The Regulator Named the Assistant

The named object

On July 16 2026, the European Commission adopted a final decision under Article 6(7) of the Digital Markets Act specifying the measures that Google must implement to give third-party AI assistants effective access to eleven named features of the Google Android operating system. The decision closes a specification proceeding opened on 27 January 2026. It is the first time a European regulator has named the AI assistant as a class of software entitled to interoperability with the operating system layer of a mobile platform. The class is not defined by the market position of any specific product. It is defined by function.

The list of eleven features is not incidental. It is grouped into four capability buckets: invocation, context, actions on apps and the operating system, and access to on-device resources. Under invocation, the decision names the long-press home button and always-on hotword detection. Under context, it names centralised access to on-device application data, context-aware intelligence, and ambient data. Under actions, it names structured on-device integration, screen automation, and system integration. Under access, it names system-level on-device models, on-device model implementation, and background execution. The eleven names, taken together, describe what an AI assistant is understood to require in order to be an AI assistant on a phone.

The specification is an ontology

Most Article 6(7) interoperability decisions specify measures against a technical standard. This one specifies measures against a functional definition of the object being regulated. The measures are keyed to eleven capabilities the Commission has judged crucial for a feature-rich and performant AI service. The Commission has, in the language of the decision, decided what an AI assistant requires. The regulated party is required to give the object those capabilities, on non-discriminatory terms, whether the object is Google's own assistant or a competitor's.

This is a substrate move rather than a competition move. The competition move would have specified access to a store of assistants, a payments rail, or a search default. The substrate move specifies access to invocation, context, action, and on-device resource capabilities that together constitute the operating environment for the class. The Commission has decided that the operating environment is the governance surface. Whether the resulting environment produces meaningful competition among assistants is downstream of whether it produces a shared substrate.

The tempo is written into the decision

The measures have a schedule. Google must implement them in the next major Android release, Android 18, and no later than 1 August 2027. Concurrent hotword detection, which would let more than one assistant respond to a wake word without pre-emption by Google's own hotwords, must be implemented in Android 19, no later than 1 August 2028. The decision also specifies a compliance sub-schedule for the eligibility program that gates access. By 1 February 2027, Google must publish draft terms for third-party consultation. By 1 May 2027, Google must publish the final terms and begin accepting certification applications. Each application must be assessed within four weeks.

The schedule is doing structural work. It commits the regulator, in writing, to an implementation tempo that runs from January 2026 through August 2028. During that thirty-one-month window, the class of software the decision is written to serve continues to evolve at a tempo that is not the regulator's. Assistants shipping today already invoke through means other than the home button, gather context through means other than the platform application programming interfaces, and act on the operating system through means other than the platform hooks the decision enumerates. By the time Android 19 ships, the enumerated features may describe an object different from the one the decision was written to regulate.

This is the familiar shape of an interoperability instrument written against a moving target. What is unusual here is that the instrument names the target as a class of software rather than as a specific incumbent. Naming a class binds the regulator to whatever the class becomes.

The eligibility gate

The decision preserves a certification gate. Google is required to run an eligibility program under terms subject to third-party consultation and Commission oversight. The gate is intended to filter for security, privacy, and abuse risk. It is also, mechanically, the surface where the substrate the regulator has named will be adjudicated case by case. The party running the gate is the party required to open its own capabilities to competitors through the gate. The party overseeing the gate is the party that wrote the decision. The parties applying to pass the gate are the class of software the decision is written to serve.

Three parties, one gate. The Commission's decision text does not resolve which party's judgment binds when they disagree.

Where this sits in the arc

The Mythos Asymmetry arc has been tracking the widening gap between policy instruments and the systems those instruments are written to bind. The Article 6(7) decision is a new kind of entry in that record. Where earlier entries described taxonomies lagging deployments, this one describes an instrument that names the deployment class in advance and then commits to a two-year and three-year schedule to make the substrate available to it. The naming is prospective. The schedule is not.

The decision also lands in the same week as an EU-institution-native large language model released as an open model by the Commission's own translation directorate and the founding of a twenty-nine-country parallel intergovernmental AI cooperation body in Shanghai. Each is a sovereignty move in a different register. Together they suggest that the substrate question, who owns the operating environment that AI assistants run inside, is being addressed by multiple institutions in parallel, with implementation tempos that do not align.

What remains on the table

• Do the eleven named features remain the correct enumeration when Android 18 ships, or does the class of software the decision is written to serve evolve past the enumeration during the specification window?
• Who adjudicates the eligibility gate when Google, an applicant, and the Commission disagree on whether an assistant meets the certification criteria?
• What happens to concurrent hotword detection during the year between the Android 18 deadline and the Android 19 deadline, when the platform is required to admit competing assistants but not required to admit their wake words?
• Do parallel sovereignty moves in the same week converge on a shared substrate or fragment into competing ones?

The policy instruments and the deployment tempo are not aligned.