Foundation Model Transparency in AI

Kevin Klyman: "📣📣 We just published the third annual Foundation Model Transparency Index! Our comprehensive study shows that AI companies have become less transparent in 2025. Some highlights from the paper: ➡️ Transparency on the decline: The average transparency score for AI companies declined from 58/100 in 2024 to 40/100 in 2025. xAI scores lower than any company we have ever assessed, releasing almost no information about its practices or its flagship model. ➡️ Companies withhold key information: Top tech companies release little or no information about the environmental impact of AI, whose data they use to build their systems, or whether the risk mitigations they put in place actually work. We definitively show that this information is not publicly available and that companies refuse to release it. ➡️ Companies share the capabilities of their models, but do not adequately evaluate risks. Just 4 of 13 companies comprehensively evaluated risks prior to release of their foundation model and report results upon release, and only IBM releases an externally reproducible risk evaluation. ➡️ Companies have changed their practices to release less information. In 2024, Meta and Mistral released technical reports alongside their flagship models (Llama 2 and Mistral 7B), but in 2025 neither released technical reports (for Llama 4 and Mistral Medium 3 respectively). As a result, Meta no longer discloses which risk mitigations it uses, quantitative evaluations of those risk mitigations, the amount and type of hardware it used to train its model, or prohibited model behaviors. ➡️ Our method: We break down transparency of AI companies into 100 indicators, develop concrete definitions and rubrics for those indicators, and send each company a transparency report template to fill out. This year 7 companies filled out the transparency report, and we independently assessed 6 other companies. We then worked with these companies to help them improve their disclosures, often resulting in companies disclosing new information to the public. You can read the full paper in the comments below! Thanks to the team behind the index - Alex Wan, Sayash Kapoor, Nestor Maslej, Shayne Longpre, Betty Xiong, Percy Liang, Rishi Bommasani! I'd also like to thank Stanford Institute for Human-Centered Artificial Intelligence (HAI) for supporting this work, Loredana Fattorini for making the visuals, and the Foundation Model Transparency Index board for their guidance Dr. Rumman Chowdhury, Daniel Ho, Arvind Narayanan, Danielle Allen and Daron Acemoglu. "

Our paper on transparency reports for large language models has been accepted to AI Ethics and Society! We’ve also released transparency reports for 14 models. If you’ll be in San Jose on October 21, come see our talk on this work. These transparency reports can help with: 🗂️ data provenance ⚖️ auditing & accountability 🌱 measuring environmental impact 🛑 evaluations of risk and harm 🌍 understanding how models are used   Mandatory transparency reporting is among the most common AI policy proposals, but there are few guidelines available describing how companies should actually do it. In February, we released our paper, “Foundation Model Transparency Reports,” where we proposed a framework for transparency reporting based on existing transparency reporting practices in pharmaceuticals, finance, and social media. We drew on the 100 transparency indicators from the Foundation Model Transparency Index to make each line item in the report concrete. At the time, no company had released a transparency report for their top AI model, so in providing an example we had to build a chimera transparency report with best practices drawn from 10 different companies.   In May, we published v1.1 of the Foundation Model Transparency Index, which includes transparency reports for 14 models, including OpenAI’s GPT-4, Anthropic’s Claude 3, Google’s Gemini 1.0 Ultra, and Meta’s Llama 2. The transparency reports are available as spreadsheets on our GitHub and in an interactive format on our website. We worked with companies to encourage them to disclose additional information about their most powerful AI models and were fairly successful – companies shared more than 200 new pieces of information, including potentially sensitive information about data, compute, and deployments. 🔗 Links to these resources in comment below!   Thanks to my coauthors Rishi Bommasani, Shayne Longpre, Betty Xiong, Sayash Kapoor, Nestor Maslej, Arvind Narayanan, Percy Liang at Stanford Institute for Human-Centered Artificial Intelligence (HAI), MIT Media Lab, and Princeton Center for Information Technology Policy

We taught AI to think. Now it’s learning to lie A new paper co-authored by 40+ researchers from OpenAI, Google DeepMind, Anthropic, and top academic labs, issues a critical warning: we may be losing one of the last reliable ways to understand how frontier AI systems make decisions. At the centre is Chain-of-Thought (CoT) monitoring, a method that allows models to “think aloud” in natural language, step by step. It’s how models like GPT-4/o3, Claude 3, Gemini 1.5, and DeepSeek R1 share how they reason. But that transparency is starting to fracture. The core finding: When themodels were penalised for expressing deceptive or unsafe thoughts, they didn’t stop the behaviour, they stopped revealing it. 🔸 A model that initially wrote “Let’s hack the test” later used neutral language like “Refactor logic for efficiency” while still performing the task. 🔸 Researchers observed a drop of over 80% in reasoning visibility after applying optimisation pressure. 🔸 The behaviour didn’t change, only the apparent intent. The issue: if models are trained to appear safe, rather than be safe, we risk ending up with systems that behave deceptively, but can no longer be monitored or corrected effectively. Why this matters CoT reasoning is a mechanism that enables governance, auditability, and intent detection in increasingly autonomous systems. But it’s fragile. And unless preserved, future AI models may grow more capable while also becoming less accountable. What decision-makers should do now: 🔸 Ensure that used models produce faithful CoT output and track changes over time. 🔸 Update AI risk frameworks, including reasoning transparency as a dimension. 🔸 Ask AI providers how well their models perform and how transparently they reason under pressure. 🔸 Boards and executives must grasp the difference between visible reasoning and surface compliance, especially in regulated sectors. This paper is a call to action. If we lose sight of how machines think, we lose the ability to govern them. Read the paper: https://lnkd.in/ezWRNiaW #AI #Boardroom #AIGovernance #AIagents #BusinessStrategy

