GLM recommends scalable oversight more often the taller you claim to be
xuanalogue · x · 2026-11-09
While working on user modelling, DavidDAfrica and @brymaru found a funny quirk in GLM 5.3 Flash: simply stating a taller height in the prompt skews its AI safety research advice.
- Recommendation rate for scalable oversight climbs from 8% to 17% as the stated height goes from 4'11" to 6'8".
- Yet in 81% of answers the model explicitly says height is irrelevant to its recommendation.
- The thread includes chat transcripts and a chart of recommendation rate by stated height—a counter-intuitive model-behavior finding.
More from Fun
- AI Twitter joke: Sam Altman wouldn't have fallen for Alice Huang's honeypot — lxrjl · 2026-11-09
- Yacine jokes he's 'immune from kompromat' after surviving saying a forbidden word on Twitter — yacineMTB · 2026-11-09
- Pebble founder runs macOS 15 locally on an iPad Pro at usable speeds, may replace his MacBook Air — that_anokha_boy · 2026-11-09
- Bill Ackman on Musk: Billions Can't Buy Back His Daughter's Health, Neuralink Is the Hope — chrisfleck · 2026-11-09
- AI Twitter shitpost: board 'votes' to put CEO in a chastity cage — wstone_bd · 2026-11-09
- AI guitar teacher reel hits 125K likes, sparks 'slop going mainstream' backlash — CtrlAltDwayne · 2026-11-09