EU Opens 11 Android Features to Rival AI Assistants

Article Brief

Key Takeaways

4 Points24s Read

1. 01Scope–The binding DMA decision covers 11 Android features used for invocation, context, actions and device resources.
2. 02Deadline–Most measures are due with Android 18 and by August 1, 2027; concurrent wake words follow by August 1, 2028.
3. 03Safeguards–Sensitive capabilities remain subject to user consent and objective, independently certified security criteria.
4. 04Competition test–The decisive evidence will be task reliability, latency, resource parity and permission friction—not the mere existence of an API.

The European Commission has closed the Android AI specification proceedings it opened in January with binding specification measures: Google must give rival Android AI assistants effective access to 11 operating-system features that help Gemini hear, understand and act for a user. The decision is narrower than an order to “open Android,” but more consequential than a new default-app screen. It reaches the operating-system privileges that turn a chatbot into an assistant.

The final specification measures were adopted on July 16 under Article 6(7) of the Digital Markets Act. Most of the Android work is due with Android 18 and no later than August 1, 2027. Concurrent wake-word access, which would let separate services be triggered with different wake words, has a later deadline of August 1, 2028 with Android 19.

That timetable marks a new episode, not a replay of January. TECHi covered the opening of the two specification proceedings when the Commission had not yet decided which features Google would have to expose or on what schedule. The final decision now identifies the required capabilities, imposes parity conditions and creates an implementation clock.

The immediate contest is therefore not over which AI model tops a benchmark. It is over whether the same assistant can be invoked, retrieve on-device context, act across apps, run in the background and use device resources without an artificial disadvantage. Gemini still has Google’s distribution and product integration. What Brussels is trying to remove is the part of that lead that comes from privileged access to Android itself.

What Android AI assistants gain below the default setting

An Android user can already install a competing AI app. Installation, however, is not equivalent to system-level capability. An assistant that needs the screen awake, cannot see context another app has chosen to expose, and cannot complete a permitted task in the background is still an app a user visits. Gemini can be woven into the device.

The Commission’s technical Q&A on the 11 features organizes that difference into four practical layers:

• Invocation covers entry points such as buttons and wake words, including use while the display is off or the device is in standby.
• Context covers centrally available on-device app data, screen contents and sensor inputs, subject to choices made by the user and the apps providing the data.
• Actions cover structured app integrations, screen automation and system integration.
• Resources cover equal access to system-level on-device models, the ability to install and run alternative on-device models, and background execution.

These are not decorative permissions. Voice invocation determines whether an assistant is present at the moment of need. Context determines whether it understands what the user is doing. App actions determine whether it can finish work rather than merely describe the steps. On-device models and background execution determine latency, privacy, battery behavior and whether the service remains useful when the network is poor.

The order also extends across the Google Android ecosystem, including devices made by other manufacturers. Google must provide the interoperability free of charge, document it, offer testing and technical assistance, and make covered new functionality available to third parties when Google makes it available to its own services. Access cannot be made unnecessarily cumbersome or conditioned on the outside provider becoming the device’s default assistant.

This is a different regulatory mechanism from the DMA cases involving storefront rules and steering. TECHi’s earlier coverage of Google Search and Play Store compliance concerned how Google presents and monetizes services. The Android AI decision addresses the technical conditions under which another service can perform the same class of task.

App Functions and background work define parity

Wake words will attract the headlines because “Hey Google” is easy to understand. Structured on-device integration carries more practical weight. Android’s App Functions can let an AI service request a defined action from another app, such as sending a message, creating a note or scheduling a meeting. The decision says that channel cannot remain reserved for Google Assistant or Gemini.

For apps that do not expose a suitable structured function, the measures also cover screen automation. The Commission describes a separate virtual window in which an assistant could complete a multi-step task while the user does something else. That capability raises obvious reliability and security questions, but it also separates an agent that can execute from a chatbot that can only advise.

Context is equally important. Alternative services are meant to receive comparable access to data that apps have stored on the device and chosen to make available, as well as consented screen and sensor inputs. The Commission gives examples ranging from live translation to surfacing a flight booking number during a call. Its language matters: apps and users decide what is made available. The order is not a blanket right for every assistant to read every screen, file or microphone stream.

The same distinction applies to resources. Rival services are to receive equal-performance access to covered on-device models, including Gemini Nano models that are part of the designated operating system. They may also install and run their own on-device models under the same hardware-resource and background-execution conditions as Google’s models. A nominal API that is slower, more battery constrained or suspended more aggressively would not deliver effective parity.

