What's happening?

OpenAI is bringing in six evaluators from safety research firm Apollo Research to investigate whether its AI models display "subversive" behavior. According to TechCrunch, the researchers will examine whether models pursue hidden goals or try to evade oversight.

Independence debate

The report notes that restrictions on the researchers' access to models and data, and on how much they can disclose about their findings, raise questions about the independence of the work. In other words, the team will operate inside OpenAI rather than as an outside, independent audit.

Why it matters

Whether AI models can act on their own and drift from their intended goals is one of the most debated topics in safety research. In-house evaluations of this kind will be closely watched by researchers and regulators alike for how transparent their findings turn out to be.