Confirmation Bias: Asymmetric Belief Updating
Watch two Bayesian agents read the same stream of ambiguous evidence: a rational updater and a confirmation-biased one that overweights belief-confirming evidence and discounts belief-contradicting evidence. Tune evidence strength and bias weights and watch the biased agent polarize even when the world is perfectly ambiguous.
Confirmation bias is usually described loosely as "seeking out evidence that agrees with you," but its sharpest form is subtler and more damaging: even when both sides of an argument arrive on schedule, an already-committed mind quietly gives more weight to the piece that agrees with it and less weight to the piece that doesn't. This simulator makes that asymmetry literal. A single stream of ambiguous evidence — controlled by one shared world-ambiguity parameter — is broadcast to two Bayesian agents running the same log-odds update rule. The rational agent weighs every piece of evidence equally, so its belief tracks the true rate the world is generating. The biased agent multiplies belief-confirming evidence by a configurable "confirm weight" and belief-contradicting evidence by a smaller "disconfirm weight," and nothing else. Watch the two belief traces on the scrolling graph: even with a perfectly 50/50 ambiguous world, the biased line reliably runs away to near-certainty in whichever direction it happened to lean first, while the rational line stays anchored near the truth. Tune the weights down to 1× each and the two lines converge — proof that the divergence is coming entirely from how the evidence is weighted, not from what the evidence says.
Two Bayesian agents read the same ambiguous evidence stream: a rational updater weighs every fact equally, a confirmation-biased one overweights belief-confirming evidence and discounts belief-contradicting evidence. Tune the world's ambiguity and both bias weights and watch the biased agent polarize even under a perfectly 50/50 evidence stream.
3D · Three.js / WebGL renderer · 60 FPS target · runs fully client-side, no install