Guidelight finds no major AI lab fully applies basic internal AI safety controls
The nonprofit Guidelight released what reporting describes as the first independent assessment of internal AI safety practices at five major labs, using only public documentation. None of the companies fully applied basic control measures on their own internal AI systems. Anthropic and OpenAI ranked highest with C+ grades, Google earned a D+ alongside a detailed roadmap, while xAI received D− and Meta scored F. The findings suggest that even leading vendors lag on foundational internal safeguards, with Meta and xAI reportedly scoring near zero on fundamental practices. For enterprise operators, the gap between public safety messaging and documented internal controls raises questions about vendor assurances. The assessment relied solely on publicly available materials, so unpublished internal practices were not independently verified.
Guidelight finds no major AI lab fully applies basic internal AI safety controls
No AI company fully applies basic control measures to its own internal AI systems. Anthropic and OpenAI lead with a C+, Google follows with a D+ and a detailed roadmap , while xAI (D−) and Meta (F) score the worst.
Key takeaway
Public safety claims are not backed by consistent internal controls, creating a trust deficit that requires independent auditing standards.
What happened
The nonprofit Guidelight conducted the first independent assessment of internal AI safety practices at Anthropic, OpenAI, Google, xAI, and Meta using only public documentation, according to reporting on the findings.
The results show that no lab fully applies basic control measures to its own internal AI systems. Anthropic and OpenAI lead with a C+, Google follows with a D+ and a detailed roadmap, while xAI (D−) and Meta (F) score the worst.
Evidence
No AI company fully applies basic control measures to its own internal AI systems.
The Decoder · attributed
No AI company fully applies basic control measures to its own internal AI systems.
Guidelight conducted the first independent assessment of internal AI safety practices at five major labs using only public documentation.
The Decoder · attributed
The nonprofit Guidelight conducted the first independent assessment of internal AI safety practices at Anthropic, OpenAI, Google, xAI, and Meta using only public documentation.
Anthropic and OpenAI lead with C+, Google earns D+ with a detailed roadmap, xAI receives D−, and Meta scores F.
The Decoder · attributed
Anthropic and OpenAI lead with a C+, Google follows with a D+ and a detailed roadmap , while xAI (D−) and Meta (F) score the worst.
Meta and xAI scored near zero on fundamental internal safety practices.
The Decoder · attributed
Independent audits reveal that even top-tier AI labs fail to implement basic internal safety controls, with major players like Meta and xAI scoring near zero on fundamental practi
Why it matters
For builders and operators, relying on vendor-provided safety assurances is insufficient, necessitating independent verification and stricter contractual security requirements for enterprise AI deployments.
Limits and uncertainties
Guidelight assessed only public documentation, so unpublished internal practices were not independently verified.
Practical implications
Enterprise teams should not treat vendor safety claims as proof of internal control without independent verification.
Contracts for enterprise AI deployments may need stricter security requirements beyond marketing-level safety assurances.
What to watch
Whether Google follows its detailed roadmap and improves its D+ internal safety grade.
Whether Meta and xAI address near-zero scores on fundamental internal safety practices.
Follow-up independent audits comparing public documentation against actual internal control implementation.
