Year One

A look back on what we've accomplished together in our first year at the AI Disclosures Project

It’s been one year since the AI Disclosures Project (housed at the Social Science Research Council in New York City) was founded by Tim O’Reilly and myself, driven by the belief that “you can’t regulate what you don’t understand”. Tim meant by that the urgent need for visibility into how companies are deploying AI systems, including the downstream impacts of this on society, especially as they are optimized in commercial contexts. These are now products after all as much as they are models.

In this special Year in Review newsletter we highlight some of our most important achievements to date. Our mission is to build healthy AI markets. Increasingly, we believe this means architecting AI markets so that they can support a transparent, diverse, and permissionless system of innovation. Our argument is that open and effective protocols, and auditable telemetry, are central to ingraining these principles into the software infrastructure that wires together AI markets.

In this brief Review we cover our research, advocacy, and policy.

Research

Model Training Violations. We published research showing that OpenAI increasingly trains its models on non-public copyrighted book material. The research was covered by dozens of media outlets, including TechCrunch and FastCompany. This work highlighted the urgent need for increased corporate transparency regarding pre-training data sources as a means to develop proper licensing and remuneration ecosystems. The paper is under review with the AI and Ethics journal.

Post-Deployment AI Risks and Mitigations. In this review paper we analyzed thousands of research papers on AI governance and society. In the literature we found widespread neglect of AI risks during its deployment, including sycophancy, persuasion, copyright, healthcare, and more. We highlighted the extreme concentration of AI research among a handful of corporations, as well as the importance of telemetry in getting the needed observability into these in-market risks. The paper is under review with the ACM Journal on Responsible Computing. It was featured in HuggingFace’s Daily Papers and in Kevin Fumai’s Top Posts.

The Open Internet is Under Attack. In this paper we used a real-world dataset of AI-human conversations to estimate the ‘attribution gap’: the difference between the number of relevant websites consumed by AI systems when answering a user question and the number of citations they provide. We found major gaps in Gemini and Perplexity’s attribution practices. OpenAI’s logs were more difficult to inspect. We highlighted the importance of auditable logs and traces (as telemetry) for creating open and mutually monetizable ecosystems for these internet-enabled AI markets. The paper is under review with the Knowledge-Based Systems journal.

Protocols and Power. We published a piece on protocols in AI Frontiers framing the importance of the Model Context Protocol (MCP) as a market-shaping device, that can decentralize power through unbundling context (user data) – allowing it to move more freely between applications and services. We highlighted the importance of open APIs and portable memory. A fuller version was published on this Substack.

An extended version of this argument, with more detailed policy recommendations, was just published as a policy paper with the Social Science Research Council (available here). I encourage you to read it if you want to start to dig deeper into the policy context of our arguments.

Advocacy

Internet and Protocols. We have been involved with several initiatives to ensure the internet, in the era of AI, can be sustained. This has focused on leveraging decentralized protocols and fair market access. We participated in the Creative Common’s Signals workshops, ongoing discussion with Miso AI (led by Lucky Gunasekara), the Real Simple Licensing (RSL) framework, and Microsoft’s NLWeb initiative. This work built on Tim’s original important piece,“How to Fix AI’s ‘Original Sin’”.

Asimov’s Addendum Substack. We launched this newsletter to narrate aloud our work on AI’s deployment risks in commercial contexts, AI corporate disclosures, and protocols for AI markets. It has grown to over 1,000 subscribers and was featured in Charlie Guo’s top substacks on AI to follow for 2024 with under 1,000 subscribers.

Presentation and Lectures. Tim provided inputs remotely to a convening at Oxford University on copyright and AI (here). Tim also presented at the AI Risk Management and Governance: Lessons from Financial Regulation and Bank Supervision at The Wharton Business School (January 30-31, 2025).

Ilan presented on the cloud as a regulatory layer at Yale and at the ITRev Forum; on building “countervailing power” on the internet at the University of Johannesburg (here); and with Sruly Rosenblat at Data for Policy on making AI thinking and behavior more transparent (here).

Conferences. Ilan attended the Paris AI Summit, following on from the 2024 UK summit. Isobel attended the AI for Good summit in Geneva and participated in a number of discussions. Tim attended the NBER Conference on the Economics of AI (September 19-20, 2024). Ilan attended the 2024 Independent Tech Researchers' Summit by the Coalition for Independent Technology Research (June 11-12, 2024) and Partnership on AI's Policy & AI Forum (NYC, September 20, 2024).

We are immensely grateful to all the individuals who met with us to extend our understanding of the numerous risks and mitigations ongoing in AI markets, AI models, and AI systems.

Policy

Ongoing work on protocols. We recently published our more detailed policy recommendations on how to structure protocols and APIs for open AI markets here. We have provided confidential inputs to several AI labs and government policy divisions on open APIs, AI memory interoperability, and leveraging protocols for healthy AI markets. These are ongoing. We presented to Alexandra Bensamoun, advisor to the French Culture Minister, on the potential role of protocols in fostering a competitive market for “cultural data providers” online. Our brief background note can be found here.

EU Digital Services Act (DSA) Submission. We made recommendations on how the DSA should mandate third-party research access to platforms through the Delegated Act. This submission applied insights from accounting-based approaches to algorithmic regulations and disclosures, based on several Asimov’s Addendum pieces, and our previous work on algorithmic disclosures. Available here.

Our Team. Moving into our offices housed at the Social Science Research Council, we also hired two superb researchers, with diverse skill sets. Ilan is an economist after all, and Tim a technologist, publisher, and tech evangelist.

Isobel Moure. Isobel is a recent graduate of Barnard College of Columbia University, where she studied Cognitive Science and Computer Science. Isobel led the writing of our recent work on Protocols published in AI Frontiers. She coordinated the weekly roundup for the Asimov Addendum substack and contributed to the working paper: “Real-World Gaps in AI Governance Research: AI Safety and Reliability in Everyday Deployments.”

Sruly Rosenblat. Sruly graduated from Hunter College with a bachelor’s degree in computer science, graduating at the top of his classes. He brings technical expertise in programming and machine learning techniques to the project. Sruly led the technical work for “Beyond Public Access in LLM Pre-Training Data: Non-public book content in OpenAI’s Models”, and constructed the dataset for our Attribution Gaps paper on the citation practices of internet-enabled AI systems.

Look Ahead

We expect the next 12 months to be our most important yet as we dive deeper into how to architect healthy, open, transparent AI markets through protocol-level mechanisms. This will also include how governments can use existing corporate telemetry (logs, traces, metrics) on their AI systems as a basis for third-party confidential auditing.

Thank you for your ongoing support! We look forward to the next part of our journey together with you.