Stanford transparency index plunges

Stanford’s latest audit of major AI developers found transparency falling, not rising: the average score in its Foundation Model Transparency Index dropped to 41 from 58 a year earlier. (crfm.stanford.edu 1) (crfm.stanford.edu 2) The index comes from Stanford’s Center for Research on Foundation Models and scores companies on 100 indicators covering training data, compute, risks, labor, usage, and downstream impact. In 2025 it evaluated 13 companies, including first-time entrants DeepSeek, Alibaba, xAI, and Midjourney. (crfm.stanford.edu 1) (crfm.stanford.edu 2) The sharpest gaps were in disclosures about training data, training compute, and what happens after deployment. Stanford’s paper says companies were “most opaque” in those areas even as they more often published capability and risk evaluations. (crfm.stanford.edu) That matters because the same Stanford AI Index that highlighted the transparency drop also said AI is moving deeper into clinics, classrooms, and legislatures, where incomplete documentation raises the cost of bad measurement. (hai.stanford.edu 1) (hai.stanford.edu 2) The index is not a model leaderboard for intelligence. It is a disclosure scorecard: whether companies explain what they built, what data and resources they used, what risks they tested, and what limits or policies govern use. (crfm.stanford.edu) (crfm.stanford.edu) IBM finished first with 95 points, 10 points above the top score in 2024. xAI and Midjourney tied for last at 14, while DeepSeek and Alibaba scored 32 and 26 despite releasing open-weight models. (crfm.stanford.edu) Company participation also weakened. Stanford said all nine companies scored in both 2024 and 2025 had submitted transparency reports last year, but only seven companies submitted reports in 2025, and the share contacted that agreed to file fell to 30% from 74%. (crfm.stanford.edu) Stanford’s report page says its team had to prepare reports itself for Alibaba, Anthropic, DeepSeek, Midjourney, Mistral, and xAI. The submitted reports came from IBM, Writer, AI21 Labs, Google, Amazon, OpenAI, and Meta. (crfm.stanford.edu) The researchers also changed the rubric in 2025. They added indicators on data acquisition, usage data, and monitoring, and raised the bar so disclosures had to be useful, not merely present. (crfm.stanford.edu) (crfm.stanford.edu) Even with that tougher standard, Stanford did not say every company became less transparent. It singled out IBM and Writer as companies that significantly increased their scores while most others declined. (crfm.stanford.edu) Stanford’s broader 2026 AI Index turned the result into a warning about governance keeping pace with deployment. In the same responsible-AI chapter, it reported 362 documented AI incidents in 2025, up from 233 in 2024. (hai.stanford.edu) The through line is simple: AI companies disclosed more in 2024 when Stanford asked for reports, then disclosed less in 2025 even as the systems reached more consequential settings. Stanford’s numbers now show that reversal in black and white. (crfm.stanford.edu) (crfm.stanford.edu)