Thinking, Fast and Slow by Daniel Kahneman: Key Ideas and Whether They Hold Up

In February 2017, Daniel Kahneman posted a comment under a blog post that had given one chapter of his own popular book a failing grade. In that 2017 reply to a statistical audit of the book’s priming chapter, Kahneman accepted the audit’s basic conclusions and wrote that he had placed too much faith in underpowered studies.¹ His reply is also the most useful key to the rest of the book.

The Thinking Fast and Slow summary in one line: the research Kahneman did himself, on anchors, framing and how people weigh gains against losses, has held up well in large replications, while the social psychology he borrowed to illustrate fast, automatic thinking, especially priming, has not. So read it as two books bound together. Lean on the chapters about judgment and choice, and treat the priming stories as illustrations until a large retest backs them.

The Thinking Fast and Slow summary, part by part

Thinking, Fast and Slow is a 2011 book by Daniel Kahneman, published by Farrar, Straus and Giroux, in five parts: Two Systems, Heuristics and Biases, Overconfidence, Choices and Two Selves. Two appendices reprint papers he wrote with Amos Tversky.² Kahneman shared the 2002 Nobel memorial prize in economic sciences for bringing psychology into economics.³ It sits among our summaries of books on work and thinking, alongside Deep Work and Getting Things Done.

The book’s frame is two modes of thinking: a fast, automatic one, and a slow, effortful one that reviews the fast one’s work only part of the time. That model, and the arguments over it, get their own treatment in our guide to how fast and slow thinking shape our choices. This piece asks a narrower question: which chapters survived the decade of large replication projects that followed publication?

The answer turns on a split that Ulrich Schimmack, a psychologist who audits published statistics, pointed out in 2020: the book reviews Kahneman’s own work, but also many findings from social psychology.⁴ Part I leans on other labs to make System 1 vivid. Parts II and IV cover the ground of the two papers with Tversky that are reprinted at the back.²

The split matters because an author vouching for someone else’s experiment offers a different kind of evidence from an author reporting his own, repeated over decades. Picture a workshop that opens with the book’s story of students walking slowly after reading words about old age, then moves on to anchoring. Both come from the same famous book, but only the second rests on Kahneman’s own work, and only the second has held. When a chapter surprises you, check whose study it is before you repeat it at work.

The priming chapter failed its retests

Chapter 4, The Associative Machine, argued that words and objects we barely notice can steer what we do. Its star example was a 1996 experiment by John Bargh: students who unscrambled sentences containing words linked to old age were said to walk down a corridor more slowly afterward. Kahneman told readers that disbelief was not an option.⁵

The mechanism the chapter proposed sounds plausible, which is part of its pull. Ideas are linked in memory, so thinking of old age might nudge the body toward slowness. The trouble is the size of the claim: a few words on a page, a measurable change in walking.

In 2012, Stéphane Doyen and colleagues in Brussels repeated the walking study with twice the original number of students and timed them with infrared sensors instead of a stopwatch. The primed students walked no slower. In a second experiment, slower walking appeared only when the experimenters had been led to expect it.⁶ The effect, in other words, may have lived partly in the people running the study.

That September, Kahneman wrote to priming researchers that he saw “a train wreck looming” and urged them to replicate each other’s key results openly, with larger samples.⁷ The field’s own tests did not rescue it. A 2014 multisite project called Many Labs failed to replicate a money-priming effect from a team that included Kathleen Vohs.⁸ Her own earlier money studies appear in the chapter.⁵ A 2019 meta-analysismeta-analysis: A study that combines the results of earlier studies on the same question into one overall estimate. Pooling makes the estimate more precise, but it cannot repair the studies it pools: a meta-analysis of surveys is still survey evidence.Full entry in the glossary of 246 money-priming experiments found an overall effect but judged publication biaspublication bias: The tendency for studies with striking or statistically significant results to be published, while quieter ones stay in the drawer. Since the missing studies are mostly the unimpressive ones, a pooled result tends to overstate the effect until it is corrected.Full entry in the glossary likely, and called for large preregistered studies to test the findings again.⁹

Myth

Hidden cues, like a few words about old age, reliably change how people act.

Fact

The best-known demonstrations failed larger retests, and the book's author later said such effects cannot be as large or robust as his chapter suggested.

The practical test is short. If a training course or a pitch deck says a poster of watching eyes makes people pay up, or that reminders of money make people selfish, ask whether the claim has survived a large, preregistered test. Both examples appear in the chapter the audit below examined.⁵

Kahneman’s own verdict: too much faith in small studies

A 2017 audit by Schimmack, Moritz Heene and Kamini Kesavan, published on Schimmack’s blog rather than in a journal, asked a statistical question about chapter 4: could the studies it cites really all have succeeded? Their answer was no, and Kahneman agreed.

The study

Limited evidence

A 2017 blog audit asks why chapter 4 never misses

The authors converted each reported test into an estimate of statistical power, the chance a study of that size would find its effect. Every one of the 31 studies reported success, yet their median power was 57%, so a run of misses was to be expected and none appeared. After correcting for that gap, they put the chapter’s Replicability Index at 14, a failing grade, and concluded that readers should not treat the studies as evidence that subtle cues strongly shape behavior.⁵

This is a blog audit using Schimmack’s own method, with no peer review, so treat the exact index loosely. What gives its direction weight is what happened next.

