The DMA’s Gatekeeper Rules and Default AI Assistants

A phone that ships with the manufacturer’s own AI assistant bound to the side button looks like the exact conduct the Digital Markets Act was written for. Whether the DMA reaches it depends on a designation decision that may not exist for that assistant, and on a definition written in 2022 for a different generation of software.

Virtual assistants are a listed core platform service

Regulation (EU) 2022/1925, the DMA, does not regulate large companies. It regulates named services. Article 2 lists the categories that can be core platform services — online intermediation services, search engines, social networking, video sharing, number-independent interpersonal communications, operating systems, web browsers, virtual assistants, cloud computing, and advertising services. Virtual assistants are on that list explicitly, which is more than can be said for most software categories, and it is the hook everything else in this page hangs from. The text is at EUR-Lex.

The definition in Article 2 describes software that can process demands, tasks or questions, including those based on audio, written input, gestures or motions, and that provides access to other services or controls connected physical devices on that basis. Read literally it is broad enough to cover a modern chat assistant that can call tools and drive device functions. Read historically, it was drafted with voice assistants in mind, and the “provides access to other services or controls connected devices” limb is doing real work: a generative assistant that only answers questions in a text box, without acting on other services, arguably falls outside it.

Designation is per service, not per company

Nothing in the DMA applies to an undertaking until the Commission designates it a gatekeeper in respect of a specified core platform service, under Article 3. The presumption thresholds are three cumulative tests: a significant impact on the internal market (quantified by annual Union turnover in each of the last three financial years or an average market capitalisation, with the service provided in at least three Member States); an important gateway for business users, evidenced by monthly active end users in the Union and yearly active business users; and an entrenched and durable position, presumed where the user thresholds were met in each of the last three financial years. An undertaking that meets the thresholds must notify the Commission; the Commission may also designate on a qualitative assessment where the thresholds are not met.

This is why “company X is a gatekeeper, therefore its assistant is covered” does not follow. A company designated for its operating system and its search engine is not thereby designated for its assistant. The obligations in Articles 5, 6 and 7 attach to the designated services listed in the decision. Two separate questions therefore have to be asked of any bundled assistant: is this assistant itself a designated core platform service, and does the assistant’s behaviour engage an obligation owed in respect of some other designated service, typically the operating system it ships on.

The second question is usually the productive one, and it is the reason the defaults rule below matters even where no assistant has been designated in its own right.

What Article 6(3) requires about defaults

Article 6(3) is the choice-architecture provision, and it names virtual assistants directly. In outline it requires a gatekeeper to allow and technically enable end users to easily change default settings on its operating system, virtual assistant or web browser that direct or steer them to the gatekeeper’s own products and services; to allow uninstallation of pre-installed applications, subject to a carve-out for applications essential to the functioning of the operating system or device that cannot technically be offered standalone by third parties; and, when an end user first uses a designated search engine, virtual assistant or browser, to prompt them to choose from a list of the main available providers.

Three things follow that are easy to miss. The uninstallation carve-out is narrow — “deeply integrated” is not the test, “essential and not technically offerable standalone” is. The choice-screen duty is a first-use duty, so a design that buries the choice behind a settings menu does not satisfy it. And the easily-change duty covers steering as well as literal defaults, which reaches interface patterns that make switching nominally possible and practically unattractive.

If a designated operating system routes a hardware button, a system search field or a share sheet to the gatekeeper’s own assistant, Article 6(3) is engaged in respect of the operating system even if the assistant itself has never been designated. That is the mechanism worth understanding, because it does not wait for the Commission to decide anything about generative AI.

What is genuinely unsettled

It is not yet clear whether the DMA as drafted reaches generative AI assistants as a category. The Commission ran consultations on the question rather than asserting the answer, and the Regulation contains its own review clause: Article 53 obliges the Commission to evaluate the DMA and report to the co-legislators, with the first evaluation due in 2026 and further ones periodically thereafter, expressly including whether the list of core platform services should change. That review, or an amending regulation following it, is what would settle the scope question. A Commission speech does not.

Two positions are visible in the debate. One holds that the Article 2 definition is technology-neutral and already covers a tool-using generative assistant, so no amendment is needed. The other holds that generative assistants are typically integrated into designated services rather than offered as standalone gateways, so the user-number thresholds are hard to apply and an amendment naming them would be cleaner. Both are arguable on the text.

Consequences if an assistant is designated

Designation is not a label; it is a compliance programme with a six-month clock. The obligations that bite hardest on an assistant are the Article 5(2) restriction on combining personal data across a gatekeeper’s services and with third-party services without specific consent — awkward for an assistant whose value comes from context across mail, calendar and search — the Article 6(5) prohibition on self-preferencing in ranking, and the Article 6(7) interoperability duty, which is the subject of its own page. Penalties under Article 30 run to 10% of worldwide turnover, and 20% for a repeated infringement, with structural remedies available for systematic non-compliance under Article 18.

None of this overlaps neatly with the AI Act, which regulates the same assistant on entirely different grounds — risk classification, transparency to users, and general-purpose model obligations. A designated assistant owes both sets of duties independently; see the risk tiers and the general-purpose model obligations for that half.