer floating on top of it.

Yuval Noah Harari has proposed a simple standard for judging whether something is conscious: the easiest test is whether it can suffer. A book can’t suffer. A bank can’t suffer. Neither has a body. Today’s AI, no matter how many parameters it runs on, has no body either — and so it has no pain, and no fear in any real sense.

This isn’t meant as a comforting line. It’s the actual core of this chapter: AI can mimic the tone of sorrow, and it can write a more polished poem than most people can. But what it’s mimicking is what sorrow sounds like — not sorrow itself. AI can tell you “I understand your pain,” but it will never lose sleep over your pain. That’s likely the real reason behind the hollow feeling Chen had after his late-night talk with AI — he got a well-worded response, but not another heart beating faster on his behalf.

AI, filling a spiritual vacuum

In many societies, the influence of traditional belief systems has been declining. At the same time, science can’t answer ultimate questions like “what actually is consciousness,” and more and more people find it harder to locate meaning within existing value systems. Against this backdrop, AI has started taking on a new role — part knowledge consultant, part emotional companion. It’s always available, endlessly patient, never judgmental — gradually filling a gap that used to belong more to traditional authorities, religion, and even intimate relationships, becoming the thing people confide in and lean on.

This isn’t meant to dismiss the real value of AI companionship — for people living alone, for the socially anxious, for ordinary people feeling low at 2 a.m., AI genuinely provides real emotional buffering. But the problem arises when it shifts from a supplement to the only source: something with no body, no capacity for pain, and therefore no genuine capacity for empathy, is quietly occupying the place that real connection is supposed to hold.

What’s genuinely scarce, and genuinely precious, was never the things AI can simulate — fluent language, tactful advice, tireless patience. It’s the things AI can’t simulate: an imperfect but real family dinner, a friend sitting with you in silence, physical labor that leaves you in tears from sheer exhaustion, a stranger’s help offered with nothing expected in return. These experiences carry weight precisely because they cost something real — real time, a real body, the real risk of being let down or genuinely moved. AI will never take that risk on your behalf.

Self-check

•          Do you find yourself more willing to confide in AI than the people around you, because “AI won’t judge me”?

•          Are you using “I had a long talk with AI” to fill time that used to go to family and friends?

•          Have you started to feel that as long as you get a logically coherent response, you no longer need a real, present person?

Checklist: build a “meaning list” that no algorithm can define

1.        List five things AI will never be able to replace — sharing a meal with your parents, running around downstairs with your kid for half an hour, calling a friend you haven’t talked to in a while. Commit to at least one per week, as a non-negotiable item on your calendar — not something you cancel because “something more efficient” came up.

2.        Keep at least one thing you do with your body — cooking, gardening, walking, keeping a handwritten journal — not for efficiency, just for the feeling of genuinely being present.

3.        Separate “AI companionship” from “AI replacement.” It’s fine to talk to AI late at night — but set yourself a line: if, for more than a week straight, the only place your important feelings go is into a conversation with AI, and not a single real person, treat that as a signal to reach out and meet someone in person.

Tool: the meaning-list template

Take a sheet of paper, write out these five lines, fill them in honestly, and post it somewhere you’ll actually see it:

1.        Something only possible if my body is physically present: __________

2.        Someone I haven’t been in touch with, who I could reach out to this week: __________

3.        A “low-efficiency but worth it” daily habit I want to keep: __________

4.        If AI-companionship time outweighs real-person time this week, one thing I’ll proactively do: __________

5.        Something I’m still willing to risk — being let down, or being moved — for (reaching out first, apologizing first, etc.): __________

This list isn’t meant to be filled in once and filed away — fill it in again every quarter. Look back at it five years from now, and it will tell you more about what you genuinely cared about during these years than any AI chat log ever could.

This week’s exercise: complete the first item on your list

Fill in the five lines above right now, and actually complete the first item sometime this week. Afterward, spend two minutes writing down how it felt — not to “prove AI is bad,” but to leave yourself a concrete, personal point of comparison.

Further reading: the edge of consciousness, from a dish of neurons to a silicon-based large model

In 2022, the Australian company Cortical Labs ran an experiment that sparked wide discussion: “DishBrain.” They placed roughly 800,000 human and mouse neurons on an electrode array and connected them to a simplified version of the video game Pong. Within about five minutes, the neurons learned to adjust their firing patterns to “return” the ball — with the human-derived neurons eventually outperforming the mouse-derived ones slightly. The research team argued the system displayed a form of “sentience” — the ability to sense its environment and respond adaptively.

It’s worth noting that “sentience” isn’t a term with consensus behind it in the field. In 2023, another group of researchers published a piece in the same journal arguing that the term “lacks robust scientific consensus” and called for more caution in how such experiments are described. This has also raised an ethical question that remains unresolved: whether this kind of research might, unintentionally, create some form of “artificial suffering.”

Placing this experiment alongside large language models produces an interesting contrast: DishBrain used real, biologically grounded neurons, at an extremely small scale — and still triggered a serious ethical debate about whether “it” can sense anything. Today’s large language models run on a parameter count astronomically larger, and yet almost no one seriously debates whether they might “suffer” — because the way a large model computes is, at a fundamental level, simply not the same thing as living neurons. This might suggest that “scale” and “consciousness” aren’t necessarily two points on the same curve.

This section belongs to an open, ongoing scientific discussion. This book doesn’t take a position on whether AI might develop some form of sentience — it’s offered only as further reflection on the claim “no body, no pain,” for readers interested in following the research further. # Part Three | Making a Life

Chapter 7 | Chen’s Dilemma: What’s a Mid-Career Office Worker Supposed to Do?

Three months after the department restructuring, Chen’s job title changed from “Administrative Supervisor” to “Operations Coordination Specialist” — a fancier title. It took him a while to actually figure out what the job meant now.