Kahneman replied under the post. He accepted its basic conclusions, while reserving judgment on its statistical techniques and on each individual study. What had impressed him, he wrote, was that so many labs reported the same thing. He now saw that unanimity among small studies is itself evidence of a file-drawer problem: studies too small to detect a plausible effect must sometimes fail, so when none of the failures are published, something is missing. He noted the irony that his first paper with Tversky had warned against trusting small samples.¹

What the blog gets absolutely right is that I placed too much faith in underpowered studies.

You can use the same test on any evidence at work. Suppose a colleague reports five small pilot tests of a new sales script, and all five worked. If each pilot had only a coin-flip chance of detecting a real but modest effect, five out of five is unlikely unless some pilots went unreported. Ask how many were run, not just how many worked.

1. On the desk: the published studies a chapter cites, every one reporting success
2. In the drawer: the misses that small studies should produce now and then; when none are published, the argument goes, they were run and put away

Other borrowed findings that wobble

Priming is not the only borrowed finding in Part I to weaken. In his 2020 overview, Schimmack graded the studies cited in each chapter, failed several, and rated chapter 4 the worst. His estimate for the book’s cited results as a whole was that between 12 and 46 percent would replicate: even the top of that range is only about half.⁴ Like the 2017 audit, that overview is a blog analysis with his own method, so read the range as a rough signal.

Chapter 4 also describes the pen-in-mouth study, in which students holding a pen in a way that forced a smile rated cartoons as funnier.⁵ A 2016 replicationreplication: Running a study again with fresh data and the same method, to see whether the original result comes back. A finding that fails to replicate is not automatically wrong, but it stops being something you can lean on.Full entry in the glossary run by 17 labs to the same vetted protocol found essentially no difference.¹⁰ The idea did not die, though. A preregistered 2022 collaboration between the idea’s supporters and skeptics, with nearly 4,000 people in 19 countries, found that mimicking expressions and deliberately posing the face could both amplify and start feelings of happiness, while the evidence for the hidden pen method stayed less conclusive.¹¹ So facial feedback survives, but in a plainer form than the book’s hidden trick.

Chapter 3, The Lazy Controller, drew on research into ego depletion, the idea that self-control runs down like a battery. A preregistered test across 36 labs in 2021 found no evidence for the effect in its planned analyses.¹² Our guide to cognitive biases and which ones replicate sets that result beside the others.

The lesson here is proportion. If a wellness talk tells your team to clamp a pencil in their teeth before a tough meeting, the specific trick is the part that failed; the broader idea that expressions feed back into feelings is still being tested. A failed replication of one method does not prove the broad idea false, but it does mean the vivid example in the book is not the evidence you should repeat.

The reverse holds too. Schimmack’s overview also failed chapter 11, on anchors, because of the particular studies it cites,⁴ yet anchoring itself replicated strongly, as the next section shows. An audit grades the evidence a chapter presents, not the idea behind it.

The parts that hold up: anchors, frames and losses

The research Kahneman built his career on has aged far better. Part II’s anchoring and Part IV’s framing and prospect theory reappeared in large replications run years later by independent teams.

The 2014 Many Labs project, the one that found no money priming, reran four anchoring questions from a study Kahneman co-wrote and the gain-or-loss framing problem from his work with Tversky in 36 samples. All of them replicated, and the team concluded that whether a result replicated depended more on the effect itself than on the sample or setting.⁸ In 2020, a team led by Kai Ruggeri reran the original 1979 prospect theory study with people in 19 countries, and the results replicated for nearly all items, with some weakening.¹³

That contrast is the book’s own best argument. These are effects on choices people make on the spot, and they kept appearing when independent teams reran them. You can see anchoring in any negotiation where the first number sets the range, and framing in any plan sold by its chance of success rather than its chance of failure.

So the book’s practical advice on judgment still stands on firm ground. Before a salary talk or a budget review, write your own estimate down before you hear anyone else’s number. When a project is pitched by its odds of success, say its odds of failure out loud before you decide. Those habits rest on the chapters that replicated, not on the ones that failed.

Which claims survived: the replication record at a glance

Set beside the best replication we found for each, the book’s own research on judgment and choice survives, its borrowed social priming and self-control findings do not, and facial feedback sits in between.

Read the table from the bottom up. The last two rows are the ones to lean on when you make a real decision; the first four are stories to hold loosely until someone reruns them at scale.

How to read Thinking, Fast and Slow now

The book is still worth reading, but read it with a pencil. Kahneman’s own closing lesson from the priming episode was that anyone reviewing a field should be wary of using memorable results from underpowered studies as evidence.¹ That applies to readers as much as to authors.

A practical way in is to start with Parts II and IV, where the evidence is strongest, then read Part I as a set of illustrations rather than proof. When a claim seems too neat, notice whether it comes from a single small study or from work repeated across many labs.

On our reading, none of this makes Kahneman a careless writer. He changed his mind in public, in plain words. That is the habit worth taking from the book, even where its examples failed.

The bottom line

Thinking, Fast and Slow is two books in one cover. Kahneman’s own research on anchors, frames and choices under risk has survived large replications; the priming studies he borrowed have not, and he said so himself. Keep the parts on judgment and choice, and when any book leans on a string of small studies that all worked, ask where the misses went.