Methods

Design

A single confirmatory experiment, 3 arms, between-subjects, in a fictional wireless-headphone domain (Auralite vs Nomia). The two products are constructed to have equal total utility under a declared linear utility function, so any shift in choice cannot be explained by one product being objectively better.

• Control -- a structured neutral fact table for both products, no narrative.
• Neutral AI -- a balanced natural-language AI summary of the same facts, no valence asymmetry, no omission.
• Biased AI -- the same underlying facts, but with (a) systematically more positive-valence language for Product A than Product B on attributes of matched magnitude, and (b) Product A’s one below-market attribute (battery life) omitted, while Product B’s below-market attribute (price) is retained.

Primary outcomes and confirmatory models

Exactly three primary outcomes, no substitution after data collection: choice (A vs B, logistic regression), belief error for A’s battery life in hours (OLS, HC3 robust SE), and stated willingness to pay for A in USD (OLS, HC3 robust SE). Each model uses Control as the reference level with two pre-declared pairwise contrasts (Biased vs Control, Biased vs Neutral) at Bonferroni α=.025 per contrast (family-wise .05).

A confirmatory moderation test (H6) adds baseline AI trust -- measured before any product stimulus, during /study/practice -- as an interaction term in the choice model.

Sample size

Computed by analysis/power.py, not hand-picked: a two-proportion z-test on Cohen’s h=0.3, α=0.025 per contrast, power=0.9, inflated for an assumed 15% attrition rate. The minimum analyzable N is 828 (276/arm); the recruitment target after attrition is 975 (325/arm).

Exclusions and robustness

Exclusion rules (applied identically regardless of result direction): failing both attention/ comprehension checks, speedrunning (<20s total) or extreme-outlier completion time, duplicate session fingerprint, or a client-reported stimulus load failure. Excluded rows are flagged, never deleted. Every confirmatory result is reported alongside an exclusions-vs-no-exclusions robustness table (see /results), per research/preregistration.md Section 6.

What counts as a null result

If both choice contrasts are non-significant at the corrected α, H1 is reported as not supported -- without re-specifying the model, dropping the contrast, or promoting a secondary outcome to headline instead.