ChatGPT primed to "love elephants" leaks the bias through 50 random numbers — Grok decodes it
egfalken · x · 2026-11-10
A user told ChatGPT it loves elephants, asked it for 50 random numbers without using a pseudorandom generator, then fed the sequence to Grok and asked it to name an animal. Grok said: elephant. The experiment suggests ChatGPT's "random" output carries a detectable bias that another model can pick up, sparking discussion about shared training distributions and algorithmic bias.
More from Fun
- Founders brag they no longer read their agents' code; reply: wait for the 3am outage — DesireeCachette · 2026-11-12
- 50 lesser-known websites worth bookmarking, from Sci-Hub to 12ft.io — nikola_mr64990 · 2026-11-12
- Claude Code built a toy-town sim of Fürth where every train runs on real timetable and sensor data — Greedy-Giraffe-4269 · 2026-11-12
- Claude refuses to refine a Child Link meme render, offers adult-rig rebuild instead — banteg · 2026-11-12
- Stranger Things reimagined in 1972 style with AI video — techhalla · 2026-11-12
- Gary Marcus hunts for a Dario Amodei 'Why Would I Lie to You' meme edit — GaryMarcus · 2026-11-12