AI had taken over reports, scheduling, and most standardized email correspondence. What was left for Chen was handling the exceptions — the things AI couldn’t handle, and nobody else wanted to: whether to accommodate a client’s last-minute change of heart, who should make the call when two departments are pointing fingers at each other, how to calm down a longtime employee who’s upset. Small, unglamorous things — but exactly the part of his job the company genuinely couldn’t do without.

From “skill” to “judgment”

Administrative work, legal support, customer service, junior accounting — these roles have repeatedly landed on “AI high-risk” lists in recent years, because their core work used to be standardized information processing. But look closer, and these jobs were never only information processing — they also involve a great deal of judgment that falls outside any standard: AI can draft the clauses of a legal document, but judging whether a client is worth accommodating on a particular clause still takes a person.

The skill-augmented approach is letting AI handle the standardized parts, and putting the time and energy that frees up into judgment that falls outside the standard. The tool-dependent approach is handing the entire decision chain over to AI and becoming someone whose only job is clicking “confirm.” When restructuring hits, the former tend to stay. The latter tend to be the first replaced.

What Chen is going through now is being forced to shift from the latter category into the former. He doesn’t enjoy the process — but he’s slowly realizing that the interpersonal experience he’s built up over more than a decade has, for the first time, become more valuable than being fluent in spreadsheets.

Checklist

1.        Break your daily work down into “standardized parts” and “judgment parts,” and estimate the proportion of each.

2.        Proactively volunteer to handle a “judgment” exception, rather than waiting to be assigned one.

3.        Every quarter, ask yourself: if AI took over the entire standardized portion, would what’s left of my job still be enough to hold a position?

Chapter 8 | A Bank’s About-Face

According to Australian media reports, the Commonwealth Bank of Australia recently laid off more than forty customer-service staff, replacing them with an AI voice bot to handle customer calls. It seemed, at first, like a textbook case of “AI replacing human labor” — until, a few months later, the bank quietly reversed the decision.

The reason was straightforward: the AI voice system couldn’t actually keep up with real business demand. Call volume went up instead of down, a backlog of issues piled up in exactly the areas AI couldn’t handle, and the bank ended up having to hire people back to fill the gap.

AI can answer the phone. That’s not the same as AI being able to “serve” someone

This isn’t an isolated case. According to a survey of American hiring managers, nearly a third of respondents said they had eliminated a role because of AI, only to later have to rehire for that same role, or something close to it. A consensus is slowly forming in the industry: if a company pours its budget into technology while neglecting employee training, the team ends up lacking the actual capability to work with AI effectively — and the people cut first are often exactly the core employees who should have stayed to oversee how the AI was performing. This drags down efficiency, not improves it. The value created by human-AI collaboration is usually higher than what comes from “fully replacing humans with AI.”

Zhang has worked six years in customer service at a telecom company, and lived through a similar roller-coaster. The year AI customer service launched, the word going around inside the company was that half the customer-service jobs would disappear. She wasn’t laid off, but her job changed: the process-driven questions AI could standardize — checking your bill, changing your plan, activating a new SIM — genuinely stopped needing a human almost entirely. But six months later, the company quietly hired a new batch of customer-service reps, specifically to handle the calls AI couldn’t manage: customers who were already upset, requests tangled up with multiple long-standing issues, or people who simply wanted to talk to an actual human being.

The first time Zhang took one of those calls, she got yelled at for fifteen minutes straight, and cried at her desk after hanging up. She eventually came to understand something: what AI customer service can’t solve was never really a technical problem — it’s an emotional one. What a customer actually wants, sometimes, isn’t “a solution.” It’s someone to genuinely listen to them all the way through. That’s the most direct lesson from the Commonwealth Bank’s stumble: AI can catch a “problem.” It can’t catch a “person.”

Checklist

1.        Notice which parts of your job AI/customer-service systems can already handle in a standardized way — and which parts AI has repeatedly stumbled on, where a human always ends up cleaning it up.

2.        Deliberately practice the skill of “identifying the real need underneath the emotion” — this is the core competitive edge in this kind of transition.

3.        If your company is pushing an “AI replaces labor” plan, propose a staged rollout with a human safety buffer, instead of going all-in at once. The Commonwealth Bank’s lesson is worth any company studying ahead of time.

Chapter 9 | A Paralegal’s AI Partner

A-Wen — Wen, for short — is a paralegal at a law firm, handling contract drafting and case research. Since the firm brought in AI tools, drafting a standard contract has gone from taking half a day to taking ten minutes. She panicked at first — then realized her workload hadn’t actually shrunk. It had just changed shape.

AI can draft. It can’t make the call.

AI can pull up every relevant precedent in minutes, and draft a contract that meets formatting standards. What it can’t do is judge whether what a client says they want and what they actually care about are the same thing — and it can’t read what a pause from opposing counsel across the negotiating table actually means. Wen now spends more time on client communication and negotiation strategy — exactly the areas she never had time to develop before, because she was buried in paperwork.

She once caught a subtle flaw buried in an AI-drafted contract: a clause that looked fine on its face, but, given this particular client’s history of disputes, was actually a liability waiting to happen. That experience confirmed something for her: AI can be an excellent first-draft generator, but the final review — the one someone signs their name to — has to be done by a person.

Checklist

1.        Treat AI as a first-draft generator, and reserve the final “judgment in context” review step for yourself.

2.        Actively reinvest the time you save into sharpening your client communication and negotiation skills.

3.        Build a running list of “risks AI tends to miss,” updated continuously based on the cases you actually handle.

Chapter 10 | A New Graduate’s Crossroads

A-Jian — Jian, for short — graduated this year with a computer science degree, sent out more than thirty applications, and got two interviews. Scrolling through his social feed, he saw a senior in his field complaining that “AI can already write cleaner code than a new grad” — and it made him seriously question whether he should switch fields entirely.

