The EU has regulated ChatGPT as a search engine. Calling it a platform would have handed OpenAI a liability shield.

The European Commission designated ChatGPT a Very Large Online Search Engine under the Digital Services Act, exposing OpenAI to fines of up to 6% of global revenue but covering only the parts of the tool that retrieve rather than converse. The article’s argument is that the alternative label, Very Large Online Platform, carried safe harbour, the protection platform law gives companies hosting other people’s content, and that classifying a chatbot that way would have implied a model’s output is third-party material.

The European Commission designated ChatGPT a Very Large Online Search Engine under the Digital Services Act on Monday, imposing transparency and risk-mitigation duties comparable to those on Google and Bing, and exposing OpenAI to fines of up to 6% of global annual revenue. Eliza Gkritsi and Pieter Haeck reported for Politico that the designation is narrow enough to leave the chatbot’s actual conversations largely uncovered.

The distinction is concrete. Asking ChatGPT to name candidates in a local election should now fall under the DSA, while a conversation with the user about who to vote for, in which misinformation might surface, may not.

Why not call it a platform

The Commission’s other option was Very Large Online Platform, the category covering social media and marketplaces, which brings content moderation obligations. It also brings something OpenAI might have welcomed.

Platform law worldwide is built on safe harbour, the principle that a company is not liable for content because users uploaded it. Applying that frame to a chatbot raises the question of whether a two-way exchange between a person and a machine counts as user-generated content at all.

Answer yes, and a model’s output starts to look like third-party material its maker merely hosts. That is a considerable prize to hand a company whose product generates every word it returns.

The cost of the safer choice

Avoiding that concession has a price, and children are where it lands hardest. Christel Schaldemose, the Danish socialist MEP who was among the DSA’s key negotiators, said ChatGPT is “much more than a search engine” and that risks connected to the chatbot itself fall outside the regulation’s strongest obligations.

She named emotional dependency and manipulative or addictive design, and asked the Commission to clarify how existing rules cover them. Those are design questions about the conversational layer, which is the layer the designation does not reach.

The scale is not marginal. As many as 60% of adults globally use chatbots for therapy, according to AXA’s 2026 mind health report, and the companionship use case has been at the centre of litigation including a suit brought by the parents of 16-year-old Adam Raine in California.

A category problem, not a drafting error

The DSA was finalised in 2022 and did not anticipate this. João Pedro Quintais, associate law professor at the University of Amsterdam, describes ChatGPT as a hybrid with the functions of a search engine, of an online platform, and of something closer to a publisher of its own content.

That third function is the one no existing category handles. The DSA was written to govern how companies handle material that other people produce, and a model producing the material itself sits outside that architecture.

It took the Commission just under a year to reach a decision, which is a reasonable indication of how hard the classification was. Experts told Politico that without the full designation text it is difficult to know precisely what OpenAI must now do.

The parallel track, and its blind spot

The Commission has a second instrument pointed at the same company. Under the AI Act, builders of general-purpose models have been obliged since last August to assess and mitigate systemic risks, and enforcement began at the end of August with the Commission questioning AI companies on their security procedures.

Look at what the expert group singled out: nuclear and biological weapons, loss of control over models, models autonomously hacking, and large-scale manipulation. Those powers were sharpened this year around rogue-agent scenarios.

Daniel Leufer of AccessNow identifies the gap precisely, telling Politico the guidance is “more focused on so-called existential risks than risks to fundamental rights”. A teenager forming a dependency on a chatbot is not a systemic risk in the AI Act’s sense, and is now not clearly a search risk under the DSA either.

Where the harms actually get regulated

Brussels does have a working method for this, developed elsewhere. The Commission has pursued Meta over addictive design aimed at children under the DSA, treating interface choices as the regulated object.

Leufer argues the designation is an opportunity to do the same here, to get into the design decisions and treat ChatGPT as what it is, which is a product. That framing has been gaining ground across jurisdictions.

Regulators from Brazil to California have converged on design rules rather than access bans, having found them more enforceable. Whether a search-engine designation gives Brussels the leverage to apply that method to a conversational interface is the open question.

What to watch

The full text is the thing. Until the designation is published, the scope of OpenAI’s obligations on the conversational layer is guesswork, and TNW has not seen it.

Watch also whether the Commission tests the boundary rather than accepting it. Designation is the beginning of DSA supervision, not the end, and the first formal request for information will reveal how broadly Brussels intends to read its own decision.

The safe harbour question will not stay theoretical for long. Some regulator, somewhere, will have to decide whether a sentence a model wrote is content the model’s owner is responsible for, and the answer will matter more than the label attached this week.