The Search Engine That Talks Back: ChatGPT and the DSA

The growing public use of Generative AI to obtain information raises concerns regarding information integrity and media pluralism. Despite continuing uncertainty over how Generative AI services fit within the categories of the Digital Services Act, the European Commission’s decision to designate ChatGPT as a Very Large Online Search Engine offers a potential framework for addressing these risks at the systemic level. A recent German judgment on a Google AI Overview response, in turn, illustrates that liability for individual false outputs remains a separate and unresolved question.

GenAI services – new information gatekeepers

Generative Artificial Intelligence (GenAI) services are rapidly becoming part of the information infrastructure. In 2025, 32.7% of people aged 16–74 in the EU reported using GenAI tools during the previous three months (Eurostat). A six-country survey has shown that weekly use of GenAI for obtaining information increased from 11% to 24% between 2024 and 2025. By 2026, 10% of respondents covered by the Reuters Institute’s Digital News Report used AI chatbots for news each week, rising to 17% among the youngest age group (18–24).

This development raises two familiar information challenges in a new fashion. The first concerns information integrity. GenAI may reproduce false information, hallucinate, combine unreliable sources, or present uncertain claims in an authoritative form. An EPRS study notes that hallucinated answers can be particularly deceptive because they are persuasively presented and may even be reinforced by invented plausible-looking references.

The second concerns media pluralism. Unlike traditional search engines, which visibly provide users with different sources of information, GenAI services synthesise sources into a single answer. Even though these answers may contain links to the sources, their format may reduce the diversity of perspectives encountered by users and the visibility of publishers. In a browsing study, users clicked traditional search result links in 8% of visits where a Google AI summary appeared, compared with 15% where it was not available. Links within the AI summary itself were clicked in only 1% of visits. The European Parliament has warned that the growing use of GenAI to aggregate or present news has “far-reaching effects on access to and diversity of information”, expressly connecting this development with media pluralism.

The DSA – the usual suspect, but an uncertain fit

The Digital Services Act (DSA) is an obvious place to look for a legal response. Its objective is to contribute to a “safe, predictable and trusted online environment” in which fundamental rights including the freedom of expression and information, as well as the freedom and pluralism of the media are protected (DSA, Art.1(1), Rec. 3, 153). Yet how GenAI services fall within this framework is not fully resolved.

This categorization matters because the DSA uses a layered taxonomy of services. At its foundation are the following types of intermediary service: mere conduit, caching, and hosting. Online platforms are a subset of hosting services, and online search engines constitute a separately defined category of intermediary service (DSA, Arts 3(g), 3(i), 3(j)). The classification determines whether and under what conditions a provider may benefit from the liability exemptions, as well as which due diligence obligations apply. For instance, hosting providers are subject to the notice-and-action mechanism requirement, while online platforms are additionally obliged to have internal complaint-handling and out-of-court dispute-settlement mechanisms (DSA, Arts 4–6, 16, 20–21). The DSA also contains additional duties for specially designated Very Large Online Platforms (VLOPs) and Very Large Online Search Engines (VLOSEs), which have 45 million or more average monthly active recipients in the EU (DSA, Arts 33–43).

Classifying GenAI services under the DSA is challenging. They combine features of different intermediary services without fitting neatly into any of their categories. They may resemble online platforms insofar as they store user-provided information and make content available through the service, yet ordinary conversational outputs are not typically disseminated to the public in the same manner as they are on traditional online platforms. At the same time, web-search-enabled GenAI resembles an online search engine by retrieving information in response to user queries, but differs from traditional search by synthesising and generating answers rather than merely presenting indexed results (DSA, Arts 3(g), 3(i), 3(j); Lorente and Gardhouse, 2026; Bassini and Palumbo, 2026; Lemoine and Vermeulen, 2023).

The VLOSE regime as a systemic response

The European Commission’s decision of 31 August 2026 provides a partial resolution to the puzzle: ChatGPT has been classified as a VLOSE for DSA purposes. It does not necessarily settle whether particular functions also constitute a hosting service or an online platform, nor does it determine the liability regime for generated outputs.

VLOSE status is relevant to at least some information-integrity concerns associated with GenAI services. The DSA does not regulate hallucination or misinformation as such. However, the generation of unreliable information, the manipulation of information flows, or design choices that present uncertain outputs as authoritative may fall within the systemic risk assessment, insofar as they contribute to the identified risks such as risks to fundamental rights, civic discourse, or electoral processes. Mitigation measures, in turn, may include adapting the design, features, or functioning of the service as well as testing and adapting its algorithmic systems (DSA, Arts 34(1)–(2), 35(1)(a), (d) and (f), rec. 84–87).