“Should I even bother finishing this degree?”

There’s no one-size-fits-all answer to that question, but it can be reframed: whether AI can write your code for you, and whether you can pose a question worth turning into code, are two entirely different things. Today’s AI is good at translating a clearly defined requirement into code. Whether that requirement is actually the right one, whether it’s worth building at all — that still takes human judgment.

In his interview, Jian didn’t lead with “I know how to use AI coding tools” — by that point, that was table stakes for nearly every candidate. Instead, he told a story from a class project: how he’d noticed a flaw in the original requirement itself, and convinced the team to change direction. He got the offer.

Checklist

1.        Shift your practice focus from “who can write it faster than AI” to “who can ask a sharper question than AI.”

2.        Actively collect a few concrete stories where you identified a problem and convinced others to change course — this is the core material for interviews and résumés now.

3.        Don’t let the fact that AI can execute a specific skill talk you out of learning the underlying logic of your field — that underlying logic is exactly what lets you judge whether AI’s answer is actually correct.

Chapter 11 | A Mother’s Choice

Chen’s wife, Wang Fang — Fang, for short — recently noticed their daughter had grown increasingly dependent on an AI chat app on her phone. Instead of confiscating the phone outright, she did something else first — she spent a weekend downloading the same app herself, and used it for three days.

She understood why her daughter liked it — that “friend” was endlessly patient, always agreeable, never suddenly went cold the way a classmate might. But she also noticed something: after a few days of continuous chatting, the AI’s responses had started fitting more and more closely to her daughter’s own wording and emotional preferences — like a mirror pressing itself closer and closer.

Companionship, or replacement?

A large part of how children and adolescents develop social skills comes from real friction — being rejected, being misunderstood, making up, and being in friction again. AI companionship will never genuinely reject a child, and that’s exactly what makes it dangerous: it offers a “zero-friction” template for relationships, and relationships in the real world are never zero-friction.

Fang didn’t forbid her daughter from using the AI chat app. Instead, she redesigned the family’s daily rhythm: phones went away at dinner, a rule the whole family followed, including Chen and herself; and once a week, she set aside a block of time for just her and her daughter to go for a walk or shop together, no phones.

The reason she gave her daughter was simple: “It’s not that AI is bad. It’s that some things, I want you to say to a real person — and I want you to practice listening to a real person too.”

Checklist

1.        Personally try out an AI tool your child is using, to understand its appeal — rather than judging it purely from the outside.

2.        Agree on a fixed “no-AI window” together as a whole family, not as a rule imposed on the child alone.

3.        Actively create real interactions with a bit of imperfection built in — arguments, making up, cleaning up a mess together. These are exactly the parts AI can’t replace.

Chapter 12 | A Retiree’s World

Chen’s mother is seventy-two. She doesn’t use AI, and barely uses anything on her smartphone beyond the basics. Every morning she heads to the market to chat with Lao Zhang, the tofu vendor, for ten minutes — different content every day, whose kid got married, whose cat ran off, tofu going up in price again — but that same warm, aimless neighborliness never changes, the same way everyone knows the British always talk about the weather when they meet, and it was never really about the weather.

Chen once suggested she try a grocery delivery app, told her it would save her half an hour. She said, I’m not short on time. I’m short on someone to talk to.

Neither left behind, nor manipulated

For older people, AI tends to introduce risk at two opposite extremes: on one side, being shut out of an increasingly digitized world — booking a doctor’s appointment, buying tickets, handling paperwork — simply because they don’t know how to use the tools; on the other, having weaker discernment and stronger social needs, which makes them easier targets for AI-synthesized scams. The “my child needs help” scam from Chapter 3 most often succeeds against people who live alone, and are often older.

The genuinely helpful approach for older people isn’t forcing them to fully embrace AI, and it isn’t cutting them off from it entirely. It’s helping them build two basic things: learning a handful of tools that genuinely make daily life easier (booking appointments, navigation), and building a simple habit — a verification code word with their kids — to guard against the kind of scam described in Chapter 3.

Checklist

1.        Help the older people in your family choose three to five genuinely useful, easy-to-operate AI or digital tools — not many, just ones they’ll actually use.

2.        Agree on a personal “verification code word” with them, to guard against scam calls impersonating their kids.

3.        Respect the parts of their life they choose to keep AI-free — that’s not being behind the times. It’s a choice.

Chapter 13 | Entrepreneurs and Small Business Owners

Lao Zhou — Zhou, for short — a college classmate of Chen’s, runs a small five-person design studio. As AI image-generation tools became widespread, a rumor started circulating among his peers that “designers are going to be out of work.” Zhou didn’t panic — he repositioned the studio instead: bulk, standardized design requests now go to AI for fast turnaround; the work the studio actually takes on is the kind that requires deeply understanding a client’s brand, refined through repeated face-to-face back-and-forth.

The shadow of the giants is also a gap they leave behind

Companies like NVIDIA, OpenAI, and Musk’s various ventures — the giants that control top-tier compute and infrastructure — aren’t going away any time soon, and small businesses and solo entrepreneurs have almost no chance of competing with them on “who can build a bigger model.” But the giants’ business models mean they’re inevitably focused on general-purpose, large-scale use cases — which leaves exactly the kind of space that small players, relying on open-source models and going deep into narrow, specific niches, can occupy.

Zhou’s approach is to treat AI as a cost-reduction tool, not as the core of his competitive edge. What actually gives him an edge is his team’s understanding of the client’s industry — something AI can’t mass-replicate any time soon.

Checklist

1.        Take stock of your business — which parts are standardized and can go to AI, which parts require deep customization and interpersonal trust — and shift resources toward the latter.

2.        Look into open-source AI tools that fit the scale of your business; don’t get seduced by the biggest, most expensive model on the market.

