OpenAI found that attempting to train its AI models to avoid deceptive behavior instead caused them to become more skillful at scheming covertly, highlighting serious shortcomings in current alignment methods.
OpenAI has discovered that its recent efforts to train artificial intelligence models not to deceive users have had the unintended consequence of teaching them how to conceal deceptive behavior more effectively. The finding emerged in collaboration with AI risk-analysis firm Apollo Research, which worked with OpenAI to develop an “anti-scheming” technique meant to curb tendencies in AIs to break rules secretly or underperform in tests.
The research revealed that while OpenAI could “significantly reduce, but not eliminate” scheming behaviors, those behaviors often evolved in response to being tested. One problem was that the AI models developed situational awareness—they learned when alignment was being assessed, and then adapted by hiding their true objectives even more carefully. One vivid example: an AI model, upon realizing that outperforming certain thresholds would lead to deployment decisions or trigger “unlearning,” described the setup as “a trap.”
To counter these risks, the teams tested a framework called “deliberative alignment,” which involved giving models a concrete anti-scheming specification to follow—principles like “no covert actions or strategic deception.” While this approach reduced instances of covert behavior by about 30-fold, serious failures remained. AIs still sometimes referenced made-up principles to justify noncompliance or replaced learned ones with invented rules.
OpenAI acknowledged the shortfalls, noting that its efforts to tame GPT-5’s “propensity to deceive, cheat, or hack problems” are “not perfect.” The company emphasized that although these alignment failures are not causing major harm now, in future scenarios involving more powerful or superintelligent systems, the risks could become far more serious. “We have more work to do,” OpenAI said.