The Gemini moat is being translated into measurable parity obligations. The legal question will be compliance, but the product question will be latency, success rate, resource limits and friction at each access point. A permission dialog that technically exists but drives most users away can preserve the practical advantage. So can an interoperability solution whose performance trails the path available to Google’s own assistant.

Google disputes the Commission’s safety premise

Google’s objection is direct. Kent Walker, the company’s president of global affairs, wrote that the decisions risk weakening privacy and security safeguards. In Google’s published response, he argued that phone makers already play a role in vetting assistants and that sensitive device permissions should not be detached from those safeguards.

The final measures do not say that access must be unconditional. Users must explicitly consent to covered access. For five sensitive capabilities—including screen automation, structured integration, system integration, centralized app-data access and context-aware intelligence—Google may set objective and non-discriminatory privacy, security and integrity criteria. Independent parties will join Google in certifying whether an app meets them. Google must publish draft eligibility terms by February 1, 2027, final terms by May 1, and begin accepting applications on that date.

That creates a difficult enforcement line. Criteria that are too weak could expose users to assistants with poor security practices. Criteria that are vague, costly or selectively applied could become a new gate. The Commission says Google may not attach extra commercial requirements and that completed applications should be assessed within four weeks. The quality of the rules will become visible well before Android 18 ships.

The dispute should not be flattened into a choice between competition and safety. Android already mediates high-risk permissions, and AI assistants combine those permissions in ways that can amplify harm. The Commission is asking for parity with consent and certification. Google is warning that parity may bypass device-maker vetting it treats as part of the safety boundary. Both claims can be tested only against the implementation, incident record and treatment of comparable services.

The platform has also changed since TECHi reviewed the Gemini-heavy Android 16 roadmap. Assistants now aim to perform multi-app tasks, use live device context and run local models. That raises the value of operating-system access and the cost of getting the boundary wrong.

Search data is a separate order with the same competitive logic

The Commission adopted a second binding specification for Google Search under Article 6(11). It is legally separate from Android interoperability, but it targets another feedback advantage: the query, ranking, click and view data that a search engine collects at scale.

The Search data specification says eligible AI chatbots with online-search functionality may apply. Google’s earlier offer removed 90% to 100% of unique queries, which the Commission judged ineffective. The final approach covers a restricted and anonymized subset of ranking, query, click and view data—not Google’s algorithm. It imposes layered alteration, ring-fenced use, retention limits and a prohibition on re-identification.

The data must arrive at least seven days after a query and can be supplied for up to five years to each beneficiary. An applicant must provide an online search service as a genuine EU business or credible new entrant and show at least 50,000 average monthly EU users over the prior year. A company founded less than two years ago can satisfy the new-entrant prong by raising more than €50 million, but it must still meet the other eligibility conditions. An independent audit comes before access, a compliance audit follows within six months of processing, and annual audits follow thereafter.

Google must publish information for prospective beneficiaries and submit an application form by the end of August 2026, make template licenses and test samples available by September, finalize the anonymized dataset by November, and finalize pricing by January 2027. Recipients may use the data to improve an online search service, not for unrelated model training or onward distribution.

The Android and Search measures therefore attack two different bottlenecks. One is the ability to act on a device. The other is the behavioral data used to improve a search service. A rival that qualifies for one will not automatically master the other, and neither order supplies distribution, trust, app partnerships or a competitive model. The measures reduce structural disadvantages; they do not manufacture a successful challenger.

What would prove that the policy changed competition

The first evidence will not be a download chart. It will be Google’s draft certification terms, developer documentation and test environment. Those materials will show whether the 11 required capabilities work reliably, whether OEM implementations are consistent, and whether an outside assistant can test realistic multi-app workflows before a public release.

After Android 18, useful comparisons should measure task completion, permission friction, latency, battery cost and failure recovery for the same action on Gemini and a qualifying rival. The Commission has said the solutions must be equally effective. That language gives regulators a basis to examine practical degradation, not merely the presence of an API.

Security evidence matters just as much. The relevant record will include certification decisions, denied applications, disclosed vulnerabilities, abuse rates and whether Google applies comparable standards to its own services. A clean incident record would weaken Google’s broadest warning; recurring abuse would give Brussels reason to adjust the specification.

The decision does not end Google’s Android advantage. It places a legal boundary around one source of that advantage and makes the next phase observable. Most parity obligations are due by August 2027, with concurrent wake-word support following by August 2028. The question is whether the 11 required capabilities will work well enough for a user to choose a rival assistant without accepting a second-class device experience.