Both Google and Apple are cramming new AI features into their phones and other devices, and neither company has offered clear ways to control which apps those AI systems can access. Recent issues around WhatsApp on both Android and iPhone demonstrate how these interactions can go sideways, risking revealing chat conversations beyond what you intend. Users deserve better controls and clearer documentation around what these AI features can access.

After diving into how Google Gemini and Apple Intelligence (and in some cases Siri) currently work, we didn’t always find clear answers to questions about how data is stored, who has access, and what it can be used for.

At a high level, when you compose a message with these tools, the companies can usually see the contents of those messages and receive at least a temporary copy of the text on their servers.

When receiving messages, things get trickier. When you use an AI like Gemini or a feature like Apple Intelligence to summarize or read notifications, we believe companies should be doing that content processing on-device. But poor documentation and weak guardrails create issues that have lead us deep into documentation rabbit holes and still fail to clarify the privacy practices as clearly as we’d like.

We’ll dig into the specifics below as well as potential solutions we’d like to see Apple, Google, and other device-makers implement, but first things first, here’s what you can do right now to control access:

Control AI Access to Secure Chat on Android and iOS

Here are some steps you can take to control access if you want nothing to do with the device-level AI features’ integration and don’t want to risk accidentally sharing the text of a message outside of the app you’re using.

How to Check and Limit What Gemini Can Access

If you’re using Gemini on your Android phone, it’s a good time to review your settings to ensure things are set up how you want. Here’s how to check each of the relevant settings:

• Disable Gemini App Activity: Gemini App Activity is a history Google stores of all your interactions with Gemini. It’s enabled by default. To disable it, open Gemini (depending on your phone model, you may or may not even have the Google Gemini app installed. If you don’t have it installed, you don’t really need to worry about any of this). Tap your profile picture > Gemini Apps Activity, then change the toggle to either “Turn off,” or “Turn off and delete activity” if you want to delete previous conversations. If the option reads “Turn on,” then Gemini Apps Activity is already turned off. 
• Control app and notification access: You can control which apps Gemini can access by tapping your profile picture > Apps, then scrolling down and disabling the toggle next to any apps you do not want Gemini to access. If you do not want Gemini to potentially access the content that appears in notifications, open the Settings app and revoke notification access from the Google app.
• Delete the Gemini app: Depending on your phone model, you might be able to delete the Gemini app and revert to using Google Assistant instead. You can do so by long-pressing the Gemini app and selecting the option to delete. 

How to Check and Limit what Apple Intelligence and Siri Can Access

Similarly, there are a few things you can do to clamp down on what Apple Intelligence and Siri can do: 

• Disable the “Use with Siri Requests” option: If you want to continue using Siri, but don’t want to accidentally use it to send messages through secure messaging apps, like WhatsApp, then you can disable that feature by opening Settings > Apps > [app name], and disabling “Use with Siri Requests,” which turns off the ability to compose messages with Siri and send them through that app.
• Disable Apple Intelligence entirely: Apple Intelligence is an all-or-nothing setting on iPhones, so if you want to avoid any potential issues your only option is to turn it off completely. To do so, open Settings > Apple Intelligence & Siri, and disable “Apple Intelligence” (you will only see this option if your device supports Apple Intelligence, if it doesn’t, the menu will only be for “Siri”). You can also disable certain features, like “writing tools,” using Screen Time restrictions. Siri can’t be universally turned off in the same way, though you can turn off the options under “Talk to Siri” to make it so you can’t speak to it. 

For more information about cutting off AI access at different levels in other apps, this Consumer Reports article covers other platforms and services.

Why It Matters 

Sending Messages Has Different Privacy Concerns than Receiving Them

Let’s start with a look at how Google and Apple integrate their AI systems into message composition, using WhatsApp as an example.

Google Gemini and WhatsApp

On Android, you can optionally link WhatsApp and Gemini together so you can then initiate various actions for sending messages from the Gemini app, like “Call Mom on WhatsApp” or “Text Jason on WhatsApp that we need to cancel our secret meeting, but make it a haiku.” This feature raised red flags for users concerned about privacy.

By default, everything you do in Gemini is stored in the “Gemini Apps Activity,” where messages are stored forever, subject to human review, and are used to train Google’s products. So, unless you change it, when you use Gemini to compose and send a message in WhatsApp then the message you composed is visible to Google.

If you turn the activity off, interactions are still stored for 72 hours. Google’s documentation claims that even though messages are stored, those conversations aren’t reviewed or used to improve Google machine learning technologies, though that appears to be an internal policy choice with no technical limits preventing Google from accessing those messages.

The simplicity of invoking Gemini to compose and send a message may lead to a false sense of privacy. Notably, other secure messaging apps, like Signal, do not offer this Gemini integration.

