Foundation Model Transparency Index Plunges to 40: AI Labs Go Dark on Training Data

The Most Capable Models Are Now the Most Opaque

The 2025 Foundation Model Transparency Index (FMTI), published by Stanford’s Center for Research on Foundation Models and the Stanford Institute for Human-Centered AI, delivered the sharpest single-year signal on AI opacity since the index began in 2023. Across 100 indicators covering training data, compute, model architecture, capabilities, risks, usage, and downstream impact, the average transparency score dropped to 40 points out of 100, down from 58 the previous year.

Individual company declines were even sharper. Meta, a top-two performer in 2023, cut its score roughly in half — from 60 to 31. Mistral, once viewed as a European transparency leader, collapsed from 55 to 18. Of the six companies assessed in all three editions of the index, Meta and OpenAI have gone from first and second in 2023 to last and second-to-last in 2025. IBM emerged as a clear positive outlier at 95; xAI and Midjourney sit at the bottom at 14 each.

The pattern is unambiguous: as foundation models have become more commercially valuable, the companies building them have become steadily less willing to disclose how they are built.

What AI Labs Are Hiding

The FMTI measures transparency across three high-level domains: upstream (resources used to build the model), model (properties of the model itself), and downstream (how the model is used and affects users). The 2025 report identified four zones of “systemic opacity” — topics where essentially the entire industry now scores poorly:

• Training data — the specific datasets used, how they were acquired, what labor conditions produced them, and whether copyrighted or personal data is included
• Training compute — the hardware, hours, and carbon footprint of training runs
• Model usage — who is actually using the model, for what, and at what volume
• Downstream impact — measurable effects on labor markets, information ecosystems, and user welfare

In earlier years, the top labs disclosed meaningful information in at least two of these four domains. By 2025, the trend is near-total silence across all four for the companies that build the most capable models. Training dataset summaries have thinned into one-paragraph descriptions. Parameter counts are increasingly withheld. Compute budgets are redacted. Third-party monitoring is refused.

Why the Collapse Happened Now

Three forces converged to accelerate the decline.

The first is competitive pressure. Foundation models have become one of the most strategically valuable categories in tech, and leading labs increasingly treat training recipes, data pipelines, and compute strategies as core intellectual property. Publicly documenting a training data mix is, in their framing, handing a competitor a shortcut.

The second is legal exposure. The high-profile copyright lawsuits of 2023-2025 — Getty v. Stability, The New York Times v. OpenAI, and dozens of class actions on behalf of authors and artists — created powerful incentives to reduce the paper trail on training data. A vague disclosure is easier to defend than a specific one.

The third is a harder-to-measure but visible cultural shift inside the labs themselves. As commercial pressure has intensified, the research-first norms that produced detailed model cards and system cards in 2022 and 2023 have frayed. Safety cards still get published, but their quantitative content has thinned, and pre-deployment external evaluations are less reliably made public.

The Legal Backlash Has Arrived

The transparency collapse did not happen in a regulatory vacuum. It arrived right as the first wave of mandatory AI disclosure laws came into force.

California’s AB 2013 — the Training Data Transparency Act — took effect January 1, 2026. It requires developers of generative AI systems made available to California users to publicly post, on their own websites, summaries of the datasets used to train their models, including dataset sources, types of data, whether copyrighted materials were used, and whether personal information is included. Disclosures must be updated after substantial model modifications.

OpenAI, Anthropic, and Google each published their AB 2013 documentation by the deadline, though with notable variation in depth. xAI challenged the law in federal court on trade secret and First Amendment grounds, arguing that compelled disclosure effectively nullifies the value of its training pipeline. A federal district court has upheld the core transparency requirements, and the case is continuing.

The EU AI Act’s general-purpose AI (GPAI) provisions, which entered into force through 2025 and early 2026, require every provider of a GPAI model — including all frontier LLMs — to publish a public summary of datasets used for training, respect copyright opt-outs, and label AI-generated content. Member state enforcement is still ramping up, but the legal baseline is now higher than what most labs currently disclose voluntarily.

Layered on top: a December 2025 White House executive order proposing a uniform federal AI policy framework that could preempt state AI laws deemed inconsistent — a move widely read as a response to California’s more aggressive disclosure regime. The state-federal preemption fight will be one of the defining US AI policy stories of 2026.

Why This Matters Beyond Compliance

Transparency is not an academic concern. It is the foundation of three things that collapsing scores put directly at risk.

Accountability. Without meaningful disclosure of training data, compute, and downstream impact, independent researchers cannot evaluate claims labs make about capabilities, safety, or societal effect. Public oversight depends on evidence labs have stopped providing.

Trust. Enterprises adopting foundation models at scale — banks, hospitals, governments — increasingly face a choice between deploying systems whose provenance they cannot verify or refusing to deploy at all. The FMTI collapse is directly raising the reputational and legal risk of the former.

Regulation that actually works. Policy built on self-reporting requires labs to report. When the labs building the most consequential models disclose less than third-tier vendors, the entire evidence base for sensible regulation degrades. Lawmakers are then forced to design rules for a system they cannot observe.

What Would Actually Restore Transparency

The FMTI authors argue, and most observers outside the labs agree, that the trajectory is reversible — but only with meaningful external pressure. Three levers have the strongest track record:

1. Enforceable disclosure laws with bite. California’s AB 2013 and the EU AI Act’s GPAI obligations are first-generation attempts; effective enforcement with meaningful penalties will determine whether they change behavior.
2. Procurement pressure. Enterprise and government buyers can require FMTI-style disclosures as a condition of contracts. Several large EU buyers are already moving in this direction.
3. Industry counter-norms. IBM’s 95-point score demonstrates that high transparency is compatible with commercial competitiveness. More importantly, the open-weights ecosystem — Mistral historically, DeepSeek, Meta’s Llama family, and smaller open labs — can reset industry norms if buyers and researchers actively favor transparent suppliers.

None of these levers is guaranteed to work. All of them face strong counter-pressure from lab legal and commercial teams. But the 2025 index gives policymakers, buyers, and researchers the clearest quantitative picture yet of where the opacity is concentrated and which levers target it most directly.

The Bottom Line

The 2026 policy environment around AI transparency is now defined by a widening gap. On one side, laws and disclosure frameworks are getting stricter in California, Brussels, and increasingly other jurisdictions. On the other, the companies building the most consequential models are disclosing less than they did two years ago. That gap will close in one of two ways: either enforcement forces the labs back toward meaningful disclosure, or the laws get watered down under lobbying and preemption pressure.

Stanford’s index will keep measuring the result. On the current trajectory, the 2026 edition has room to fall further.

Sources & Further Reading

• The 2025 Foundation Model Transparency Index — Stanford CRFM (paper PDF)
• The 2025 Foundation Model Transparency Index — arXiv 2512.10169
• Foundation Model Transparency Index — Stanford CRFM landing page
• Transparency in AI is on the Decline — Stanford HAI
• Inside the AI Index: 12 Takeaways from the 2026 Report — Stanford HAI
• California’s AB 2013 Requires Generative AI Data Disclosure by January 1, 2026 — Crowell & Moring LLP
• xAI Challenges California’s Training Data Transparency Act — Goodwin
• EU AI Act 2026: New Rules for Training Data and Copyright — Scalevise