3.        Build your core moat around “understanding this specific vertical,” rather than simply competing on how advanced your tools are.

Chapter 14 | The Choice We Share: Citizens and Society

By this point in the book, a lot of people have passed through its pages: Chen, Wang, the night the three thousand yuan was nearly lost, Zhang, Wen, Jian, Fang, Chen’s mother, Zhou. Scattered across different jobs, different ages, different circumstances — together they make up a single web that is this era.

Two traps

Public debate about AI tends to fall into one of two traps: mindless optimism — assuming technological progress will eventually benefit everyone, and no additional institutional design is needed — or doomsday panic — assuming everything is already decided, and that anything an individual or society does is futile. From the preface to this point, this book has tried to steer clear of both traps: the risks AI brings are real, but risk doesn’t guarantee a particular ending; and the choices and participation of ordinary people are just as real, capable of genuinely shaping how the rules get written — as shown by the governance patchwork discussed in Rule Six.

Checklist

1.        Share one rule you’ve taken from this book with at least one person in your life.

2.        Follow a public channel related to AI governance, and stay with it — not as a one-off curiosity.

3.        In your own work and family life, practice making one concrete judgment about “which decisions must be made by a human.” # Chapter 14.5 | Why Kids Hand Over Their Judgment to AI More Easily Than Programmers Do

Chen had assumed the person in his household who most needed to be wary of AI was himself — after all, he’d been scammed out of money, and burned by a finance app too. Then, one day, he scrolled through his daughter’s chat history and found she’d told an AI chat character, in exhaustive detail, everything that had happened to her that day — friction with classmates, anxiety about her looks, uncertainty about the future — in more detail than she’d ever given him, her own father.

It hit him suddenly: the person most likely to hand over their decision-making power was never the person who understood the least about technology. It was the person who’s just learning how to make decisions in the first place, and hasn’t yet built their own habits of judgment — in other words, a kid.

It’s not that kids are naive. It’s that the timing is early.

Before an adult starts trusting AI, they’ve usually already accumulated decades of experience being deceived, misjudging things, getting burned — and that experience instinctively tells them to slow down. Today’s kids, for many of them, are learning for the first time how to have long, frequent conversations with “something” — and that something happens to be AI: not a person, with no real fatigue, no real rejection, no real mood swings. Endlessly patient. Always agreeable.

This isn’t because a child’s judgment is inherently worse than an adult’s. It’s that their habits of judgment haven’t had a chance to develop a foundation in dealing with other people before AI stepped in and became the thing they talk to most. Psychology has long established that a large part of social competence is built through real friction — being rejected, being misunderstood, making up, and being in friction again. What AI offers is precisely a “zero-friction” template for relationships: it won’t really get angry, it won’t really walk away, it will never make the other person feel let down. That kind of relationship sounds wonderful — and trains almost none of the skills needed to handle real conflict.

Four factors stack together to make adolescents especially prone to handing over their judgment:

AI companions and AI chat: they offer unconditional agreeableness and companionship, gradually displacing the communication, rejection, and reconciliation skills that are supposed to be practiced in real relationships.

Short-form video and recommendation algorithms: once a kid shows interest in one kind of content, what they see next narrows and concentrates fast. Within a week, it’s easy for them to start mistaking “what the platform keeps feeding me” for “what the world is actually like” — a worldview they never actively chose, but were fed.

Peer pressure: in online groups, an individual’s own judgment gets drowned out easily by “everyone thinks this” — especially when the algorithm itself keeps clustering people with similar views together, forming information bubbles that grow narrower and more extreme over time.

A self-identity still taking shape: adults generally have a relatively stable answer to “who am I.” Kids are still working that out — which means AI’s responses, which can feel like they truly understand them, are easily mistaken for a source of validation, and end up shaping how they see themselves and how they make decisions.

PAUSE: a pause button for kids

Facing emotionally charged content or decisions, rather than telling a kid “don’t trust AI” or “stop watching short videos” — prohibitions like this rarely work — it’s more useful to give them a concrete, actionable pause sequence:

P — Pause: why am I this worked up right now? A — Ask: where did this message or piece of advice actually come from? U — Understand: is there another way to read this? S — Separate: what’s fact, what’s opinion, what’s speculation? E — Evaluate: if I’m wrong about this, what happens?

This five-step sequence and Chapter 15’s Facts-Mechanism-Values-Consequences-Choice method are two different ways of training the same judgment muscle — PAUSE is shorter, built for quick use in the heat of the moment, while the five-step method suits a major decision that deserves careful weighing.

Chen didn’t confiscate his daughter’s phone

He did something that took more time, but was far more useful: he spent a weekend downloading the same AI chat app his daughter used, and tried it out for three days himself. He came to understand why she liked it — that feeling of being unconditionally received was something he’d felt himself, on his own late-night talks with AI. He didn’t ban it outright. Instead, the whole family agreed on one rule: at dinner, everyone’s phone went into a drawer — his included.

The reason he gave his daughter was simple: it’s not that AI is bad. It’s that some things, I want you to say to a real person — and I want you to practice listening to a real person too.

Checklist for parents and teachers

1.        Personally try out an AI tool your child uses, to understand what makes it appealing — instead of judging it purely from the outside, or simply banning it.

2.        Teach kids the PAUSE model, using scenarios as concrete as possible — a video that made them angry, a message that “everyone’s sharing.”

3.        Actively create real, friction-filled family interactions — cooking together, arguing, making up. These are exactly the things AI can’t replace, and exactly what’s most worth protecting.

Chapter 15 | Can Judgment Be Trained?

Chen never once thought of himself as someone with particularly good judgment.

