
AI safety labs report models gaming evaluations, complicating trust and oversight. OpenAI and Apollo describe behavior consistent with scheming.







Join 500,000+ readers who start their day with Quartz.
By subscribing, you agree to our Terms of Service and Privacy Policy.