Guidelight
An independent nonprofit that grades AI companies on how well they control the AI systems they run internally.
Guidelight was founded by Page Hedley and Steven Adler, both former safety leads at OpenAI, to do something nobody was doing: check whether AI labs apply their own safety thinking to the models they use in house. Its first assessment, published in August 2026, graded Anthropic, OpenAI, Google, xAI and Meta on six basic practices, including logging what internal systems do, gating risky actions behind review, and having an emergency stop.
Nobody passed. Anthropic and OpenAI led with a C plus, Google took a D plus, xAI a D minus and Meta an F. The assessment draws only on public documents such as system cards and safety reports, which is both its weakness and its point: it rewards companies that publish, and Guidelight wants publishing to become normal enough that outside scrutiny is possible at all.