Updating beliefs in the right direction when shown new evidence, but not nearly enough — clinging to the old view.
Imagine you think your friend always picks vanilla ice cream. Then someone tells you he actually picked chocolate the last five times. Instead of saying 'Oh wow, he likes chocolate now!' you think 'Well, he still probably likes vanilla... maybe he just tried chocolate a few times.' You barely change your mind even though the new clues are really strong.
Conservatism bias describes the systematic tendency of individuals to give disproportionate weight to their existing beliefs, predictions, or base rates when confronted with new, relevant evidence. Unlike outright denial or confirmation bias (which selectively seeks supportive evidence), conservatism bias acknowledges the new data but fails to adjust beliefs by the magnitude that rational, Bayesian updating demands. This leads to sluggish, incremental revisions where large shifts are warranted, causing people to remain anchored to outdated assessments. The effect is especially pronounced when new information is abstract, statistical, or complex, as opposed to vivid and concrete, because the cognitive effort required to fully integrate abstract data into existing mental models is substantial.
The same glitch looks different depending on the terrain. Finance, medicine, a relationship, a team — same mechanism, different costume.
Investors systematically under-react to corporate events such as earnings announcements, dividend changes, and stock splits. When a company reports earnings substantially above or below expectations, stock prices adjust, but typically not enough — a pattern known as post-earnings-announcement drift — because investors fail to fully incorporate the magnitude of the new information into their valuation models, remaining anchored to prior estimates.
Clinicians may cling to an initial diagnosis even as new test results and symptoms accumulate that point toward a different condition. The initial diagnostic impression becomes a cognitive anchor, and subsequent evidence is under-weighted, leading to delayed treatment adjustments. This is particularly dangerous in conditions with evolving presentations, such as cancers initially misidentified as benign conditions.
Teachers form early impressions of student ability based on initial assignments or classroom behavior. When students subsequently improve or decline in performance, teachers tend to adjust their expectations and grading patterns more slowly than the objective evidence warrants, perpetuating early assessments long after they have become inaccurate.
People form initial impressions of a partner's character traits early in a relationship and then under-adjust those impressions when faced with contradicting evidence over time. A partner who was initially perceived as reliable may continue to receive the benefit of the doubt long after repeated unreliable behavior, because the prior belief about their character is slow to update.
Product teams anchor on initial user research or feature assumptions and fail to sufficiently pivot when A/B test results or usage analytics present strong contradicting signals. This leads to continued investment in features that data shows users are abandoning, because the team under-weights the new behavioral data relative to their original design thesis.
Managers anchor on first impressions formed during onboarding or early performance reviews. When an employee's performance dramatically shifts — either improving or deteriorating — the manager's formal evaluations and promotion decisions lag behind the actual performance trajectory, reflecting the original assessment more than current reality.
Voters and media consumers form early impressions of political candidates or policies and then under-react to new information such as policy reversals, scandal revelations, or updated economic data. Polling shifts tend to be smaller than the magnitude of new events would predict, as partisans especially resist updating priors that conflict with their established political narratives.
Ward Edwards, 1966–1968. Edwards formalized the concept in his landmark bookbag-and-poker-chip experiments and published the key paper 'Conservatism in Human Information Processing' in 1968 (in B. Kleinmuntz, Ed., Formal Representation of Human Judgment). Phillips and Edwards (1966) also published 'Conservatism in a Simple Probability Inference Task' in the Journal of Experimental Psychology.
In ancestral environments, prior beliefs were usually built from direct, repeated personal experience and were generally reliable. Rapidly abandoning well-tested priors based on a single new data point could be dangerous — a hunter who ignored years of experience about a safe trail because of one anomalous report might walk into a predator's territory. Weighting accumulated experience heavily protected against noise and deception in an environment where information sources were unreliable.
Machine learning models trained on historical data can exhibit conservatism-like behavior when fine-tuned or updated with new data, particularly when regularization techniques heavily penalize deviation from pre-trained weights. In recommendation systems, models may be slow to update user preference profiles when user behavior shifts, continuing to serve recommendations based on outdated patterns. LLMs trained with reinforcement learning from human feedback can also reflect conservatism if evaluators systematically under-weight novel or surprising information in their ratings.
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.
