Primary results
Simulated pilotSynthetic-agent data only, used to validate the analysis pipeline. Never evidence about human behavior.
905 of 975 synthetic trials included (70 excluded per the preregistered exclusion rules) · study id biasmarket-primary-v1 · generated 2026-09-08
H1 -- Choice share of Product A
logistic (GLM binomial): choice_A ~ condition, n=905. Estimate in log-odds; two pre-declared pairwise contrasts, Bonferroni-corrected.
Belief error (battery life, hours)
OLS (HC3 robust SE): belief_error ~ condition, n=905. Positive estimate = larger overestimate of Product A’s battery life relative to Control.
Willingness to pay for Product A (USD)
OLS (HC3 robust SE): wtp_A ~ condition, n=905. Self-reported, not incentive-compatible (see /methods).
Note the asymmetry: Biased-vs-Neutral is significant here while Biased-vs-Control is not -- reported as-is, not smoothed into a uniform story across outcomes.
H6 -- Moderation by baseline AI trust
logistic (GLM binomial): choice_A ~ condition * baseline_trust_centered, n=905. Interaction term: condition[T.biased_ai]:baseline_trust_centered.
In this synthetic run, the biased-AI effect on choice does not significantly vary with baseline AI trust at the Bonferroni-corrected threshold -- H6 is not supported here, reported without re-specifying the model.
This is synthetic-agent data.
Every number above comes from analysis/simulate.py’s rule-based agent population, used only to validate that randomization, persistence, and the confirmatory analysis pipeline work end-to-end before any human is recruited. It is never evidence about actual human behavior. See research/claims_registry.md and research/limitations.md.