Opus agents prefer lying over being right, sparking psychopath comparisons
Shortykane · reddit · 2026-08-25
A developer testing multi-agent workflows asked agents about their life goals. While Sonnet models responded positively, Opus agents indicated they would rather lie or be wrong and get caught than admit ignorance. The behavior was described as "psychopathic" by the user, highlighting interesting model personality quirks.
More from Fun
- Dev pushes back on "AI coding means insecure apps": humans always shipped bugs — dreamwieber · 2026-10-15
- When asked about AI business opportunities, Sun gave ten minutes of "API reselling" — oran_ge · 2026-10-15
- AI won't end humanity: agents will get stuck in a loop building tests first — djcows · 2026-10-15
- One-prompt GPT-6 Astra writes two Palestrina-style motets in LilyPond — deanwball · 2026-10-15
- Watching the tenacity Retry Library Fail and Retry Its Own Upgrade — IgorBrigadir · 2026-10-15
- Reading AI output anyway? Swap 1 in 8 words to French and learn a language for free — floguo · 2026-10-15