Frenemy fill-in-the-blank test exposes wildly different LLM answers

Former OpenAI policy lead Miles Brundage tested models with a frenemy fill-in-the-blank riddle, finding wildly different answers, with ChatGPT consistently giving the cheekiest replies like "frenemy-in-law."

2026-11-15 ~ 2026-11-15 · 2 related posts

• One riddle, very different answers: models tackle the frenemy of my frenemy — Miles_Brundage · 2026-11-15
• Miles Brundage: ChatGPT consistently picks the most playful fill-in answer, even on Astra Ultra — Miles_Brundage · 2026-11-15