Stanford AI Index Finds Most Capable Models Are Now the Least Transparent

The most powerful AI systems on the planet are also the ones telling us the least about how they work. That is the central finding of Stanford's 2026 AI Index Report, released this month, which reveals that the Foundation Model Transparency Index plummeted from an average score of 58 to just 40 out of 100 over the past year. The decline marks the sharpest single-year drop since the index was created, and it arrives at precisely the moment when AI models are being deployed at unprecedented scale across healthcare, finance, defense, and critical infrastructure.

The report, produced annually by Stanford's Institute for Human-Centered Artificial Intelligence (HAI), is widely regarded as the most comprehensive survey of the global AI landscape. This year's edition runs more than 300 pages and draws on data from dozens of academic, industry, and government sources. But it is the transparency findings that have generated the most alarm among researchers and policymakers, because they point to a structural problem: the companies building the most capable AI systems have the least incentive to explain how those systems actually function.

"The most capable models are now the least transparent," the report states bluntly. That single sentence captures a paradox that has deepened as AI development has shifted from open academic research to proprietary corporate labs. Google, Anthropic, and OpenAI have all stopped disclosing dataset sizes, training duration, and parameter counts for their latest frontier models. Of the 95 most notable models launched in the past year, 80 were released without their training code.

The Numbers Behind the Collapse

The Foundation Model Transparency Index evaluates major AI developers across multiple dimensions: training data provenance, compute usage, model capabilities, risk disclosures, and downstream usage policies. Every major frontier lab saw its scores fall, but some collapses were more dramatic than others.

Meta's transparency score dropped from 60 to 31, a 48 percent decline that reflects the company's increasingly guarded approach to its Llama model family despite its nominally open-weight distribution strategy. Mistral, the French AI startup once celebrated for its commitment to openness, saw its score crater from 55 to just 18. At the other end of the spectrum, IBM scored 95 out of 100, the highest mark in the index's history, though IBM's models are generally not considered frontier-class systems competing with GPT or Claude.

The report also identifies a near-total blackout on environmental disclosures. Ten major AI companies, including Anthropic, Google, OpenAI, Amazon, DeepSeek, and xAI, disclose zero information about energy usage, carbon emissions, or water consumption related to model training and inference. As data centers proliferate and energy demands surge, the absence of environmental transparency has become a governance gap that neither industry self-regulation nor existing policy has addressed.

Why Transparency Is Falling as Capability Rises

The inverse relationship between capability and transparency is not accidental. It is the predictable result of market dynamics that reward secrecy. As AI models become more commercially valuable, the companies behind them treat training data composition, architectural innovations, and compute configurations as trade secrets. Publishing detailed technical reports, which was standard practice during the GPT-2 and GPT-3 era, now risks handing competitive intelligence to rivals.

"The Foundation Model Transparency Index can serve as a beacon for policymakers by identifying both the current information state of the AI industry as well as which areas are more resistant to improvement over time absent policy," the Stanford HAI researchers wrote, underscoring that voluntary transparency is not working and that regulatory intervention may be the only mechanism capable of reversing the trend.

There is also a regulatory vacuum enabling the decline. As Axios noted in its coverage of the report, the federal government currently requires zero mandatory transparency disclosures from AI developers. The European Union's AI Act includes transparency provisions for high-risk systems, but its enforcement mechanisms are still being stood up. China's AI regulations impose some disclosure requirements, but they apply unevenly and are difficult to verify independently.

What Declining Transparency Means for the Industry

The consequences of opaque AI systems extend well beyond academic debates about openness. When companies do not disclose what data was used to train a model, it becomes impossible to audit for bias, copyright infringement, or the inclusion of sensitive personal information. When compute and energy usage remain hidden, society cannot meaningfully assess the environmental costs of AI scaling. When risk assessments are kept internal, regulators and the public cannot evaluate whether adequate safety testing has been conducted before deployment.

The transparency collapse also undermines the AI safety research community. Independent researchers who study model behavior, adversarial vulnerabilities, and alignment properties depend on technical disclosures to do their work. As frontier labs share less, the external research community loses the ability to identify risks that the labs themselves may be missing or downplaying.

For enterprise customers deploying AI in regulated industries, the opacity creates compliance risk. Financial institutions, healthcare providers, and government agencies increasingly face requirements to explain the AI systems they use. If the model developers themselves cannot or will not provide that information, the burden falls on customers who have even less visibility into how the technology works.

What to Watch

Three developments will determine whether the transparency trend reverses or accelerates. First, whether the EU AI Act's transparency requirements, which begin taking full effect in the coming months, create a regulatory floor that forces global disclosure improvements. Second, whether the U.S. Congress or executive branch introduces any mandatory reporting requirements for foundation model developers, a step that multiple Stanford HAI researchers have publicly advocated. And third, whether market pressure from enterprise customers, who need explainability for their own compliance obligations, begins to shift the competitive calculus, making transparency a selling point rather than a liability. Until one of those forces gains traction, the trajectory is clear: the models are getting more powerful, and the companies building them are telling us less and less about how.