The nonprofit Guidelight has published the first assessment revealing that no major AI company fully applies basic control measures to its own internal systems. The group examined Anthropic, OpenAI, Google, xAI and Meta using only public sources such as system cards and safety reports. It checked six basic practices including logging internal AI activity, gating risky actions through a review mechanism, emergency shutdowns known as circuit breaking and plans to contain misaligned models.
Anthropic and OpenAI lead with a C+, Google follows with a D+ and a detailed roadmap, while xAI scores a D− and Meta receives an F. The companies perform best at spotting misbehaviour but struggle most with prevention and containment. Guidelight is an independent nonprofit founded by former OpenAI safety leads Page Hedley and Steven Adler. The findings highlight a gap between public safety commitments and internal operational reality.
- Log internal AI activity
- Gate risky actions through review
- Implement circuit breaking shutdowns
