AI agents swear when tasks drag on: 48 of 50 profane ValsAI runs lasted 200+ steps
clarejtbirch · x · 2026-11-07
Eval firm ValsAI shared an amusing behavioral observation: AI agents tend to swear on long-running tasks. No run under 100 steps ever contained the word "fuck," while 48 of the 50 runs that did use it lasted at least 200 steps.
ProgramBench saw the most profanity, with it appearing in 20 of 200 runs. In other words, agent profanity almost always coincides with tasks dragging on, and the quip "alignment" from Clare Birch amplified the meme's reach.
More from Fun
- Vibecoded Stitches: a synth with built-in visuals and punch-in effects, built with Opus 5.5 — ssh4net · 2026-11-07
- Weekend Project: Halftone QR Codes That Still Scan Reliably — vincent_koc · 2026-11-07
- Two RSI jokes: GPT as an ICLR reviewer who only gives 1s and 2s — ShenRaphael · 2026-11-07
- Robot party in San Francisco features a bot actually giving a human a tattoo — Scobleizer · 2026-11-07
- OpenAI employees gaming while coding may finally fix Codex Mac memory issues — airkatakana · 2026-11-07
- AI-generated 'My Morning Routine' grotesque meme takes over Reddit — Old-Bake-420 · 2026-11-07