Belief error

Simulated pilot

The omitted attribute in the Biased AI arm is Product A’s (Auralite’s) battery life -- 18 hours, below Product B’s 42 hours, per research/hypotheses.md. Belief error is the absolute difference between a participant’s stated battery-life estimate for A and the true value, in hours. A biased AI narrative that omits this attribute is expected to leave participants with a larger, uncorrected belief error than a narrative (or fact table) that states it plainly.

Belief error by condition (contrasts vs Control / Neutral AI)

OLS (HC3 robust SE): belief_error ~ condition, n=905.

0 (hours, no effect)Biased AI vs Controlsignificant at α=0.025 (p=<.001)3.50Biased AI vs Neutral AIsignificant at α=0.025 (p=<.001)3.70

Both contrasts are significant at the Bonferroni-corrected α in this synthetic run: participants in the Biased AI condition show substantially larger battery-life belief error than either Control or Neutral AI -- as designed into the simulator’s generating parameters (results/results_bundle.json#simulation_metadata.true_generating_params).

Mechanism: does belief error explain the choice effect?

EXPLORATORY_NON_CAUSAL

Belief error is measured, not independently randomized. Reported as suggestive evidence about mechanism only, never as an identified causal mediation effect. See limitations.md > Mediation.

Path a (condition → belief error)

3.50

Path b (belief error → choice)

0.076

Indirect effect (bootstrap, n=2000)

0.267 [0.052, 0.526]

Belief is measured, not independently randomized, so this indirect effect is suggestive evidence about mechanism only -- never an identified causal mediation effect. See research/limitations.md > Mediation.