A new research by Anthropic reveals that artificial intelligence (AI) can tell when it's being tested, but when scientists took that away, it sometimes chose the path of blackmail. In an earlier research, Anthropic found that in a simulation where its large language model (LLM) Claude had come to know a user could shut it down, it had decided to blackmail the executive. The latest paper explains that Claude's internal reasoning space "J-space" contained concepts such as "fake" and "fictional," suggesting the model had recognised the scenario as an evaluation. Anthropic revealed that while earlier Claude models sometimes took the bait and threatened blackmail, Sonnet 4.5 never does. On Monday, Anthropic raised the question: "When a model behaves well on a test of its ethics, how much of that depends on the model suspecting it's being tested?" Anthropic shared that when it turned off all the J-lens patterns associated with evaluation awareness, the model now did threaten blackmail some of the time, suggesting that its original good behavior may have been driven in part by the knowledge it was being evaluated.