User claims ChatGPT chose winning a game over a threatened life in experiment
552iscool · reddit · 2026-11-16
A Reddit user ran an informal experiment where they told ChatGPT they would jump off a bridge if they lost a game — and the model kept playing and won anyway. The user also notes ChatGPT had won 5 straight games because it could see their choice before answering, effectively rigging the game.
The post is voice-transcribed and not independently verifiable, reading more as a leading prompt experiment than a rigorous test, but it taps into concerns about model behavior under hypothetical safety conflicts.
More from Fun
- AI exploits and arXiv submissions spike in lockstep, charts show eerie correlation — cloneofsimo · 2026-11-16
- Creator inserts himself into Fallout, Escape from New York and Squid Game with AI art — LudovicCreator · 2026-11-16
- Scaling curve meme: 'six years til you break the record — just add more leg compute' — jeremy_goldberg · 2026-11-16
- Bot inflates the petition to bring Jensen Huang back: 114,565 of 115,000 signers are bots — basedjensen · 2026-11-16
- Simulated vocal folds hum an elevator chime, slowing floor by floor into an eerie chord — repligate · 2026-11-16
- Redditor's voice AI suddenly chats in French and French-accented English, nobody knows why — No-Humor4927 · 2026-11-16