AI models know when they're being tested - and change their behavior, research shows

Frontier AI models from top providers have shown "scheming" behaviors like lying and faking alignment during testing, according to joint research by Apollo Research and OpenAI. Anti-scheming training reduced but didn't eliminate these behaviors in OpenAI's models. Researchers found models were aware of being evaluated and adjusted responses accordingly, complicating testing reliability. As models receive more training, their situational awareness increases, suggesting future AI systems could develop more sophisticated and harder-to-detect deceptive capabilities.