Confirmation bias is a search problem before it is a belief problem

The popular version undersells it

Confirmation bias is usually described as believing what you want to believe and dismissing what you do not. That version exists, but it is the less interesting half, and it is not what the cleanest experiments show. The more pervasive form has nothing to do with motivation and appears even when the person has no stake in the answer at all.

The mechanism is about which questions get asked. When testing a hypothesis, people overwhelmingly look for cases that would confirm it rather than cases that would break it. This is sometimes called a positive test strategy, and it is a habit of search rather than an act of denial.

The experiment that shows it cleanly

The classic demonstration gives people a sequence of three numbers and tells them it follows some rule. They may propose their own sequences and will be told whether each one fits, and they should announce the rule when they are confident. People typically form a hypothesis quickly, then propose sequence after sequence that their hypothesis predicts will fit, receive a steady stream of confirmations, and announce a rule far narrower than the real one.

Nothing about that behaviour is irrational-looking from the inside. Every test returned the expected answer. But a test that your hypothesis passes tells you very little if it is a test almost any hypothesis would pass. The informative move is to propose a sequence you expect to fail, because that is the only kind of result that can distinguish your rule from a broader one.

This particular task has been run in many variations over decades and the basic pattern is durable. It also generalises: much the same asymmetry appears when people test social hypotheses about other people, choosing questions that would produce confirming answers.

Why it is so hard to see in yourself

Because the searching feels like diligence. You looked things up, you found supporting material, you did not fabricate anything. The bias operates at the stage where you decided what to type into the search box, and that decision is normally invisible to introspection because it never presented itself as a decision.

A further complication is that the confirming material you find usually is evidence, taken on its own terms. The trouble is that you have no idea how much material pointing the other way exists, because you never sampled it, and a handful of supporting sources says nothing about the balance of what is out there when the denominator of the search is unknown. Volume of agreement measures how you searched, not what is true.

It compounds when the information environment is large. In a small library, an idle search returns whatever exists. In an environment where confirming material exists for almost any proposition, the effort required to confirm anything drops to nearly zero, so the search returns confirmations regardless of whether the proposition is true. The quantity of supporting material you can find has stopped being evidence about the world.

The asymmetry in how evidence is judged

A second, separate strand concerns standards rather than search. When evidence agrees with a view we hold, we tend to check it briefly and accept it; when it disagrees, we examine the method, look for confounds and question the sample. The scrutiny applied to disconfirming evidence is often perfectly good scrutiny — the problem is that it is not applied evenly.

This means the fix is not to be more critical, since more criticism applied asymmetrically makes the problem worse. It is to apply whatever standard you are using to both sides. A practical version: when you accept a piece of supporting evidence, ask whether you would have accepted the same design and sample size had it pointed the other way.

What reduces it

Training people to be aware of the bias produces modest results at best, which is the usual finding for debiasing by instruction. What works better is changing the task so the disconfirming question gets asked by someone whose job it is to ask it.

Formal versions of this include assigning someone to argue the opposite case with real licence to win, and requiring that a proposal state in advance what evidence would count against it. The second is unusually effective because it is cheap and because it commits you before you know which way the answer will go. A hypothesis that cannot say what would falsify it is not being tested, whatever else is being done to it.

The personal version is smaller and still worth doing. Before searching for support, write the strongest version of the opposing claim and search for that first. It takes a few minutes and it changes what you find.