Gemini Screen Reading: 11 Taps to Turn Off Across 3 Menus

👁 118 reads

The change Google is rolling out on Android phones looks from the outside like swapping one app for another. Google Assistant is leaving, Gemini is arriving. But the difference runs deeper than the logo: the old assistant worked like an advanced remote control - it listened to a command, performed an action, and finished. The new model works like a participant in what you’re doing. It can look at your screen content, summarize an email you have open, read a PDF in your browser, suggest a reply in a messenger. To do that, it needs to send that content somewhere.

The security blog Niebezpiecznik announced a livestream with the co-founder of Proget, asking the question directly: when does information from your phone leave the device and reach the model provider. That’s a good question, so we decided to break it down into measurable parts. Not impressions, but the number of switches and the number of screen taps.

Eleven taps, three different settings screens, zero single switch

We sat down with a phone running Android 15 and went through the entire path of limiting AI access to screen content. We counted eleven taps spread across three unrelated locations - and that’s in a scenario where we knew exactly what we were looking for. The first location is the Gemini app and its activity settings: that’s where you turn off conversation recording and consent to human review. The second is system settings and the digital assistant section, where the switch for using screen context lives. The third is your Google account in a browser or account settings, because some personalization options don’t live on the phone, but on the service side.

The number itself isn’t shocking. The proportion is. Enabling these same permissions takes one tap - one consent button in the wizard that appears when you first launch Gemini after the update. The ratio is therefore 11:1. For every move needed to grant access, you need to make eleven to revoke it. This isn’t a design accident, it’s standard interface asymmetry: consent is in the user’s path, refusal is in a corner.

The second thing we counted concerns the default state. When transitioning from Assistant to Gemini, a user who simply clicked “next” has three out of four data collection mechanisms enabled: recording conversation history with the model, the ability to use fragments of those conversations to improve the service, and access to screen context on demand. The only thing disabled by default is constant, continuous monitoring without invocation. In other words: by default AI doesn’t watch all the time, but by default it has the right to look when you call it, and by default it records what it saw.

Old Assistant saw commands, Gemini sees documents

To show the scale of the difference, we compared the scope of data available to both solutions. This isn’t about the list of permissions in the app manifest, because that looks similar, but about what actually leaves the device during typical use.

The difference comes down to one thing: the unit of data is changing. Assistant processed a query, meaning a few words you consciously spoke yourself. Gemini processes a view, meaning everything currently in frame - including what you didn’t intend to show anyone, because it was in the same window. An account number next to a transfer summary. A patient’s name above a test result. A contract fragment below the paragraph you asked about.

The business cost: $3 per phone per month to have control

There’s another side to this story, purely about cost, and here the number makes the biggest impression. An ordinary user can go through those eleven taps and be done with it. An organization that issued a hundred phones to employees doesn’t have that option - it won’t go through three settings screens a hundred times and won’t verify whether someone accidentally clicked “next” after another update. It needs an MDM system, meaning centralized device management, which enforces policy from the top down.

We calculated what this costs at typical market rates for MDM licenses, which range from $2-4 per device per month. That gives $2 to $4 per phone per month, or about $3 in the middle of the range. For a hundred devices that’s $300 per month and $3,600 per year. For five hundred phones - $18,000 per year. This is a real amount that companies will start entering in their budgets not because of a new hacking threat, but because of a change in the default setting of an app they didn’t install.

It’s worth calculating this another way. One mid-range business phone costs around $475-660, usually amortized over two or three years. An MDM license over three years at $3 per month is $108, or about 20 percent of the device cost itself. The cost of controlling what the phone sends externally is therefore one-fifth the cost of the phone.

Apple took a different path and that’s also measurable

On iPhones the architecture looks different, though it shouldn’t be idealized. Apple split processing into two levels: some tasks are performed by a model running locally on the phone’s chip, some go to server infrastructure, and the most complex queries can be passed to an external provider after separate confirmation. That last confirmation is key, because it means an additional, visible step before sending data outside the ecosystem.

From our counting, on iOS full limitation of AI features requires six or seven taps, because most switches sit in one settings section instead of being scattered across three locations. That’s less than eleven, but not because of greater privacy care - Apple simply controls the entire system and doesn’t have to split settings between phone, app, and cloud account. With the same phone, Apple’s AI features work in a narrower scope in Poland and in Polish than in English, which paradoxically means less data going out - not from user choice, but from lack of language support.

GDPR: consent to something an employee couldn’t refuse is questionable

Here we enter ground where European readers have a stronger position than users in the United States. Article 4 point 11 of GDPR defines consent as freely given, specific, informed and unambiguous indication of will. Clicking “next” in a wizard that appears after an automatic update and doesn’t offer an equivalent refusal option is hard to consider meeting that standard. The European Data Protection Board has pointed this out when evaluating interface design, writing about patterns that steer user choice.

“Interfaces should not be designed in a way that influences users to make decisions that are not in their interest in terms of data protection, for example by highlighting options that are less protective of privacy.” - European Data Protection Board, guidelines 03/2022 on dark patterns

For a company the conclusion is concrete and costly. If a work phone contains customer data, patient data, job candidate data or anything the employer is the controller of, then sending a screen fragment with that data to an external model is data processing. A legal basis, a processing agreement and an entry in the activity register are needed. No employee click will handle that. Data protection authorities have repeatedly fined controllers precisely for lack of control over where data physically leaks - and here the mechanism is identical, only the channel isn’t a lost USB drive, but an AI assistant.

What to do today, before the next update resets everything

For a personal phone the sequence is as follows. Open Gemini, go into app activity settings and turn off conversation recording - this cuts off history accumulation and limits review of fragments by people verifying quality. Then go into system settings, find the section responsible for default apps and digital assistant, and revoke consent to use screen context. Finally look at your Google account settings in a browser, because that’s where personalization options invisible from the phone live. Write down somewhere the date of this operation, because with major system updates some switches return to default state and need to be checked again.

For a phone where work and private life coexist, the only sensible answer is profile separation. Android has a work profile that keeps company apps in a separate container - the assistant from the private profile shouldn’t see content from the other side. This isn’t absolute protection, but it moves the problem from “everything in one frame” to “two separate frames”, and that’s in practice the cheapest change you can make without buying a license for $3 per month.

Finally, something no number can measure, but which is most important in this story. For years we’ve learned to ask what permissions a given app has. That question stops being enough, because an AI assistant doesn’t need access to your gallery to see a photo - it’s enough that you open it and it looks at the screen. The boundary is no longer the permissions list, but what you have in front of your eyes at a given moment. The new habit of digital hygiene is therefore the question: do I want a third party looking at this with me.

Facts checked October 3, 2026. Sources: Niebezpiecznik - AI in phone can read your screen, EDPB, guidelines 03/2022, Gemini help documentation, iPhone user guide. Numbers regarding tap count and default switch states come from our own walkthrough of settings on a device and may vary depending on manufacturer overlay.

// read moreHow Many Screens Can Gemini See on Your Phone Daily? Over 2,600State of AIHow to Turn Off Gemini on Android: 4 Ways and One TrapGuidesCLAUDE.md and AGENTS.md: What I Use Across 42 ProjectsGuides

Kinetyra