What changes

The designation brings the DSA’s heavier obligations: risk assessments, mitigation measures, independent audits, data access, and transparency reporting. These include vetted researcher access to data. OpenAI has until January 2027 to comply. The scope is narrower than the headlines suggest. This is a VLOSE designation, not a VLOP one. It rests on ChatGPT’s web search function, and the chat interface that lets users interact with the model does not necessarily search the web. Whether LLMs without search functionality fall under the DSA is still open. 

The role of AI4POL

AI4POL seeks to supports European regulators and policymakers with the tools, insights, and frameworks to develop and enforce AI regulation in line with European values and human rights. The ChatGPT designation shows why this mission matters. The EU is stretching a regime built for platforms and search engines to cover a technology that generates content rather than just hosting it. Regulators making that move need evidence on how these systems actually behave. That evidence is what AI4POL aims to provide. Its work on next-generation AI governance covers tools for identifying online consumer manipulation and detecting self-preferencing in digital markets. Its work on AI development and institutional constraints looks at how governance structures shape AI worldwide. A designation only changes outcomes if regulators can assess what happens inside these systems, and risk assessments and audits need methods behind them.

A teaser from our upcoming paper

This is the question behind a paper co-authored by Kun He, Pratiksha Ashok, and Jens Prüfer, currently under review. We study how LLMs moderate content, which differs from what we know from social media. Social media moderation mostly happens after content is published. In an LLM it happens before a response is generated, and the rules behind it are largely undisclosed.

 We test the same prompts, drawn from real user reports of refusals, across ChatGPT (United States), DeepSeek (China), and Mistral (European Union). We vary the subscription tier and the language of the prompt. The patterns are systematic: what a user can access depends on where the model comes from, how they pay for it, and which language they ask in. We call restrictions that go beyond what the law requires voluntary algorithmic censorship. The term does not accuse anyone of bad intent. It marks the space where providers make discretionary choices that shape what users can know.

The paper also asks whether the DSA can address that discretion. Our early reading is that its transparency and systemic-risk provisions fit the problem better than rules built around individual pieces of content, and the ChatGPT designation gives that argument institutional weight. The designation also leaves gaps: tiered access is not directly addressed, nor are cross-language disparities, and of the three models we study, only ChatGPT has been designated. More on this once the paper is out.

What to watch for? 

The coming months will show what these obligations mean in practice. Will the first systemic risk assessments cover how moderation choices affect access to information? Will vetted researcher access give outside scholars real visibility into model behaviour? And will other services, including those without search functions, follow ChatGPT into the DSA? At AI4POL, we will keep following these questions. The designation means AI governance is now enforcement as well as principle, and enforcement depends on evidence.

Read the EU Commission's press release here