Think Again — About AI

admin

Using Adam Grant’s ideas about rethinking to look at how we build, fund, and use artificial intelligence

In his book Think Again, Adam Grant makes a simple point: in a world that changes fast, the most useful skill is not how smart you are. It is how willing you are to rethink what you believe.

Grant says we often slip into three habits. The preacher defends beliefs like they are sacred. The prosecutor attacks anyone who disagrees. The politician just says what the crowd wants to hear. He suggests a better habit: think like a scientist. Treat your opinions as guesses to be tested. Stay open to being wrong. Hold what he calls confident humility — believe in your ability to learn, but stay humble about what you know right now.

AI seems like a good place to try this out. The technology is moving faster than our opinions about it can settle. That makes it easy to get stuck in one of those three habits. So let me try Grant’s approach in two settings: first the business and investing world, and then the individual person.

A quick note before I start. I have tried to separate two things in this article. When I state a fact, I have included a reference so you can check it. When I share a guess or an opinion, I have tried to say so clearly. Those personal views are collected at the bottom of the article.

Part One: Business and Investing

Preaching, prosecuting, and politicking about AI

Listen to many earnings calls or sit in many strategy meetings, and you may hear Grant’s three habits at work.

The preachers treat AI like a creed. “AI-first” becomes something you are not allowed to question, and money flows toward the belief instead of the evidence.

The prosecutors are the opposite. They have decided the whole thing is a bubble, and they treat every failed project as proof. They may be right that there is hype. The risk is that they treat their doubt as a closed case instead of a view they keep testing.

The politicians may be the trickiest in a market. They adopt whatever AI story analysts or the press seem to want this quarter. When the mood shifts, the plan shifts too — not because the evidence changed, but because the audience did.

The scientist’s version

Grant’s scientist does not ask “How do I defend my position?” Instead, the scientist asks “What would change my mind?” For a leader or an investor, that could reframe the questions they ask:

  • Instead of “Is AI overhyped or underhyped?” — a question built to be argued — you might ask “What specific, measurable result would tell me my current bet is wrong, and by when?”
  • You might treat a strategy as an experiment with a clear stopping point set in advance. If a company punishes people for ending a project, it may keep funding weak bets long after the evidence says to stop.

There is one tension worth naming. Markets often reward confident, exciting stories. Good thinking, in Grant’s view, asks for confident humility instead. In my opinion, the gap between those two pressures is where a lot of bubbles and crashes seem to live — though that is my read, not a proven claim.

Part Two: The Individual Person

If the business risk is being too sure, the individual risk may be quieter. It may be the slow handing-off of the thinking itself.

Handing off your thinking

Grant’s whole idea assumes you actually do the first round of thinking. You form a view, then you have the humility to look at it again. But AI offers a shortcut: skip forming the view at all, and accept a smooth, confident answer instead.

This matters because rethinking is something you can only do to a thought you already had. If the first pass is handed off, there is no first draft of your own to go back and improve.

There is research that points in this direction. A 2025 study by Michael Gerlich, published in the journal Societies, surveyed 666 people and found that heavier use of AI tools was linked to lower scores on critical thinking, and that “cognitive offloading” — letting a tool carry the mental work — appeared to explain much of that link (Gerlich, 2025, Societies). The effect looked stronger among younger users in that sample.

It is worth being careful here. This kind of study shows a correlation, not proof that AI use causes weaker thinking. People who already think less may simply lean on AI more. My own view is that the risk is real and worth taking seriously, but the science is still young, and I would not state it as settled.

Losing the first-principles habit

There is a second concern. AI tools usually hand you a result, not a chain of reasoning you can check. You get the answer without the steps that led to it.

That makes it harder to think from first principles. To use Grant’s scientific approach, you need to ask “What is this claim built on, and is that base solid?” That question is hard to ask when the answer arrives as a finished package. You may end up judging how smooth and confident the answer sounds rather than whether it is correct — and sounding fluent is exactly what these tools do well, whether or not the claim is true.

In my opinion, a sensible response is to treat AI output as a guess to test, not a verdict to accept. You can use the tool to make a rough draft, then take it apart yourself: ask what would make it wrong, and rebuild the reasoning in your own head.

When AI trains on its own output

The third concern is bigger than any one person, and it is the one I find most interesting.

As AI-made content fills up the web, future AI models end up training on the output of earlier models instead of on original human work. Researchers have a name for the damaging version of this: model collapse.

This one is more than a guess. In a 2024 paper in the journal Nature, Ilia Shumailov and colleagues showed that when models are trained over and over on data made by earlier models, they begin to lose the rare and unusual cases first, and over several rounds their output drifts away from the real range of human data and degrades (Shumailov et al., 2024, Nature, “AI models collapse when trained on recursively generated data”). In one example, a passage that started out about medieval architecture turned into nonsense about jackrabbits after several generations (reported in Nature; see also coverage in The Register, 2024).

I want to be fair about the limits, though. Later research suggests collapse is not guaranteed in the real world. If fresh human data keeps getting added to the mix, the worst outcomes can be reduced or avoided (see, for example, work presented at ICLR 2024 on retraining models that include their own data, summarized by Inria, 2025). So this is a documented risk, not a certain fate.

Here is my own thought, offered as a guess rather than a fact: this looks like the big-picture version of the individual problem. A mind that stops making its own first drafts, and only edits borrowed ones, may drift toward the average and lose its surprises. A model trained on its own output may do something similar. Both are systems that may have stopped thinking again because they stopped thinking first.

Pulling It Together

What links the business world and the individual may be that both are struggles with the same skill, pointed in opposite directions. Companies can lean toward being too sure — preaching the thesis, prosecuting the doubters, following this quarter’s story. Individuals can lean the other way — never forming a view at all, letting the tool write the first draft and forgetting there was meant to be a second one.

Grant’s advice is the same for both: hold your views as guesses, decide ahead of time what would change your mind, and treat being wrong as something you learned rather than something you lost.

There is a real choice here, and in my opinion it is not decided by the technology itself. AI could help us all become better scientists — a patient partner to argue against and a machine for generating guesses to test. Or it could help us all become better preachers, handing us smooth certainties we never earned and never question. Which way it goes may depend on whether we are still willing to think again.

A note on facts and opinions

I have tried to keep two things separate in this article. Statements of fact are tied to a named source so you can check them yourself. Statements marked with phrases like “in my opinion” or “my own view” are my personal take, offered in the spirit of Grant’s book — as guesses to be tested, not conclusions to be defended. I may be wrong about them, and I would welcome being shown why.

References

  • Grant, A. (2021). Think Again: The Power of Knowing What You Don’t Know. Viking.
  • Gerlich, M. (2025). “AI Tools in Society: Impacts on Cognitive Offloading and the Future of Critical Thinking.” Societies, 15(1), 6.
  • Shumailov, I., Shumaylov, Z., Zhao, Y., et al. (2024). “AI models collapse when trained on recursively generated data.” Nature, 631, 755–759.
  • The Register (2024). “AI models face collapse if they overdose on their own output.” (Coverage of the Shumailov et al. study.)
  • Inria (2025). “Could we see the collapse of generative AI?” (Summary of related ICLR 2024 research on retraining stability.)

Inspired by Adam Grant’s Think Again (2021). The personal opinions in this article are my own.