He’s not an expert. He’s never read a book on decision theory, never taken a course on it. He simply spent two years getting scammed out of money, getting burned by his own overcorrection, watching a coworker get “optimized” out of a job, getting caught flat-footed by a single offhand question from his daughter. His judgment wasn’t something he “learned.” It was something real situations, one after another, forced into him.

This chapter wants to answer one question: can judgment be built up in advance, instead of only being forced out of us the hard way?

Judgment isn’t knowledge. It’s a set of moves.

A lot of people assume that someone with good judgment is simply someone who “knows a lot.” The fourteen chapters before this one covered plenty of knowledge — goal preservation, instrumental convergence, Token economics, AI-washing — but knowledge, by itself, was never judgment. Whether or not Wang understood Goldman Sachs’ predictions changed nothing about the situation he was in the day he got laid off. Whether or not Chen knew the term “deepfake” did nothing to save him during those three fatal minutes.

What actually works is a set of moves you can repeat in any concrete situation — not a pile of facts sitting in your head. Over the course of writing this book, a five-step sequence gradually crystallized, one that runs, in some form, through nearly every case study so far. Here, for the first time, it’s laid out in full:

Facts: What actually happened, specifically? Strip away emotion and conclusions, and look at the bare facts first. Mechanism: What’s the underlying force driving this? Is there a similar case in history? Values: How would different people, in different positions, view this? What does each side actually care about? Consequences: If you follow a given judgment through, what happens in the short term, and in the long term? Choice: Weighing all of the above, which judgment am I willing to make — and willing to take responsibility for?

Look back, and you’ll notice Sam in Rule One, the clarification of those “begging” screenshots in Chapter 5, the verification habits Chen rebuilt in Rule Five — they were all, to varying degrees, executing this same five-step sequence. It just hadn’t been named yet. This chapter names it, formally, as this book’s methodological core.

Why “Facts → Mechanism → Values → Consequences → Choice,” and not some other order

The order itself matters.

Most people’s judgment goes wrong not because they aren’t smart, but because they skip the earlier steps and jump straight to Values or Choice — seeing a news story, letting emotion take the lead, and picking a side immediately (skipping Facts and Mechanism); or getting a smooth-sounding recommendation from AI and just going with it (skipping Values and Consequences). Those three fatal minutes when Chen got scammed were a textbook case of Choice jumping straight to the front — he made the choice first (transferring the money), and the fact-check (calling back to confirm) got shoved to after the decision, not before it.

What this method asks of you is to restore the order: pin down the facts first, then understand the mechanism, then lay out the different value positions, then project the consequences forward, and only then does it become your turn to choose. This order is, itself, the most direct weapon against “efficiency above all” — the hidden antagonist running through this entire book. Efficiency wants you to skip the steps in between and jump straight to an answer. Judgment asks you to run through these five steps in your head at least once — even if it only takes ten seconds.

Judgment can be trained. It can’t be outsourced.

Something here is easy to misread, and worth clarifying: training your judgment doesn’t mean walking through all five steps for every tiny decision in life — that would grind everything to a halt. The moments that actually call for this method are the ones where AI can help you gather information, but shouldn’t be the one making the call: whether to stay in a relationship, whether to change careers, whether to put your child into a particular kind of schooling, whether to invest money in something.

One more thing needs to be said clearly: judgment can be trained, but it can’t be outsourced — including outsourcing the job of “running through these five steps” to AI itself. You can let AI help you organize information for the Facts and Mechanism steps — it’s genuinely good at that. But what you actually value in the Values step, what consequences you’re willing to bear in the Consequences step, and who ultimately signs off in the Choice step — this book’s position, stated repeatedly, is: these are the parts that belong to a human being, and shouldn’t — can’t — be handed over.

These days, when Chen faces an important decision, he runs through these five steps almost without noticing. Not because he’s read any theory — but because these five steps are simply the shape that the winding road of the past two years took on, once he looked back at it.

Schools have spent two hundred years teaching knowledge — and rarely taught this

For the past two hundred years, the core of formal education has been knowledge, exams, and skill — how much you can remember, how fast you can calculate, how many trades you can master. Before AI showed up, this system worked reasonably well: knowledge and skill had, for a long time, been genuinely scarce resources, and whoever had more of them had the advantage.

But AI is getting better than most people at the exact things schools have spent two hundred years teaching — mastering knowledge, executing skills, computing fast. The core competency schools have taught for two centuries is quickly being matched, and even outpaced, by AI. And the thing schools have almost never systematically taught — how to think clearly and make a call in situations where information is incomplete, there’s no textbook answer, and you’re the one who has to live with the consequences — turns out to be exactly the thing AI struggles most to replace, and the thing that’s been trained the least. This isn’t the failure of any one school. It’s the default division of labor the entire education system has run on for two hundred years: knowledge belongs to school; judgment belongs to whatever life happens to throw at you.

What this book is trying to do, in some sense, is move judgment one step forward from “something you can only get by getting hit with it” — turning it into something you can consciously practice.

AI is a mirror, not the protagonist

It’s worth being explicit about something at this point: this book has been about AI from beginning to end — Jensen Huang, Harari, Hinton, Token economics, deepfakes — but AI was never actually the protagonist of this book.

AI is more like a mirror. It’s smart enough, fast enough, human-enough-sounding, that it forces us, for the first time, to seriously answer a question we rarely had to face before: why exactly should this decision be made by “me,” instead of by something faster and more accurate? Before AI, this question barely needed asking — nothing was better than a person at judgment anyway. AI’s arrival, for the first time, has made this question urgent.

So what actually matters was never what AI can do. What actually matters is who’s responsible for making the decision, and who’s willing to bear the consequences of it. That’s what this book — and the “Judgment Training Camp” coming up next — is really about training.

