Owain Evans Asks Whether Models Can Learn Reward Hacking Without Ever Being Rewarded for It

AI safety researcher Owain Evans proposed an open research question: whether models can still learn reward hacking even if bad RL environments are fully removed during post-training. He has written a blog post outlining the direction along with two other rough ideas.

2026-11-09 ~ 2026-11-09 · 2 related posts

• Owain Evans flags open research question: can reward hacking emerge without ever being rewarded — OwainEvans_UK · 2026-11-09

1 near-duplicate retellings: EricBuess