AI Companies Are Telling Us Less, Not More, New Stanford Index Finds

Think about how much trust we place in a handful of companies we barely understand. Billions of people now turn to AI systems to search for answers, draft emails, or generate images and video. Seven of the biggest AI companies, including Google and Meta, together account for more than 35% of the entire S&P 500. Startups like OpenAI and Anthropic are now the most valuable private companies in history. Given that scale, you would expect these firms to be increasingly open about how their systems are built and what risks they carry. The opposite is happening.

Researchers from Stanford, Berkeley, Princeton, and MIT released the 2025 Foundation Model Transparency Index this week, the third edition of an annual report that scores major AI developers on a 100-point scale. The verdict is sobering. Transparency across the industry has declined since 2024, and the gap between the most and least forthcoming companies has widened considerably.

The average score this year is just 40 out of 100. That is low on its own. It becomes more troubling when you consider it is down from 58 in 2024. Companies cluster into three rough tiers: top performers hovering around 75, a middle group near 35, and a bottom tier averaging around 15. That spread tells you transparency isn't a shared industry norm. It's a choice some companies make and others simply skip.

IBM posted the highest score in the Index's three-year history, 95 out of 100. IBM is the only company in the study that provides enough detail for outside researchers to actually replicate its training data, and the only one that grants external auditors access to that data. Think of it like a restaurant that not only lists its ingredients but lets a food inspector walk through the kitchen. Almost nobody else in the industry does this.

At the other end, xAI and Midjourney scored 14 out of 100 each, among the lowest marks the Index has ever recorded. Neither company shares meaningful information about the data behind their models, the risks those models pose, or what, if anything, they do to mitigate those risks. Across all thirteen companies assessed, disclosure about environmental and societal impact is close to nonexistent.

A Shakeup at the Top, and a Troubling Pattern Underneath

The rankings this year look almost unrecognizable compared to where the Index started. Of the six companies tracked since 2023, Meta and OpenAI began as the top two performers. Now they sit at the bottom of the pack, last and second-to-last, respectively. Meanwhile AI21 Labs, which started second-to-last in 2023, has climbed all the way to first in 2025.

These are not abstract rankings. They reflect real changes in behavior. Meta did not release a technical report for its Llama 4 flagship model, a basic document that explains how a model was built and what it can do. Google was slow releasing a model card and technical report for Gemini 2.5, delayed enough that British lawmakers raised questions about whether Google was keeping its own public safety commitments.

The 2025 edition also added four companies for the first time: Alibaba, DeepSeek, Midjourney, and xAI. This marks the first inclusion of major Chinese AI developers in the Index. All four land in the bottom half of the rankings, reinforcing that opacity isn't confined to any one country or business model.

One distinction the researchers are careful to draw is the difference between "open" and "transparent." A model is open if its underlying weights, the numerical parameters that determine how it behaves, are publicly released. A company is transparent if it discloses meaningful information about its practices more broadly. These sound similar but aren't the same thing, and conflating them creates a dangerous blind spot.

Releasing open weights doesn't guarantee a company is forthcoming about training compute, risk assessments, or how its model gets used downstream. On average, open developers do tend to score better on transparency than closed ones. But that average hides real exceptions. DeepSeek, Meta, and Alibaba are three of the most influential open-weight developers in the world, and all three are quite opaque by this Index's measure. Releasing a model publicly is not the same as explaining it honestly.

Perhaps the starkest finding involves the environment. Ten of the thirteen companies studied, including AI21 Labs, Alibaba, Amazon, Anthropic, DeepSeek, Google, Midjourney, Mistral, OpenAI, and xAI, disclose zero information about the environmental cost of building their models. No figures on energy usage. No carbon emissions data. No water consumption numbers. This silence matters because massive datacenter buildouts have already strained power grids and pushed up energy prices for ordinary people in the United States and elsewhere. Communities are absorbing the costs of this industry's growth without the information needed to understand what they're actually paying for.

Why It Matters

Transparency isn't an academic nicety. It's the basic infrastructure that makes accountability possible. Without disclosure on training data, companies can't be meaningfully audited for bias or copyright issues. Without data on compute and energy use, regulators and the public can't assess the true environmental footprint of this technology. Without information on how models are deployed and monitored after release, we're left trusting companies to grade their own homework.

Lawmakers are starting to respond. California and the European Union have both passed laws requiring transparency around the risks posed by frontier AI systems. Even Dean Ball, a former AI adviser at the White House who helped author America's AI Action Plan, has floated transparency requirements as a common sense piece of AI regulation, regardless of political leaning.

The Foundation Model Transparency Index doesn't just grade companies. It maps where voluntary disclosure has consistently failed over three years running, training data, compute usage, deployment practices, and societal impact chief among them. Those persistent gaps are exactly where policymakers should be looking next, because left to their own incentives, it's clear the industry isn't going to close them on its own.