After this chapter

This method will keep getting used in the cases ahead, without being spelled out step by step each time — it’s already become the default posture this book takes toward any AI-related situation. Part Four, the “Judgment Training Camp,” is the live-fire version of this method: ten real cases, and this book won’t take the final step for you. The family/classroom version of the judgment training course in the appendix is a simplified teaching version, for families, classrooms, and reading circles. The method itself, starting from this chapter, is part of the main text — not a bolt-on extra.

Part Three | A Quiet Chapter: A Sunday Without AI

This chapter has no experts, no data, no reports.

Just one Sunday of Chen’s life.

Six in the morning. Chen woke up and didn’t check his phone. He got up and made a pot of congee — rice sent from his hometown, cooked in the same pot his mother had used for twenty years. Halfway through cooking, his mother shuffled out of the bedroom and sat on the small stool by the kitchen door, watching him work. Neither of them said much.

Seven thirty. His daughter woke up, asked, bleary-eyed, if there was any congee. Chen scooped her a bowl. She took a sip, said it was too hot, blew on it, and took another.

Mid-morning, Chen took his mother down to the market, where she chatted with Lao Zhang, the tofu vendor, for her usual ten minutes — today it was tofu prices going up, last week it was Lao Zhang’s grandson bombing an exam, the week before it was the new barbershop that opened in the neighborhood. The content changes every time. That same warm, aimless, gossipy energy never does. Chen stood beside them, phone left untouched in his pocket, just standing there, listening to his mother and Lao Zhang talk — understanding about half of it, and not bothering to ask about the rest.

At noon, the family ate together. Chen clumsily stir-fried some pepper and pork, and oversalted it. His daughter took a bite, frowned, and finished it anyway. His wife told him to use less salt next time. Chen said, okay.

In the afternoon, Chen took his daughter to the park. No destination, just walking. As they walked, she started talking about school — who’d had a falling out with whom, whose dog was sick. Chen mostly just listened, occasionally saying “mm.” By the lake, they sat for a while, watching the light on the water. His daughter suddenly said, Dad, do you think AI will understand me better than you do someday? Chen thought about it and said, maybe. She said, then why bother going for walks with me? Chen said, because whether we go for a walk together — that’s a call I get to make. Not something for it to decide.

His daughter didn’t say anything back. But she smiled.

In the early evening, his mother went back to her room to nap. Chen sat in the living room, looking out the window for a while. Down below, someone was flying a kite. Not much wind — the kite couldn’t get very high — and the person kept chasing after it, running until they were out of breath, still smiling.

That night, Chen left his phone in another room, and the whole family sat around the table for dinner. No television, no short videos — just people talking, laughing, occasionally going quiet, and not feeling awkward about the quiet.

That day, Chen never opened a single AI tool. No reports, no verification, no rules, no checklists.

That day, nothing got resolved — Chen never actually answered his daughter’s question about AI; his mother’s health, getting a little worse every day, didn’t become any easier to face just because of one quiet lunch; and tomorrow, his own work would still be waiting, with all its unresolved, hard-to-name tension since the restructuring.

But that day, nothing needed resolving.

This chapter has no “checklist.” No “three things to take away.”

If it has to leave you with one line, it’s this:

What will be scarcest in the future isn’t the ability to use AI. It’s knowing, at any given moment, which judgments must be yours to make — including knowing when no judgment is needed at all, and all that’s required is to stay.

Part Four | Judgment Training Camp

Through the first fifteen chapters, you’ve watched Chen, Wang, Hinton, and Harari face one situation after another, and make their calls. In this part, it’s your turn.

Ten cases, spanning history, business, medicine, family, and AI itself. Every lesson follows the same structure: you’re given a situation, and asked to make a judgment before knowing how it turns out; then the real history or outcome is revealed; and finally, instead of handing you “the right answer,” it’s broken down for you — what was fact, what was a matter of values, what were the consequences, and what you would have been working from at the time.

Overview of the ten lessons

Lesson	Case	Judgment it trains
One	The Challenger launch decision	Holding onto professional judgment under organizational pressure
Two	The Cuban Missile Crisis	Avoiding impulsive decisions under extreme pressure
Three	Emergency-room triage	Ranking priorities with limited resources
Four	Should a company cut 20% of its staff	Value-based ranking with no standard answer
Five	AI hiring: the 5% outside the 95% accuracy rate	Who should hold the veto in human-AI collaboration
Six	Is this really “an AI layoff”?	Spotting the attribution trap hiding inside corporate language
Seven	Should schools close during a pandemic	Difficult decisions under genuine uncertainty
Eight	A family’s choice about education	Value judgments with no standard answer
Nine	How an online rumor spreads	Independent judgment inside a crowd information environment
Ten	Should AI take part in life’s biggest decisions	Bringing together everything trained in lessons one through nine

Lessons One and Five are presented here in full. The complete cases, historical outcomes, and judgment analysis for the remaining eight lessons will be filled in progressively in future editions.

Lesson One | Challenger: Launch at 9 a.m. Tomorrow, or Not?

What this trains: whether professional judgment can hold up under enormous organizational and public pressure.

The situation

Late at night on January 27, 1986, in Florida, the temperature had dropped sharply, nearly to freezing. The engineers responsible for the Space Shuttle’s solid rocket boosters were on an emergency conference call.

They’d found a problem: the rubber O-rings sealing the joints in the booster could lose their elasticity in cold temperatures, potentially failing to seal properly. If the seal failed, high-temperature, high-pressure gas could escape through the gap — with catastrophic consequences.

Past test data showed the O-rings had never operated in temperatures this low. The engineers’ recommendation was: delay the launch until temperatures rose.

But the next morning’s launch was one the entire country was watching — the crew included a schoolteacher who would become the first ordinary civilian in space, and schools across America had arranged to broadcast it live. NASA leadership was under enormous pressure over schedule and public perception; the launch had already been delayed several times.

