Beliefs that the 2020 Presidential election was fraudulent are prevalent across the U.S. despite substantial contradictory evidence. We surveyed 1642 Americans during the U.S. Presidential vote count on November 4-5, assessing fraud beliefs and presenting hypothetical election outcomes before key states were decided. Participants’ fraud beliefs increased when their preferred candidate lost and decreased when he won, and this effect scaled with preference strength. A Bayesian model accounts for this bias as reflecting a rational attribution process operating on biased prior beliefs about the true election winner and beneficiary of fraud. Our findings suggest that a systems approach targeting multiple beliefs simultaneously may be more fruitful in combating false beliefs than direct “debunking” attempts.