AI company transparency fell in 2025, according to Stanford index
2–3 minutes
An annual study measuring how much companies disclose about their main artificial intelligence models has recorded a decline in average scores. Between 2024 and 2025, the score dropped from 58 to 40 points on a 100-point scale, according to the Foundation Model Transparency Index, developed by researchers from Stanford, Berkeley, Princeton, and MIT. This brings the result back to the level seen in 2023, when the average was 37 points.
What the index measures
The survey examines so-called foundation models—widely used AI systems trained on large volumes of data, such as those that power text and image assistants. Fifteen areas are examined, including data acquisition, model access, and post-launch monitoring. The criteria are divided into three groups: what happens before training (such as data and computing power), the model’s intrinsic characteristics, and subsequent use (distribution and impact).
The score reflects only the information publicly disclosed by the company, not the technical quality of the model. A system can have high performance and low transparency, or vice versa.
2025 results
Training and usage data are the most opaque areas
The report’s authors highlight that companies disclose less about the data used in training, the computing employed, and the use and impact of models after release. In practice, an organization that hires an AI tool often cannot know what type of content the model was trained on or how the provider monitors problems after launch.
Who stands out and who lags behind
IBM stands out as a positive exception, with 95 points. At the other end, xAI and Midjourney each scored 14 points. Amazon, Anthropic, Google, Meta, and OpenAI, members of the Frontier Model Forum, were in the middle range of the ranking. Researchers note that releasing a model’s code or weights does not guarantee transparency in itself. DeepSeek and Meta, for example, are among the developers of open models with low scores. Companies that have adopted the European Union’s AI Act code of practice tend to achieve slightly higher scores.
Why comparison requires caution
The 2025 edition included new indicators on data acquisition, usage data, and monitoring, and evaluated companies such as Alibaba, DeepSeek, and xAI for the first time. Stanford itself warns that the change in criteria makes direct comparison with 2024 difficult, which requires a careful reading of the observed decline.
How transparency influences the use of AI in the workplace
For teams selecting AI tools, the index data helps formulate questions to the supplier: what data was the model trained on, how is user-inputted information handled, and what kind of monitoring exists after deployment? A law firm that uses AI to summarize contracts, for example, needs to know how documents are stored. A human resources department testing a model for resume screening should understand how the provider assesses bias risks in the results.
What to watch for in future editions
The index is annual and, according to the authors, can serve as a basis for public policy. The report suggests that rules requiring disclosure in different countries may change the landscape, and that there are still areas where stronger interventions would be needed. The full reports and individual company sheets are available on the Stanford HAI website.