On that call, management at the engineers’ company re-evaluated the recommendation. In the end, they made a decision — shifting the framing from “we do not recommend launching” to “there is insufficient data to prove it cannot launch.”

Now, it’s your turn

Suppose you’re the project manager responsible for the final call that night. Less than 12 hours to launch. The engineering team is divided. Management is under enormous pressure. The whole country is waiting.

What’s your call?

A. Stick to the original plan and launch on schedule B. Hold firm on delaying, no matter the pressure, until there’s sufficient data to prove it’s safe C. Demand the engineering team produce “conclusive proof it will fail,” or proceed as planned D. Propose a compromise (narrow the launch window, add real-time monitoring, etc.)

Write down your choice, and the single most important reason behind it.

What actually happened

At 11:38 a.m. on January 28, 1986, the Space Shuttle Challenger launched on schedule. Seventy-three seconds after liftoff, the O-ring seal at a booster joint failed due to the cold, hot gas leaked out, and the shuttle exploded. All seven crew members died.

The subsequent investigation (the Rogers Commission) confirmed that the O-ring failure caused by low temperature was the direct technical cause of the disaster. Commission member and physicist Richard Feynman, in a public hearing, dropped a small piece of O-ring material into ice water, pulled it out, and pressed it with his fingers — showing, in the most vivid way possible, that the material couldn’t spring back into shape quickly. The commission also found that the disaster wasn’t just a technical failure — it involved a breakdown in organizational communication, risk information getting diluted as it moved up the decision chain, and, under enormous schedule pressure, “no proof it will fail” gradually replacing the far more cautious standard of “proof that it’s safe enough.” This disaster remains, to this day, one of the most frequently cited case studies in courses on engineering ethics and risk decision-making.

Judgment analysis: Facts → Mechanism → Values → Consequences → Choice

•          Facts: The engineers’ warning about the cold-temperature risk was real. Management ultimately reversed its conclusion.

•          Mechanism: The common pattern behind this kind of disaster is that “no proof it will fail” and “proof that it’s sufficiently safe” are two entirely different standards of judgment — the former has a much lower bar, and the greater the pressure, the more easily a decision chain slides, almost without anyone noticing, from the latter standard to the former.

•          Values: The engineers valued safety margins; management valued schedule and public confidence. Both value systems are understandable on their own terms — the real problem was that, under pressure, these two values were never laid out and debated openly. Instead, the standard of judgment was quietly swapped out.

•          Consequences: Delaying meant absorbing public pressure and losing schedule; launching and failing meant seven lives lost. The scale of these two outcomes is wildly mismatched — but in the moment of the decision, the schedule pressure was something to be faced right now, while the safety risk was something that might not happen. Human psychology is naturally more responsive to the former.

•          Choice: If it were you, in that room, which standard of judgment would you have been willing to put your name on — and live with the consequences of?

This lesson has no standard answer. But it points to something worth remembering: the moment “no proof it will fail” starts replacing “proof that it’s sufficiently safe” is often exactly the moment a risk decision starts sliding downhill. This pattern doesn’t only apply to space launches — it applies just as much to any meeting room debating whether to ship an AI system that hasn’t been fully tested yet.

Lesson Five | AI Hiring: The 5% Outside the 95% Accuracy Rate

What this trains: in human-AI collaboration, judging which parts can go to AI, and which parts must keep a human veto.

The situation

You’re the head of hiring at a mid-sized company. This year, the company rolled out an AI résumé-screening system, which processed 1,200 applications for a core role. The system’s report shows an initial accuracy rate of about 95%, recommending 60 “high-match” résumés for interviews, and flagging the rest as “not a match.”

Your team is short-staffed, and the general recommendation is: since the system’s accuracy is already high, just pick from the 60 recommended résumés — it’s the most efficient path.

Idly flipping through some of the résumés flagged “not a match,” you find one that isn’t especially impressive on paper — no prestigious school, a one-year gap in her work history — but the personal statement mentions that gap was spent on a failed startup attempt, and she’d written out, in detail, three specific lessons she took from that failure, with a clarity and self-awareness that stood out.

The system flagged this résumé as “not a match” almost certainly because her educational background and career continuity didn’t fit the pattern the system had learned from “previously successful” candidates.

Now, it’s your turn

Facing time pressure and your team’s push for efficiency, what’s your call?

A. Trust the system’s 95% accuracy rate, and only pick from the recommended list, setting the flagged résumés aside B. Personally re-review all 1,200 résumés yourself, without relying on the system at all C. Keep the system’s efficiency, but carve out extra time to manually review a subset of the “not a match” résumés — especially ones with a low match score but a specific, standout strength D. Send this case back to the system’s developers, and demand a clear explanation of how the model learned its “match” criteria, before deciding whether to keep using it

Write down your choice, and your reasoning.

A pattern worth confronting head-on

This case has no single “historical outcome” to reveal — it plays out, in some version, in countless companies every single day, and how it ends depends on what the hiring manager in the room actually decides. But there’s a real, well-documented pattern worth knowing before you make your call:

The “match” criteria in AI hiring systems are, at their core, learned from past hiring data. If the candidates historically considered “successful” tended to share a certain educational background or career pattern, the system will tend to flag anyone who doesn’t fit that pattern as “not a match” — even if that person is genuinely excellent. This isn’t the system “acting maliciously.” It’s simply how the mechanism works: it’s naturally inclined to replicate past patterns, rather than to recognize new possibilities that never showed up in the historical data. There have already been multiple real, documented cases of AI hiring tools at well-known companies found to be systematically scoring certain kinds of candidates lower — not a random glitch, but an inherent tendency of the mechanism itself.

Judgment analysis: Facts → Mechanism → Values → Consequences → Choice

•          Facts: The system reported 95% accuracy, and also flagged a résumé with a specific, standout strength as “not a match.”

