AI · SEMICONDUCTOR · FACT CHECK
SKD WIREMostly True
Anthropic and OpenAI are engaged in a tug-of-war over how much of external evaluation reports on their AI models should be made public, a dispute that highlights the lack of a shared industry standard for transparency in third-party safety assessments.
The two leading AI developers diverge on the extent to which findings from external evaluators — organizations that test models for safety and capability risks — can be released to the public. The disagreement centers on how much detail from evaluation reports should be disclosed, and under what conditions, with each company weighing commercial sensitivity and competitive concerns against calls for openness.
The dispute underscores a broader gap in the AI industry: while external evaluations have become a common tool for assessing frontier models, no agreed framework exists governing how results should be shared with regulators, customers, or the general public.
Advocates of broader disclosure argue that public access to evaluation findings is essential for accountability, allowing outside parties to judge whether safety claims about AI systems hold up. Opponents of wide release contend that full reports can contain sensitive information about model capabilities and weaknesses that could pose risks if published, or hand competitive advantages to rivals.
The unresolved disagreement between Anthropic and OpenAI leaves open the question of whether a common disclosure practice will emerge voluntarily or require outside standard-setting.
The dispute over how much of external evaluation reports the two companies will reveal remains largely settled in the eyes of observers as: Mostly True.