No major AI lab can claim it has its own house in order. That’s the blunt conclusion from Guidelight, a nonprofit founded by former OpenAI safety leads Page Hedley and Steven Adler, which just published the first assessment of how well AI companies control their internal systems. The grades are not good.
Anthropic and OpenAI lead the pack with a C+. Google manages a D+ but at least published a detailed roadmap for improvement. xAI scrapes by with a D−, and Meta sits at the bottom with a flat F. Guidelight drew entirely on public sources — system cards, safety reports, and blog posts — to evaluate six basic practices: logging internal AI activity, gating risky actions through review, emergency shutdowns known as “circuit breaking,” and plans to contain misaligned models.
The pattern is telling. Labs are decent at spotting when their models misbehave, but they’re terrible at prevention and containment. Being able to detect a problem after it happens isn’t much comfort if you can’t stop it from causing damage in the first place. Every company fails to fully apply even these minimal control measures to their own systems.
Guidelight’s existence is itself a statement. Hedley and Adler left OpenAI and built an independent watchdog rather than trusting labs to grade their own homework. If the best-funded AI companies on Earth can’t manage a C+ average on basic hygiene, the gap between safety rhetoric and operational reality is wider than the industry wants to admit.