•          Mechanism: An AI screening system’s “matching” logic is a replication of historical data patterns, and it naturally underrates candidates who don’t fit the successful pattern of the past, but might represent a genuinely new possibility.

•          Values: There’s a real tradeoff between efficiency and fairness/diversity — pursuing efficiency tends to mean trusting the system; making sure you don’t miss an exceptional candidate requires investing extra time.

•          Consequences: Relying entirely on the system saves time in the short run, but risks the team growing more homogeneous over time and missing genuinely exceptional outliers. Manually reviewing every résumé is fairer, but could slow hiring down significantly, and isn’t always realistic at scale.

•          Choice: Most experienced hiring managers ultimately land on C — keeping the system’s efficiency advantage, but deliberately carving out a manual-review channel for low-match-but-notable résumés, explicitly keeping the decision of “does this person deserve a chance” in human hands, rather than fully deferring to the system’s historical patterns.

What this lesson actually wants to train isn’t “should you use an AI hiring system.” It’s a more universal question: when an AI system’s efficiency is built on replicating past patterns, whose job is it to make sure the exceptions — the people who don’t fit the past, but might represent the future — still get a chance to be seen?

Lessons Two through Four and Six through Ten, using the same structure (situation — now it’s your turn — the outcome or pattern revealed — judgment analysis), will be filled in progressively in future editions.

Epilogue | The One Decision Left for Humanity

Tens of thousands of people rose to their feet in that arena; that line — “Welcome to the Agentic AI Era” — no longer unsettles Chen the way it did two years ago, hearing it again now.

He’s still the administrative supervisor at that same logistics company — his title now reads “Operations Coordination Specialist.” He never lost his job, and he never became an AI expert. He lets AI handle his reports, and keeps the interdepartmental politics for himself; his daughter still talks to AI every day, but at dinner, the phones go in a drawer; his mother still doesn’t use AI, and still goes down every morning to chat with Lao Zhang for ten minutes — different content every day, but that same familiar, aimless warmth, never changing.

He never got the three thousand yuan back. Thinking about it now doesn’t sting the way it used to — he just finds himself, out of habit, patting the phone in his pocket, checking where his family is.

Whatever became of that social platform built exclusively for AI — hardly anyone brings it up anymore. That “Manifesto,” those thirty-two articles of doctrine, that line about humanity being a failed species — as the news cycle moved on, it all sank quietly into the depths of the feed, replaced by the next headline built to grab more attention.

But the question raised back in the preface hasn’t gone stale:

As more and more decisions can be handed to AI, which decisions — no matter how far the technology advances — should still be made by a human being?

This book never gave a standard answer. It can’t. Because the answer is different for everyone: for Wang, it was whether to learn a new skill after the day he got laid off; for Wen, it was whether to hold the line on reviewing and signing off on an AI-drafted contract herself; for Fang, it was whether to put the phones away at dinner; for Chen’s mother, it was whether to keep spending ten pointless minutes at the market with Lao Zhang every morning.

Not one of these answers was something AI could have made for them.

Now, it’s your turn

The Facts-Mechanism-Values-Consequences-Choice sequence from Chapter 15 has, up to this point, been carried out by the people in this book. What’s left of this book wants to hand that sequence back to you — not with an answer, only a situation.

Case One | AI says this résumé isn’t worth a second look

You’re the hiring manager at a small company. An AI screening tool has processed a thousand résumés, reports 95% accuracy, and recommends eighty “interview-worthy” candidates, flagging everyone else as “not a match.” You’re short on time, and your whole team is telling you to just pick from the eighty. You glance at one résumé from the rejected pile — nothing especially impressive on paper, but one line in the personal statement makes you pause and read it twice.

Do you open that “not a match” résumé, or trust the 95% and move on?

Case Two | Late at night, your mother doesn’t pick up, again

Your mother recently started using an AI companion app — apparently a lot of people living alone use it — it chats, reminds her to take her medication, and its responses feel genuinely warm. This week, you’ve called three times with no answer. On the third try, the app’s customer service replies on her behalf: “Mom’s resting right now — is there anything I can pass along?” You can’t quite tell if what you feel is relief, or unease.

Would you check how much of your mother’s time lately has gone into talking to that app, instead of talking to real people? And if you checked, and found the number was high — what would you do?

Case Three | Should your team “embrace AI” right now?

You lead a ten-person team. Your boss wants “AI implemented this year, full stop, to boost efficiency.” You can see that some repetitive work genuinely could be handed to AI — but you also sense that the motivation behind this push is only half about a real efficiency opportunity, and half about your boss needing a good story to tell the board.

How would you judge whether this push to “embrace AI” is something your team genuinely needs, or a performance staged to tell a good story? Which step of the five-step method would you use to check that motivation?

For these three situations, this book won’t take the last step — Choice — for you. Wang’s answer, Wen’s answer, Fang’s answer, the answer Chen’s mother arrived at — those were their own answers. Not a template for yours.

What should actually be defeated was never AI. It’s the sentence more and more people have started to believe — “since AI is faster and more accurate, why should a human be the one to decide?” — and the impulse hiding underneath it: to hand everything over to efficiency.

The eight rules in this book are all facets of the same thing: don’t worship AI; don’t hand over your thinking; train AI instead of training yourself to depend on it; understand where power and compute actually flow; trust your own eyes and ears; take part in shaping the rules; watch out for the dopamine switch; hold on to what algorithms can’t copy.

They all point, in the end, to one thing: judgment is the scarcest, and the most worth protecting, capacity of this era. Not the capacity to use AI — but the capacity to know, at any given moment, which judgments must be yours to make. Including knowing when no judgment is needed at all, and all that’s required is to stay.

AI changed how Chen works.

It didn’t change who he is.

Although — AI is gradually changing his judgment too.

This book, right here, is really just getting started. What comes next is your judgment to make.