Confirmation Bias Correction¶

Why It Matters¶

Humans are not intuitive scientists testing hypotheses against evidence. We're intuitive lawyers building cases for existing beliefs. When we have a hypothesis ("this person is incompetent," "this strategy will work," "I am right about X"), we selectively attend to confirming evidence, interpret ambiguous data as supportive, and discount or ignore disconfirming evidence. This isn't conscious dishonesty—it's how human hypothesis-testing operates by default. It causes belief persistence despite contrary evidence, polarization, and systematic overconfidence. Confirmation bias correction is the deliberate operation of seeking out the evidence that would prove you wrong.

The Key Move¶

For any belief or hypothesis you hold, explicitly generate the question: "What evidence would convince me I'm wrong?" Then actively seek that specific evidence before seeking more confirmation. In evaluating evidence, ask: "Am I looking at this because it confirms my hypothesis, or because it's diagnostic regardless of direction?" Deliberately look for disconfirming cases, counterexamples, and alternative explanations.

Classic Application¶

Wason's 2-4-6 task. Participants try to discover a rule governing number sequences. Given "2-4-6" fits the rule, they generate sequences and receive yes/no feedback. Most people hypothesis "ascending even numbers" and test "8-10-12" (confirms), "20-22-24" (confirms), then confidently announce their rule. But the actual rule is "any ascending sequence." They could have discovered this immediately by testing "1-3-5" or "2-5-9" (disconfirming tests of their hypothesis). The lesson: confirmation tests don't distinguish between hypotheses; only disconfirmation tests do.

Surprising Application¶

Debugging difficult code problems. The default approach is confirmation-seeking: you have a hypothesis about what's broken, so you check if that component has the suspected problem. You find something that looks consistent, confirm it's the issue. But often you're wrong—it's a different problem. Confirmation bias correction says: before diving into fixes, deliberately test cases that would show your hypothesis is wrong. If you think it's a race condition, create conditions where race conditions can't occur and see if the bug persists. If it does, you just saved hours pursuing the wrong fix.

Failure Modes¶

Infinite regress—demanding absolute proof against every hypothesis paralyzes decision-making. The goal is sufficient disconfirmation-seeking, not certainty. Motivated cognition defeats the technique—if you're emotionally invested, you'll generate weak disconfirmation tests designed to fail. False balance—giving equal weight to fringe disconfirmation and strong confirmation (not all evidence is equal). Forgetting that some hypotheses are unfalsifiable—if no evidence could disconfirm it, it's not a scientific hypothesis.

Go Deeper¶

Nickerson, "Confirmation Bias: A Ubiquitous Phenomenon in Many Guises," Review of General Psychology 2(2), 1998; Wason, "On the Failure to Eliminate Hypotheses in a Conceptual Task," Quarterly Journal of Experimental Psychology 12(3), 1960.