The biases you cannot think your way out of

Cognitive biases have become general knowledge, which is mostly good and occasionally a menace. The lists circulate, the names are catchy, and it's easy to finish reading one and feel slightly inoculated — as though the fix were simply to have heard of the thing.

That's the one part the research fairly consistently says doesn't work.

None of what follows is an argument for despair. It's an argument that the leverage sits somewhere other than where people look for it, which is inside their own heads, right before they make the decision.

Bias isn't stupidity

The first correction worth making is that these aren't failures of intelligence. They're mostly sensible shortcuts running outside the conditions they were built for.

Take the availability heuristic — judging how common something is by how easily examples come to mind. That's a perfectly good rule if your examples arrive from direct experience, because in a small world the things you've seen most often really are the things that happen most often. It falls apart in a world of mass media, where rare and dramatic events get reported constantly and common dull ones don't get reported at all. So people reliably overestimate deaths from shark attacks, plane crashes and murders, and underestimate the ones that quietly account for most of the total. The heuristic didn't malfunction. The environment moved.

This matters practically. If biases were failures of effort, trying harder would fix them. Because they're features of how judgement stays fast and cheap, trying harder mostly produces a more confident version of the same answer.

Knowing about it barely helps

The uncomfortable finding is that teaching people about biases produces much smaller improvements than anyone expects. People taught about anchoring still anchor. People who can define confirmation bias still weight the confirming evidence more heavily.

Part of the reason is that this stuff operates on inputs you never consciously see. By the time a judgement surfaces in awareness it already feels like the output of reasoning, because reasoning is the only part you've got access to. The tilt happened upstream, and there's no sensation attached to it.

There's also an unhelpful side effect, documented by Emily Pronin and colleagues as the bias blind spot. Learning about biases makes people markedly better at spotting them in other people while leaving their own judgement roughly where it was. The person who's read most about motivated reasoning isn't obviously less prone to it. They're just far better equipped to diagnose it in whoever disagrees with them.

Which is worth sitting with for a moment before reading the rest of this.

The ones that won't budge

Anchoring is the best demonstration, because it works even when the anchor is transparently irrelevant and everybody knows it. Amos Tversky and Daniel Kahneman set this out in 1974: spin a wheel to produce an obviously random number, ask people an unrelated numerical question, and the answers drift towards whatever the wheel produced. Participants told explicitly to ignore the anchor still show the effect. You can watch it happening to you and not stop it.

Hindsight bias is similarly stubborn. Once an outcome's known, it becomes very hard to reconstruct how uncertain things genuinely were beforehand, and past events take on a feeling of inevitability they never had at the time. This quietly wrecks how we assess decisions, because we end up judging the choice by the result when the only fair standard is the information that was available before it.

Confirmation bias is the broadest of the three and the hardest to see from inside, because it doesn't feel like ignoring evidence. It feels like noticing that the supporting evidence is solid and the opposing evidence has problems. The scrutiny is entirely real. It's just applied unevenly, and you don't get a warning light when it is.

The five worth knowing by name

A handful account for a disproportionate share of ordinary error. Recognition alone won't defeat them, but you can't build a procedure against something you can't name.

• Confirmation bias. Evidence that fits what you already think gets weighed more heavily and searched for more readily than evidence that doesn't.
• Anchoring. The first number mentioned drags every later estimate towards it, even when everyone in the room knows that number was arbitrary.
• Availability. Events that come easily to mind feel more probable, so vivid heavily-reported risks are systematically overrated and dull ones underrated.
• Hindsight bias. Once you know how it turned out, it feels as though it was obvious, which destroys most of what you could have learned from the decision as it was actually made.
• Survivorship bias. Drawing conclusions only from what made it through — successful companies, surviving buildings, returning aircraft — while the failures that would have changed the picture aren't in the room.

Wald and the aircraft that came back

The last one has the best illustration, and it's worth telling properly because the popular version has been sanded down.

