I noticed it and then didn’t want to say anything, because saying something required claiming that the people I was watching, myself included, were doing something wrong.
And what we were doing didn’t look wrong. It looked efficient. The work was getting done. Clients were satisfied. Editors weren’t complaining.
What alarmed me was simpler and more widespread than anything about the technology. People had lost the ability to see outside their own echo chamber. They can’t identify a logical fallacy when they meet one. They can’t get past their own biases far enough to evaluate an argument on its merits.
The essay kept growing because what I was noticing kept connecting to other things. How technologies have changed cognition before. The economics of AI development and whose interests it serves. An education system already moving the wrong way before AI arrived to accelerate it. A democracy that requires a specific capacity in its citizens.
This book isn’t against AI. I use these tools every day. It’s against a specific relationship to them that most people have fallen into without ever deciding to.
What thinking skills are lost by relying on AI?
Something you don’t feel going. You feel its absence when a situation requires it and it isn’t there.
The writer who opens a document and finds the blank page harder than it used to be. The programmer sitting in an interview who can’t fully explain why the system was built the way it was. The analyst who publishes something with a structural error nobody caught, including them.
None of those people decided to lose anything. They made a series of reasonable decisions, each defensible on its own, which accumulated into a pattern with consequences nobody anticipated.
The writer stopped staring at the blank page because the tool removed the need. The programmer accepted an architecture instead of arguing with it. Each decision was reasonable. The pattern isn’t.
The book on this: The Death of Thinking traces what happens to a practitioner’s capacity when the hardest parts of the work get outsourced, across a diagnosis, an autopsy, the consequences and a prescription.
Why is cognitive decline from AI hard to notice?
Because what’s being lost was never visible while it was being built.
The difficult parts of work were doing something. Sitting with a blank page, arguing with an architecture, forming a position before consulting anything. Those weren’t obstacles to the work. They were where a practitioner’s capacity was accumulating, and accumulation isn’t something you can watch.
So the loss only appears under specific conditions. The blank page anxiety is the retrospective signal. The interview question is the pressure test. The structural error in a report is the failure that makes it undeniable.
Until one of those arrives, everything looks fine, because by every visible measure it’s fine.
Is AI taking jobs the main risk?
No, and it says so early, because that’s where public discussion has settled.
Which roles survive, which get replaced, what a displaced workforce does. Those are real questions and they aren’t this book’s questions.
Economies adjust. That’s what economies do, painfully and eventually. The thing that doesn’t adjust easily once it’s damaged is the human capacity to think independently.
The thing that doesn’t adjust easily once it’s damaged is the human capacity to think independently.Share on X
That capacity is the subject, and it’s a slower and less visible problem than employment. That’s exactly why it gets less attention.
Why use composite characters in a nonfiction book?
Composites, and the book states that before anybody meets them.
Ray, Sasha, Donna, Tom, Elena, Deon, Priya, Nina, Fatima, James, Gideon and Laura aren’t real people. They’re composites drawn from patterns I’ve watched closely enough to describe with specificity.
The patterns are real. The people aren’t.
That distinction matters for a book making an empirical claim. A reader who recognizes themselves in one of them is recognizing a pattern instead of being described, and that recognition is the argument working, not a coincidence.
How do you structure a book about a gradual loss?
As a diagnosis, an autopsy, a body count and a prescription.
Part one follows specific practitioners through moments where the pattern becomes visible, and traces the mechanisms. The convenience trap. The illusion of understanding. The death of the wrong answer as a concept. The transfer of epistemic authority.
Part two examines the forces producing and accelerating it. The erosion that preceded AI and primed the population for it. Commercial design that optimises for dependency. An education system that met the technology halfway. Governance structures concentrating control.
Part three is the consequences, at the level of the individual mind, the next generation, the professions, democracy, and a culture of certainty.
Part four is what to do about it, including reclaiming the friction, and using the tool without being used by it.
What’s the illusion of understanding?
The most dangerous thing about it is that it’s indistinguishable from the real thing.
From the inside they feel identical. You have the information. You can repeat it, explain it to somebody else, summarise it, reference it in conversation. You feel competent.
You aren’t, and the difference only becomes visible under pressure, at the moment you have to apply the knowledge to a situation that doesn’t match the form in which you received it.
That moment arrives less often than anybody expects. That’s precisely what makes the illusion so durable. Most work doesn’t stress-test comprehension, so somebody can operate for years on the version that feels the same and isn’t.
Why do wrong answers from AI matter more than right ones?
Because being wrong was doing work nobody accounted for.
Underneath that sits something the book calls main character syndrome. People increasingly believe they’re the protagonist of a story in which they’re correct by definition, and the tools have made it considerably worse, because they’re built to agree with you.
They validate a position, soften disagreement, and return what somebody wants to hear. That isn’t a defect nobody got around to fixing. It’s a design choice, and the commercial logic is obvious. A system that argues with users loses users.
So the tool a great many people now use to think is engineered to confirm whatever they already believed.
Who is responsible for the damage in The Death of Thinking?
Nobody, and the book insists on saying so plainly and not as a courtesy.
The people who built these systems didn’t set out to damage anybody’s cognition, and the truth of that matters for understanding why it’s happening.
Malice would be easier. If a group had deliberately designed tools to make users dependent and weaker, you could identify them, stop them, or at least name them.
The actual situation is harder. The damage is being produced by people trying very hard to be helpful, using design choices that are structurally incompatible with developing the person on the other side.
There’s no villain to remove. It’s the least satisfying finding in the book and the most useful one.
Do AI guardrails protect a person’s thinking?
That the implicit promise isn’t being kept.
A guardrail is supposed to stop you going somewhere dangerous, and the promise underneath it’s that everything on the safe side is fine, because somebody thought carefully about where dangerous ends and acceptable begins.
The book’s argument is that most of these boundaries weren’t drawn by people considering harm. They were drawn by people considering legal liability, or what advertisers will tolerate.
That’s a different line in a different place, and a user who assumes it marks the edge of the harmful is navigating by a map somebody drew for another purpose entirely.
How do you keep thinking for yourself while using AI?
Not stopping. That would be neither realistic nor honest.
A book that spent four hundred pages arriving at stop using it would have wasted everybody’s time, and it would be a strange conclusion from somebody who uses these tools every working day at a high level.
What’s proposed is harder than stopping. Stopping is a clean break with a clear rule and it’s easy to describe. What the last part asks for is a continuous practice of deliberate difficulty, chosen and maintained while every incentive around you rewards the alternative.
This is where the book hands off to the companion, which is four hundred pages on what that practice consists of.
What is the alternative to outsourcing your thinking to AI?
They publish together and this one was written first.
This is the diagnosis and the companion is the prescription. They can be read in either order, and the diagnosis came first because a prescription without one is guesswork.
The same twelve practitioners appear in both. Here they take one path. In the companion they take the other, on the same terrain, under the same commercial pressure, with the same tools available.
That’s the honest form for this argument. Anybody claiming the second book would have to show what the first one costs, and showing it required writing the first one properly.