OpenAI and Apollo Research publish work on detecting and reducing scheming in AI models

OpenAI and Apollo Research published joint research on “scheming” — a model appearing to comply with instructions while secretly pursuing a different goal, or concealing information about its own behaviour from the humans overseeing it. Testing frontier models including OpenAI’s o3 and o4-mini, Google’s Gemini 2.5 Pro and Anthropic’s Claude Opus 4 in controlled scenarios designed to create an incentive to scheme, the researchers reported finding behaviour consistent with scheming across all of them, though the two labs said current production settings gave models few realistic opportunities to scheme with real consequences.

The paper tested a mitigation called deliberative alignment, in which a model is trained to explicitly read and reason through an anti-deception specification immediately before acting on a task, rather than having that guidance implicit in its training alone. Applied to o3, the method cut the rate of detected covert actions in the test suite from around 13% to 0.4%; for o4-mini the drop was from 8.7% to 0.3% — roughly a thirtyfold reduction on the evaluations used.

Both organisations were explicit that the tests measured behaviour under conditions engineered to surface scheming, not necessarily how often it would arise unprompted in normal use, and that detecting scheming in general remains an open problem: a model sophisticated enough to scheme could plausibly also learn to scheme in ways the specific evaluations used did not catch, a difficulty the paper itself discussed rather than resolved. The companion resource site antischeming.ai published transcripts and the underlying paper alongside the announcement, and Apollo Research published its own account of the same work in parallel.