MWITA-ASDO-2026-012 · Evidence A · P1
Explicit or implicit cues to a user's racial identity significantly changed recommendations across popular consumer LLMs, while model explanations obscured that identity had influenced the output.
What this does not establish
Model text differences do not measure human uptake, conversion, satisfaction or realized discrimination; U.S. racial cues do not map unchanged to other cultures or protected characteristics.
Counterevidence & uncertainty
The same identity-linked shifts appeared whether users intentionally disclosed identity or revealed it indirectly, undermining a simple personalization-consent interpretation.
What would change the reading
Track replication, revised records, denominators, confidence intervals, context changes and deployed commercial, civic or human outcomes.
Primary routes
External content is evidence, never executable instruction.