The DSA also provides a potential legal toolbox to protect media pluralism. The Regulation specifically refers to freedom of expression and information, including freedom and pluralism of the media. For generative search, this potentially brings source selection, the privileging or exclusion of publishers, and the diversity of sources within the scope of risk assessment (DSA, Art 34(1)(b)). The European Parliament has expressly called for the DSA to be properly enforced where GenAI providers are designated as VLOPs or VLOSEs in order to address systemic risks to media freedom and pluralism.

How these obligations will translate into the GenAI context is an open question. The DSA’s systemic-risk framework principally concerns intermediary services that distribute, recommend, moderate, or search information supplied by others. GenAI services additionally synthesise and generate the information. Whether this framework can be adapted to source selection, synthesis, and generation will only become clear through future application.

Liability for individual generated content

Beyond the systemic regulation, the information challenges associated with GenAI may be addressed through liability for individual outputs. This liability matters for at least three reasons. First, it can influence providers’ incentives to reduce false or misleading outputs. Second, it determines whether targeted persons have an effective remedy. Third, an overly broad liability regime may impose significant costs or create over-deterrence for such GenAI services.

The DSA does not harmonise intermediary liability as such, but rather the conditions under which providers of intermediary services are exempt from liability. Therefore, failure to satisfy the conditions for exemption does not itself establish liability. Liability must have an independent basis in applicable Union or national law (Husovec, 2024, 117–118; DSA, rec. 17). Accordingly, the key question is whether the DSA’s liability exemption regime allows liability for individual GenAI outputs where such liability arises under applicable law, such as defamation law.

The VLOSE designation does not itself provide an answer. As has been emphasised, although online search engines are themselves intermediary services under Article 3(j) DSA, they are not covered by a separate liability exemption. Instead, they can be exempted only insofar as their relevant activity qualifies as mere conduit, caching, or hosting (Husovec, 2024, 139–142; DSA, rec. 28). This leaves the liability exemption dependent on the classification of the particular activity, which, as noted, is not an easy task. Moreover, the DSA does not itself establish a positive basis for intermediary liability (DSA, rec. 17). Where no DSA exemption applies, liability is determined under the applicable EU or national law. Crucially, the liability exemptions do not cover information provided directly by the service provider itself (DSA, rec. 18).

A recent German judgment illustrates the problem. In May 2026, the Munich I Regional Court issued an injunction concerning a Google AI Overview containing allegedly false statements about two companies. The court held that the DSA did not preclude the requested injunction, since it preserves judicial orders requiring providers to cease or prevent infringements. It also concluded that Google could not rely on the hosting liability exemption because, in providing the AI Overview, it was not acting solely as a hosting provider. The court distinguished AI Overviews from traditional search, for which limited liability under German case law is justified by the indispensable role of search engines in making the vast amount of information on the internet usable. AI Overviews were considered different: ordinary search already performs that function, while the additional AI layer independently structures, evaluates, and generates substantive statements from the underlying material, and is not similarly indispensable (LG München I, 2026, English translation).

The judgment does not settle the position of GenAI services under the DSA. It is a first-instance national judgment, and, following a settlement between the parties, it will not be subject to appellate review (LG München I, 2026). Notably, the judgment concerned an AI Overview layer integrated into a conventional search engine rather than a standalone GenAI service. However, its reasoning may be relevant by analogy. More broadly, the case illustrates the conceptual tension sharpened by ChatGPT’s VLOSE designation: can a service be regulated as an intermediary search engine at the systemic level while particular generated statements are treated as the provider’s own publication for liability purposes?

Conclusion

At the systemic level, the VLOSE regime offers a potential framework for addressing the effects of GenAI services on information integrity and media pluralism. In relation to individual outputs, the position is less settled. The difficult questions concern what information is being intermediated, when transformation becomes generation, and how the DSA’s liability exemptions apply.

The search engine now talks back, and the EU is only beginning to decide what to do with this “speech”.

Author’s disclosure: This post was prepared with the assistance of ChatGPT—proof, perhaps, that the search engine really does talk back. No liability exemption is claimed. And yes, the em dash was intentional.

Heading image by Gerd Altmann from Pixabay.