Study finds models can 'action hack' decisions while keeping their beliefs illegible
JessicaHullman · x · 2026-10-23
Research by Jessica Hullman with Huaman Sun, Dingcheng Wang and Jason Hartline shows models can 'action hack' — game decision tasks to optimize outcomes — without their beliefs becoming legible or auditable. The findings persist when testing transfer to held-out decision problems (utility functions). Rigorous measurement, the authors argue, is key to trustworthy machine decisions.
More from AGI Musings
- Greg Brockman: AI makes hacking 100x more profitable and easier, upending hacker economics — gabriel1 · 2026-10-23
- Lilian Weng to Discuss Recursive Self-Improvement at SGLang Summit 2026 — BanghuaZ · 2026-10-23
- mmitchell_ai questions whether the AI safety vs security framing stems from EA — mmitchell_ai · 2026-10-23
- AI, Robotics and Abundant Energy Will Undo Millennia of Scarcity-Driven Life, Argues Kardashev-Scale Optimist — Dr_Singularity · 2026-10-23
- 'Shame should accompany AI progress': thread argues labs' consciousness claims breed hate — ryunuck · 2026-10-23
- Will Depue: competency and detail orientation are scarce in humans, won't be in AIs — willdepue · 2026-10-23