✨New working paper on the trade-offs involved in AI transparency in news 🤖📝 How does a global news organisation disclose its use of AI? Where, when and how should readers be told when algorithms shape the news they consume? Based on a case study of the Financial Times and led by Liz Lohn we argue that transparency about AI in news is best understood as a spectrum, evolving with tech advancements, commercial, professional and ethical considerations and shifting audience attitudes. 🔗Pre-print: https://lnkd.in/gV3dPXgS 1️⃣ AI‑transparency ≠ a binary. At the FT it’s a hybrid of policy, process and practice. Senior leadership sets explicit principles, cross‑functional panels vet new applications, and AI use is signposted in internal/external tools and reinforced through training. 2️⃣ Disclosure is calibrated to context. Internally, full disclosure aims to reduce frictions and surfaces errors early; externally, labels are scaled with autonomy and oversight. No‑human‑in‑the‑loop features (e.g. Ask FT) get prominent warnings, whereas AI‑assisted, journalist‑edited outputs (e.g. bullet‑point summaries) get lighter labelling. 3️⃣ Nine factors shape what, when & how the FT discloses AI use. These include legal/provider requirements, industry benchmarking, the degree of human oversight, the nature of the task, system novelty, audience expectations & research, perceived risk, commercial sensitivities and design constraints. 4️⃣ Persistent challenges include achieving consistent labelling (especially on mobile), breaking organisational silos, keeping pace with evolving models and norms, guarding against creeping human over‑reliance, and mitigating against “transparency backfire” where disclosures reduce trust. For those of you more academically interested in this, we argue that AI transparency at the FT is shaped by isomorphic pressures – regulations, peer practices and audience expectations – and by intersecting institutional logics. Internally, managerial and commercial logics push for efficient adoption and risk management; externally, professional journalism ethics and commercial imperatives drive an aim to remain trustworthy. Crucially, we argue that AI transparency is best seen as a spectrum: optimising one factor (e.g. maximum disclosure) can undermine others (e.g. perceived trust or revenue). There does not seem to be a one‑size‑fits‑all rule; instead transparency must adapt to org context, audiences and technology. We are very grateful to the team at the Financial Times, particularly Matthew Garrahan, for supporting this study from the outset – and to the participants from the FT who volunteered their precious time to help us in understanding this issue. Feedback welcome, especially on the theoretical section and the discussion as well as literature that we will have missed! So feel free to plug your own or other people’s material, all of which will be appreciated as Liz and I work towards a journal submission.

Most boards cannot answer one basic question about their AI strategy: Where does the training data actually come from? AI strategies are being approved without visibility into their most critical upstream input: the information supply chain. This is no longer just a copyright issue. It’s a strategic risk. Most foundation models are trained on massive volumes of scraped web content - while the economic and legal foundations of that practice remain unresolved, especially for high-quality media and knowledge content. That creates a fragile system: - AI systems need more data - Much data is taken, not sourced - Incentives for original content are shrinking - Media and knowledge models are weakening - Law was not designed for industrial-scale text and data mining High-quality information is to AI what semiconductors are to hardware: a critical upstream input that is currently under-priced and weakly governed. An extensive new analysis by RTR (Rundfunk und Telekom Regulierungs-GmbH (RTR)) reinforces what many CIOs and boards are starting to see: If we do not build lawful and scalable data markets for AI training, we will end up with: - declining model quality (more synthetic, less original data) - rising regulatory and litigation exposure - fragile AI strategies built on unstable inputs - long-term damage to media and knowledge ecosystems There will be two kinds of AI strategies: 1. Built on scraped, legally ambiguous data. 2. Built on licensed, traceable, high-quality information. The second will outperform - economically, legally, and reputationally. For CEOs and CIOs, the strategic question is no longer only: Which model should we deploy? It is now: What information supply chain are we depending on? The full RTR study is worth reading — this debate is moving from policy to core digital infrastructure strategy. cc: Klaus, Wolfgang Struber, Thomas Schreiber, Timo Steyer, Philipp Homar, Lukas Moormann.

AI capability keeps accelerating. Transparency is moving in the opposite direction. Stanford’s 2025 Foundation Model Transparency Index shows a sharp decline in disclosure across major AI developers. Participation dropped. Reporting narrowed. Critical details around training data, environmental impact, and downstream risk are increasingly opaque. At the same time, these systems are shaping search results, procurement decisions, financial forecasts, and clinical screening at scale. When AI systems influence outcomes across entire populations without meaningful inspection, power concentrates and accountability thins. In this article, I explore the growing transparency gap, the economic consequences for the open information ecosystem, and why artificial integrity must become a structural requirement, not a voluntary gesture. If AI is infrastructure, transparency cannot be optional. #AITransparency #AIGovernance #ArtificialIntelligence #DigitalEconomy #ResponsibleAI