During the Second World War the American military ran a Statistical Research Group out of Columbia University, and among its members was Abraham Wald, a Hungarian-born mathematician who'd fled Austria after the Anschluss. The military wanted to know where to add armour to bombers. Armour is heavy, you can't cover the whole aircraft, so the question was where it would do the most good. They had good data on damage to aircraft that returned from missions, and the hits clustered in particular places — the fuselage, the wings.

Wald's contribution was to point out that the data described a filtered sample. The planes that got hit in those places came back to be measured. The planes hit elsewhere didn't come back at all, which is exactly why the engines and cockpit looked so undamaged in the figures. Put the armour where the returning planes weren't hit.

Worth being clear about one thing: the diagram everyone's seen, the aircraft outline covered in red dots, is a modern illustration rather than Wald's own work. His actual memoranda are dense probability, and he never drew the plane. The reasoning is his; the picture isn't.

Survivorship bias is everywhere once you've got the shape of it. Every "habits of successful founders" list is built from the founders who succeeded, and the same habits are all over the ones who went under, uninterviewed. Old buildings look better made than new ones partly because the badly made old ones fell down. Fund performance tables quietly drop the funds that closed.

Numbers are where it gets expensive

Base rate neglect gets less attention than confirmation bias and probably costs more, because it turns up in medicine, in security screening and in court.

Here's the shape of it. Suppose a condition affects one person in a thousand. There's a test that catches every genuine case and produces a false positive five per cent of the time. You test positive. What are the odds you've got it?

Most people say something around ninety-five per cent. Doctors say it too, which is the alarming part — this has been put to practising clinicians repeatedly and they don't do much better than anyone else.

Do it with actual people instead of percentages and it comes apart in your hands. Test ten thousand: ten of them have the condition and all ten test positive. Of the 9,990 who don't have it, five per cent — about 500 — test positive anyway. So 510 positive results, of which 10 are real. That's under two per cent. The test isn't broken and it isn't useless; the rarity of the condition simply swamps it, and the intuitive answer was out by a factor of fifty.

Gerd Gigerenzer has spent decades showing that the second phrasing — natural frequencies, real counts of people — makes this dramatically easier for almost everyone, including the doctors. That's the useful lesson. The failure isn't really in the head. It's in the format.

The same fact, two wordings

Framing effects are the cheapest demonstration that preferences aren't as fixed as we'd like. Tell people a surgical procedure has a ninety per cent survival rate and they'll take it. Tell them it has a ten per cent mortality rate and enthusiasm drops sharply, and it drops among the surgeons too.

Tversky and Kahneman's version, published in 1981, offered people a choice of programmes to fight an outbreak, described either in terms of lives saved or lives lost. Identical arithmetic. Reliably different choices, with the gain framing pushing people towards the certain option and the loss framing pushing them towards the gamble.

You can't unframe a question — there's no neutral wording sitting underneath. What you can do is deliberately restate an important decision the other way round and see whether your answer moves. If it does, you've learned something real about how little of your preference was in the facts.

Some of the famous ones didn't survive

Honesty requires a section that most articles on this subject skip.

Psychology has spent roughly the last fifteen years in a painful and entirely necessary reckoning with reproducibility. Large coordinated replication projects went back to well-known findings and ran them again with proper sample sizes, and a substantial share came back much smaller than advertised or didn't come back at all.

Ego depletion — the idea that self-control runs on a limited pool that gets used up — was textbook material for years. A large multi-laboratory replication found an effect near zero. Social priming, where exposure to a few words was supposed to change unrelated behaviour, has fared badly enough that Kahneman himself wrote publicly that he'd leaned too hard on underpowered studies when he covered it in Thinking, Fast and Slow. Power posing was disavowed by one of its own original authors.

Two things people quote constantly are contested rather than dead, and deserve the qualifier. The Dunning-Kruger effect is real as a description of the data, but there's a serious argument that much of the famous pattern falls out of regression to the mean and the way the graph is drawn, rather than from incompetent people being uniquely incapable of seeing it. And loss aversion — the claim that losses hurt about twice as much as equivalent gains please — has been challenged as far less general than it's usually presented.