For comparison’s sake, let’s see how this works with Apple devices.

Siri and WhatsApp

The closest comparison to this process on iOS is to use Siri, which it is claimed, will eventually be a part of Apple Intelligence. Currently, Apple’s AI message composition tools are not available for third-party apps like Signal and WhatsApp.

According to its privacy policy, when you dictate a message through Siri to send to WhatsApp (or anywhere else), the message, including metadata like the recipient phone number and other identifiers, is sent to Apple’s servers. This was confirmed by researchers to include the text of messages sent to WhatsApp. When you use Siri to compose a WhatsApp message, the message gets routed to both Apple and WhatsApp. Apple claims it does not store this transcript unless you’ve opted into “Improve Siri and Dictation.” WhatsApp defers to Apple’s support for data handling concerns. This is similar to how Google handles speech-to-text prompts.

In response to that research, Apple said this was expected behavior with an app that uses SiriKit—the extension that allows third-party apps to integrate with Siri—like WhatsApp does.

Both Siri and Apple Intelligence can sometimes run locally on-device, and other times need to rely on Apple-managed cloud servers to complete requests. Apple Intelligence can use the company’s Private Cloud Compute, but Siri doesn’t have a similar feature.

The ambiguity around where data goes makes it overly difficult to decide on whether you are comfortable with the sort of privacy trade-off that using features like Siri or Apple Intelligence might entail.

How Receiving Messages Works

Sending encrypted messages is just one half of the privacy puzzle. What happens on the receiving end matters too. 

Google Gemini

By default, the Gemini app doesn’t have access to the text inside secure messaging apps or to notifications. But you can grant access to notifications using the Utilities app. Utilities can read, summarize, and reply to notifications, including in WhatsApp and Signal (it can also read notifications in headphones).

We could not find anything in Google’s Utilities documentation that clarifies what information is collected, stored, or sent to Google from these notifications. When we reached out to Google, the company responded that it “builds technical data protections that safeguard user data, uses data responsibly, and provides users with tools to control their Gemini experience.” Which means Google has no technical limitation around accessing the text from notifications if you’ve enabled the feature in the Utilities app. This could open up any notifications routed through the Utilities app to the Gemini app to be accessed internally or from third-parties. Google needs to publicly make its data handling explicit in its documentation.

If you use encrypted communications apps and have granted access to notifications, then it is worth considering disabling that feature or controlling what’s visible in your notifications on an app-level.

Apple Intelligence

Apple is more clear about how it handles this sort of notification access.

Siri can read and reply to messages with the “Announce Notifications” feature. With this enabled, Siri can read notifications out loud on select headphones or via CarPlay. In a press release, Apple states, “When a user talks or types to Siri, their request is processed on device whenever possible. For example, when a user asks Siri to read unread messages… the processing is done on the user’s device. The contents of the messages aren’t transmitted to Apple servers, because that isn’t necessary to fulfill the request.”

Apple Intelligence can summarize notifications from any app that you’ve enabled notifications on. Apple is clear that these summaries are generated on your device, “when Apple Intelligence provides you with preview summaries of your emails, messages, and notifications, these summaries are generated by on-device models.” This means there should be no risk that the text of notifications from apps like WhatsApp or Signal get sent to Apple’s servers just to summarize them.

New AI Features Must Come With Strong User Controls

As more device-makers cram AI features into their devices, the more necessary it is for us to have clear and simple controls over what personal data these features can access on our devices. If users do not have control over when a text leaves a device for any sort of AI processing—whether that’s to a “private” cloud or not—it erodes our privacy and potentially threatens the foundations of end-to-end encrypted communications.

Per-app AI Permissions

Google, Apple, and other device makers should add a device-level AI permission, just like they do for other potentially invasive privacy features, like location sharing, to their phones. You should be able to tell the operating system’s AI to not access an app, even if that comes at the “cost” of missing out on some features. The setting should be straightforward and easy to understand in ways the Gemini an Apple Intelligence controls currently are not.

Offer On-Device-Only Modes

Device-makers should offer an “on-device only” mode for those interested in using some features without having to try to figure out what happens on device or on the cloud. Samsung offers this, and both Google and Apple would benefit from a similar option.

Improve Documentation

Both Google and Apple should improve their documentation about how these features interact with various apps. Apple doesn’t seem to clarify notification processing privacy anywhere outside of a press release, and we couldn’t find anything about Google’s Utilities privacy at all. We appreciate tools like Gemini Apps Activity as a way to audit what the company collects, but vague information like “Prompted a Communications query” is only useful if there’s an explanation somewhere about what that means.

