Analysis · Privacy and product design

Meta's €805 million fine: what on-device data, consent and AI features mean for product teams in Europe

Italy's data protection authority fined Meta €805 million because a Facebook feature, on by default, read photos, videos and metadata from Android phones without specific consent. The striking part is not the size of the fine but the principle: in the EU, reading data on a user's device needs consent even if nothing leaves the phone. As AI features look for more on-device context, that principle reaches far beyond Meta.

AnalysisSumanta Mukhopadhyay8 min read6 sources
#news#governance#ai#architecture
ANALYSIS · ANALYSIS · PRIVACY AND PRODUCT DESIGN16 Nov 2026 · 21:20 GST

What the Garante decided

Italy's data protection authority fined Meta €805 million over the Facebook Android app's gallery suggestions feature, which was on by default and proposed photos and videos from users' phones to post; the authority found it accessed photos, videos and metadata without specific, informed consent1. Meta must ask for consent through a separate, specific prompt, keep the feature off by default and report back to the regulator; Meta said it strongly disagrees and will appeal23.

The legal basis matters more than the number. The decision rests on the rule, from the EU ePrivacy Directive, that storing or accessing information on a user's device requires consent unless it is strictly necessary for a service the user asked for4. Reporting notes the authority treated that requirement as applying whether or not data left the phone1.

Why this matters beyond Meta

Many product teams assume that processing data locally is automatically privacy-friendly and needs no special permission. The European Data Protection Board's guidelines on the technical scope of the device access rule take a broader view: the rule turns on gaining access to information on the device, and the guidelines discuss local processing among their use cases5. On-device processing is often good privacy engineering; it is not, by itself, a consent exemption.

AI features make this sharper. Assistants that read photos to suggest posts, scan messages to draft replies, index files to answer questions or use sensors for context all reach into the device. The more context a feature uses, the more it looks like the gallery suggestions feature the Garante fined.

Higher sensitivity
Photos, messages, filespersonal contentContacts and calendar
Medium
Location and sensorsInstalled apps and usage
Lower
Device settingslanguage, screen size
Rule of thumb
If a feature reads it, askunless strictly necessary for what the user requested
A rough map of device data AI features tend to use. The higher up, the more care consent needs.

Designing AI features that use device data

Off by default. If a feature reads personal content from the device and is not strictly necessary for something the user explicitly asked for, ship it off and let users turn it on.

A specific, honest prompt. Ask at the moment the feature is first useful, in plain words: what will be read, why, whether anything leaves the device, and how to turn it off. The operating system's permission dialog is not enough on its own; it says what the app can access, not what your feature does with it. The same pattern works inside AI assistants: an in-context confirmation that says exactly what will happen before it does6.

Chat interface where an agent shows an interactive card summarising an action, listing what it will change and offering a Confirm button for the user
An in-context confirmation that explains exactly what will happen before it does: the pattern consent prompts for device data should follow. Source: Model Context Protocol Blog, MCP Apps · CC BY 4.0

Minimum access. Ask for the narrowest access that works: selected photos rather than the whole gallery, one folder rather than all files, approximate rather than precise location.

Records. Keep a record of what each user agreed to, when and in which version of the prompt, so you can show it later and re-ask when the feature changes.

Same rules for models. If device content is sent to a server or a third-party model, say so, minimise it and apply your normal data protection controls. If it is used to train models, that is a separate purpose needing its own basis.

On-device processing is good engineering, not a legal shortcut. In the EU, if your feature reaches into the phone, ask first, ask specifically and start switched off.

A quick review for existing features

  1. ListEvery feature that reads data from users' devices
  2. ClassifyWhat it reads and whether it is strictly necessary for a user request
  3. Check defaultsAnything not strictly necessary should start off
  4. Check promptsIs consent specific, honest and separate?
  5. MinimiseNarrow access to what the feature needs
  6. RecordStore consent with prompt version and date
A review most teams can complete in a short sprint, with legal advice on the edge cases.

CTO take

This fine is about consent and defaults, not about whether the feature was useful. As products add AI that looks for more context, the easiest way to stay on the right side of EU rules is to treat device content as something users lend you deliberately, not something you collect because the permission dialog allowed it. Product, engineering and legal should review device-reading features together, and new AI features should come with a consent design from day one. Meta has said it will appeal, so the decision may change; the design principles will not. Read alongside the 16 November tech news, the data protection programme case study and the Europe page.

How I can help

I help product and engineering teams design AI features that respect privacy rules by default, from consent flows to data minimisation. I am based in India, the United States and the UAE, working globally. See AI architecture, where I work, or analyse your challenge and bring the result to a call.

References
  1. 1
  2. 2
  3. 3
  4. 4
  5. 5
  6. 6
    MCP Apps: Extending servers with interactive user interfacesModel Context Protocol Blog · 21 Nov 2025
FAQ

Common questions

Why did the Garante fine Meta?

Reporting says the Garante found that the Facebook Android app's gallery suggestions feature, enabled by default, accessed photos, videos and metadata on users' devices without specific, informed consent, in breach of Italy's implementation of the ePrivacy Directive. Meta said it will appeal.

Does processing data only on the device avoid the consent requirement?

Not necessarily. The EU rule on accessing information in a user's device applies to gaining access to that information, and the EDPB's guidelines on its technical scope discuss local processing as a use case. Product teams should get legal advice rather than assume on-device processing is exempt.

How should AI features that use device data be designed in the EU?

Off by default, with a specific consent prompt that explains what is accessed and why, a clear way to turn it off, the minimum access needed, and records of consent and settings.

Keep reading

Related

ServiceAI & LLM architectureI design and build AI systems that leadership can trust: grounded in your data, measured by evaluations, governed by your rules and priced so they make sense at scale.→ServiceFractional CTO and AI architect for European companiesTechnology leadership for EU startups, scale-ups and enterprises, with AI and data platforms built for the EU AI Act, GDPR and the resilience rules that now apply to financial services and critical sectors.→InsightOne opt-out for AI training: what the shared robots.txt token changes, and what it does notOn 30 October, OpenAI, Google, Microsoft and Anthropic said they will honour a single robots.txt signal for opting out of AI training. For anyone who publishes content, that is simpler than maintaining a list of crawler names. But rollout is uneven, the companies treat AI answers differently from training, and the signal is voluntary. Here is how to decide whether to use it, and how to check it works.→InsightAI agents can now pay: what ChatGPT Wallet means for merchants, payment teams and anyone building agentsOn 13 November, OpenAI launched ChatGPT Wallet, letting its agents complete purchases for users with approval each time and tokenised card credentials. Card networks have been building agent payment schemes for months. Agent-initiated checkout is no longer a demo: it is traffic that will reach your store, your fraud rules and your support team. Here is how to prepare, whether you sell online or build agents yourself.→
Open to advisory, fractional CTO & select builds

Let's talk about what you're building.

The fastest route is a free 30-minute call. Prefer to write? Send a message and I'll reply personally within two business days.

Start a conversation

How would you like to reach me?

Book a 30-min callFastest. Pick any open slot on my calendar.→ Write in GmailOpens a pre-filled draft in your browser.→ Use my email appOutlook, Apple Mail or your default client.→

I reply personally within two business days. Time zones from the Americas and Europe to the Middle East and Asia-Pacific.