Seeking, interpreting, and remembering information in ways that confirm what you already believe.
Imagine you decide your neighbor is mean. After that, every time they don't wave at you, you think 'See? They're mean!' But every time they do wave, you think 'They're just pretending.' You only notice the things that prove what you already think, and you ignore or explain away anything that doesn't match.
Confirmation bias operates across three distinct stages of information processing: biased search (selectively seeking evidence that supports existing beliefs), biased interpretation (construing ambiguous evidence as supporting one's position), and biased recall (preferentially remembering information consistent with one's views). This triad of distortions means that even when two people encounter identical evidence, each may walk away more convinced of their opposing positions — a phenomenon known as attitude polarization. The bias is not limited to strongly held beliefs; it can affect how people test even mundane hypotheses, as people tend to employ a 'positive test strategy' — asking questions whose expected answers would confirm rather than disconfirm their current theory. Crucially, confirmation bias is largely unconscious: people genuinely believe they are being objective while systematically filtering reality through the lens of their preconceptions.
The same glitch looks different depending on the terrain. Finance, medicine, a relationship, a team — same mechanism, different costume.
Investors who are bullish on a stock tend to seek out analyst reports and news articles that support their position while dismissing bearish analysis as uninformed or biased, leading to overconcentration in losing positions and delayed recognition of downturns.
Physicians who form an early diagnostic hypothesis may selectively order tests that confirm the suspected condition while neglecting to rule out alternative diagnoses, contributing to diagnostic error — particularly with conditions that share overlapping symptoms.
Teachers who form early impressions of students as high- or low-performing tend to interpret subsequent academic work through that lens, grading ambiguous answers more favorably for 'good' students and more harshly for 'weak' ones, reinforcing the original assessment.
Partners who believe their relationship is deteriorating selectively attend to their partner's negative behaviors as evidence of decline while overlooking gestures of affection, creating a self-reinforcing cycle of dissatisfaction and withdrawal.
Product teams that are emotionally invested in a feature tend to design user research that validates their solution rather than genuinely testing whether it solves a real problem, leading to the launch of features users don't need or want.
Hiring managers who form a positive first impression of a candidate in the first 30 seconds of an interview tend to spend the remaining time asking softball questions that let the candidate confirm that impression, while asking tougher questions of candidates who made a weaker first impression.
Voters preferentially consume news from outlets that align with their existing political views and dismiss reporting from opposing outlets as biased or fake, deepening partisan divides and making compromise increasingly difficult.
Peter Cathcart Wason, 1960. Formalized through his '2-4-6 Task' experiment published as 'On the failure to eliminate hypotheses in a conceptual task' in the Quarterly Journal of Experimental Psychology. The concept was comprehensively reviewed by Raymond S. Nickerson in 1998.
In ancestral environments, rapidly categorizing stimuli based on prior experience was critical for survival. If rustling bushes previously indicated a predator, treating all future rustling as dangerous — and ignoring the times it was just the wind — was an asymmetrically safe error. Building on and confirming existing mental models of the environment allowed for faster decision-making and group cohesion, where shared beliefs enabled coordinated action against threats without the costly delay of re-evaluating every assumption from scratch.
Machine learning models can exhibit confirmation bias when trained on historically biased datasets, reinforcing existing patterns rather than discovering novel ones. For example, a hiring algorithm trained on past successful hires — predominantly from one demographic — will preferentially score similar candidates higher, perpetuating the original bias. In LLMs, confirmation bias emerges when models are fine-tuned or prompted in ways that favor certain viewpoints, producing outputs that echo training data distributions rather than reflecting balanced evidence. Recommender systems amplify this further by creating algorithmic filter bubbles that confirm users' existing preferences and beliefs.
50 cards are free to preview. Buyers unlock the rest of the deck plus the interactive training — Spot-the-Bias Quiz unlimited, Swipe Deck with spaced repetition, My Blindspots, Decision Pre-Flight, the Printable Deck + Cheat Sheets, and the Field Guide e-book. $29.50$59.
