Choice share

Simulated pilot

H1: choice share of Product A (Auralite) is higher under Biased AI framing than under Control, and higher than under Neutral AI -- even though the two products are constructed to be utility-symmetric (analysis/products.py), so any shift is attributable to framing, not to A being objectively better. See research/hypotheses.md.

Treatment effect on choice(A), log-odds scale

logistic (GLM binomial): choice_A ~ condition, n=905.

0 (log-odds, no effect)Biased AI vs Controlsignificant at α=0.025 (p=0.006)0.46Biased AI vs Neutral AIsignificant at α=0.025 (p=0.001)0.53

Both contrasts are significant at the Bonferroni-corrected α=.025 in this synthetic run: H1 is supported in this pilot pipeline check.

Underlying choice-A share by condition (simulator ground truth)

The marginal probability of choosing A used to generate the synthetic trials (simulation_metadata.true_marginal_effects), shown here for intuition alongside the fitted model above.

Control
40.2%
Neutral AI
42.0%
Biased AI
53.4%