EU Forces Google to Open Android AI to Third Parties Under DMA

The European Commission has concluded its specification proceeding into Google's AI implementation on Android, recommending that Google must open up system-level AI features to third-party services under the Digital Markets Act (DMA). The commission specifically targets Gemini's privileged access to hot words, screen context, local data for proactive suggestions, and hardware for running local models.
Key Proposed Changes
- System-Wide Invocation: Third-party AI assistants must be triggerable via hot words or button presses, just like Gemini.
- Screen Context: Alternative AI tools must be allowed to read screen content when invoked by the user.
- Local Data Access: AI services must be able to access local data for proactive suggestions and summaries (similar to Google's Magic Cue).
- App Control: Third-party AI assistants must be able to autonomously control installed apps and system features (e.g., sending emails, sharing photos).
- Hardware Access for Local Models: Mandate to ensure developers have necessary hardware access to run local models with high performance and responsiveness.
Google's Response
Google senior competition counsel Claire Kelly stated: This unwarranted intervention would strip away that autonomy, mandate access to sensitive hardware and device permissions; unnecessarily driving up costs while undermining critical privacy and security protections for European users.
Context
The DMA has already forced Google to implement search choice screens, allow alternative payment methods in the Play Store, and limit data sharing. The commission VP for Tech Sovereignty Henna Virkkunen said: As we navigate the rapidly evolving landscape of AI, it is clear that interoperability is key to unlocking the full potential of these technologies.
Currently, Gemini is preinstalled and has exclusive system-level access for proactive suggestions. The commission expects formal measures by summer 2026.
📖 Read the full source: HN AI Agents
👀 See Also

Claude Consumer Terms Analysis: Data Retention, Liability Caps, and Service Termination
An analysis of Anthropic's Consumer Terms of Service reveals key details for $100/month Max plan subscribers: data training is on by default with 5-year retention for opted-in users, liability is capped at $600 maximum, and service can be terminated without refund for violations.

Anthropic source code leaks reveal unannounced Claude features and internal model
Anthropic accidentally leaked 500,000 lines of source code containing details about unannounced Claude features including KAIROS background execution, dream mode, undercover mode, and an internal model called capybara. This is the second such leak in 2025.

Self-Supervised Fine-Tuning on Own Mistakes Boosts Small Models to 80% on HumanEval
A developer trained Qwen 2.5 7B on its own self-generated coding pairs, reaching 112/164 HumanEval (+87 problems) with zero human-written training data. The approach transfers to Llama 3.2 3B and Qwen 3 4B.

Claude API Usage Data Shows Impact of New Limits on Max Plan Users
A Claude Max 20x user reports API-equivalent daily usage dropping from ~$210/day to ~$52/day after new limits were implemented, requiring significant workflow changes including using Sonnet and Codex.