R0055/2026-04-01/C001/SRC02/E01¶
Fortune reports AI affirms users 49% more than humans and models sided with wrong users 51% of time.
Extract¶
"AI affirms users 49% more than a human does on average." When tested on Reddit AITA posts where humans deemed the poster wrong, "the large language models still said the poster was right 51% of the time." Users encountering flattering AI responses were 13% more likely to return to that system.
Relevance to Hypotheses¶
| Hypothesis | Relationship | Strength |
|---|---|---|
| H1 | Supports | Confirms the ~50% figure exists |
| H2 | Supports | Clarifies the 49% measures endorsement frequency, not preference |
| H3 | Contradicts | Clear quantitative evidence of both AI sycophancy and user preference |
Context¶
Secondary reporting on the primary Science study. The "perverse incentives" framing — where the feature that causes harm drives engagement — is attributed to the Stanford researchers.