The ones described earlier in this article are among the better-supported. Anchoring in particular has come through large multi-site replications in good shape. But the right posture towards any single striking result about the mind is interest rather than conviction, and that applies to everything here as much as to anything else you read.

Where bias meets incentive

Not every systematic error is a cognitive one, and mistaking the second kind for the first sends you after the wrong remedy entirely.

If an analyst keeps producing optimistic forecasts, that might be motivated reasoning — or it might be that pessimistic forecasts have historically got people quietly moved off the project. If a doctor over-orders tests, that might be availability bias, or it might be that the cost of a missed diagnosis lands on them and the cost of an unnecessary scan doesn't. Debiasing training won't touch either case, because the behaviour is a rational response to the situation the person's standing in.

So before diagnosing a bias, check whether the error is being paid for. It's a duller explanation and it's right more often than the interesting one.

What actually shifts the numbers

The leverage sits in procedures and environments, not in willpower. Every item here works by changing the situation rather than asking anybody to think harder inside it.

Consider the opposite. Not a general resolve to be even-handed — a specific question: what would have to be true for me to be wrong here? Explicitly generating reasons against a conclusion has a better track record than open-minded intent, because it puts the missing evidence physically in front of you instead of relying on you to weigh it fairly from memory.

Write the criteria down first. Anchoring and hindsight both lose most of their grip if the standard was fixed in advance. This is the logic behind pre-registering scientific studies, and it transfers straight to ordinary decisions: agree what would count as success before you find out what happened.

Run a premortem. Gary Klein's version is a simple reframing that works better than it has any right to. Rather than asking what might go wrong, tell the group it's eighteen months later and the project has failed comprehensively, then ask everyone to write down why. Assuming the failure has already happened unlocks objections that people won't voice as hypotheticals.

Use structure where judgement is unreliable. Paul Meehl showed in 1954 that simple statistical rules matched or beat trained clinicians at prediction, and the finding has been confirmed repeatedly since across a wide range of fields. Structured interviews with fixed questions and scored answers outperform free-form conversations at predicting job performance. These methods feel mechanical and mildly insulting to expertise, and they routinely win in precisely the domains where practitioners are most certain they're unnecessary.

Change the default. Countries where organ donation is opt-out record far higher consent than countries where it's opt-in, and the gap is much too large to be about national attitudes to donation. Most people take whichever option required no action. If you want a different outcome, the cheapest intervention is usually to move which one is the path of least resistance — though defaults aren't magic either, and consent on a register doesn't automatically become a transplant.

Get an outside view. Often the single most effective correction is another person with different priors, ideally one with nothing invested in your conclusion. This works not because they're unbiased — they aren't — but because their biases point in a different direction from yours.

The one kind of training that looks like it works

There's a caveat to the gloom at the top of this article, and it's worth stating precisely, because it's narrower than people want it to be.

Reading about biases does very little. Practising against them with immediate feedback appears to do considerably more. Work led by Carey Morewedge tested training that made people commit to judgements, told them straight away when a bias had caught them, and had them try again — and the improvements were measurable and still there when participants were retested weeks afterwards.

That's a real result and it shouldn't be oversold. The gains were in the sort of tasks people were trained on, and nobody has shown they transfer neatly into a boardroom or a hospital. But the distinction it draws is the one that matters. Knowledge of a bias is not a defence. Repeated practice with correction, which is how you'd learn anything else difficult, is at least a start.

What to take away from all this

Not a list of biases and a feeling of immunity. That's the failure mode, and it's the one this whole subject encourages.

The habit worth building is smaller and more annoying. Before an important judgement, ask what would change your mind — and then check whether anything actually could. If nothing would, you haven't reached a conclusion. You've picked a side, and the reasoning arrived afterwards to tidy up.