The current user options are not enough. It’s clear that the AI features device-makers add come with significant confusion about their privacy implications, and it’s time to push back and demand better controls. The privacy problems introduced alongside new AI features should be taken seriously, and remedies should be offered to both users and developers who want real, transparent safeguards over how a company accesses their private data and communications.

Reposted from the EFF’s Deeplinks blog.

The European Union Council is once again debating its controversial message scanning proposal, aka “Chat Control,” that would lead to the scanning of private conversations of billions of people.

Chat Control, which EFF has strongly opposed since it was first introduced in 2022, keeps being mildly tweaked and pushed by one Council presidency after another.

Chat Control is a dangerous legislative proposal that would make it mandatory for service providers, including end-to-end encrypted communication and storage services, to scan all communications and files to detect “abusive material.” This would happen through a method called client-side scanning, which scans for specific content on a device before it’s sent. In practice, Chat Control is chat surveillance and functions by having access to everything on a device with indiscriminate monitoring of everything. In a memo, the Danish Presidency claimed this does not break end-to-end encryption.

This is absurd.

We have written extensively that client-side scanning fundamentally undermines end-to-end encryption, and obliterates our right to private spaces. If the government has access to one of the “ends” of an end-to-end encrypted communication, that communication is no longer safe and secure. Pursuing this approach is dangerous for everyone, but is especially perilous for journalists, whistleblowers, activists, lawyers, and human rights workers.

If passed, Chat Control would undermine the privacy promises of end-to-end encrypted communication tools, like Signal and WhatsApp. The proposal is so dangerous that Signal has stated it would pull its app out of the EU if Chat Control is passed. Proponents even seem to realize how dangerous this is, because state communications are exempt from this scanning in the latest compromise proposal.

This doesn’t just affect people in the EU, it affects everyone around the world, including in the United States. If platforms decide to stay in the EU, they would be forced to scan the conversation of everyone in the EU. If you’re not in the EU, but you chat with someone who is, then your privacy is compromised too. Passing this proposal would pave the way for authoritarian and tyrannical governments around the world to follow suit with their own demands for access to encrypted communication apps.

Even if you take it in good faith that the government would never do anything wrong with this power, events like Salt Typhoon show there’s no such thing as a system that’s only for the “good guys.”

Despite strong opposition, Denmark is pushing forward and taking its current proposal to the Justice and Home Affairs Council meeting on October 14th.

We urge the Danish Presidency to drop its push for scanning our private communication and consider fundamental rights concerns. Any draft that compromises end-to-end encryption and permits scanning of our private communication should be blocked or voted down.

Phones and laptops must work for the users who own them, not act as “bugs in our pockets” in the service of governments, foreign or domestic. The mass scanning of everything on our devices is invasive, untenable, and must be rejected.

Republished from the EFF’s Deeplinks blog.

At least Florida’s SB 868/HB 743, “Social Media Use By Minors” bill isn’t beating around the bush when it states that it would require “social media platforms to provide a mechanism to decrypt end-to-end encryption when law enforcement obtains a subpoena.” Usually these sorts of sweeping mandates are hidden behind smoke and mirrors, but this time it’s out in the open: Florida wants a backdoor into any end-to-end encrypted social media platforms that allow accounts for minors. This would likely lead to companies not offering end-to-end encryption to minors at all, making them less safe online.

Encryption is the best tool we have to protect our communication online. It’s just as important for young people as it is for everyone else, and the idea that Florida can “protect” minors by making them less safe is dangerous and dumb.

The bill is not only privacy-invasive, it’s also asking for the impossible. As breaches like Salt Typhoon demonstrate, you cannot provide a backdoor for just the “good guys,” and you certainly cannot do so for just a subset of users under a specific age. After all, minors are likely speaking to their parents and other family members and friends, and they deserve the same sorts of privacy for those conversations as anyone else. Whether social media companies provide “a mechanism to decrypt end-to-end encryption” or choose not to provide end-to-end encryption to minors at all, there’s no way that doesn’t harm the privacy of everyone.

If this all sounds familiar, that’s because we saw a similar attempt from an Attorney General in Nevada last year. Then, like now, the reasoning is that law enforcement needs access to these messages during criminal investigations. But this doesn’t hold true in practice.

In our amicus brief in Nevada, we point out that there are solid arguments that “content oblivious” investigation methods—like user reporting— are “considered more useful than monitoring the contents of users’ communications when it comes to detecting nearly every kind of online abuse.” That remains just as true in Florida today.

Law enforcement can and does already conduct plenty of investigations involving encrypted messages, and even with end-to-end encryption, law enforcement can potentially access the contents of most messages on the sender or receiver’s devices, particularly when they have access to the physical device. The bill also includes measures prohibiting minors from accessing any sort of ephemeral messaging features, like view once options or disappearing messages. But even with those features, users can still report messages or save them. Targeting specific features does nothing to protect the security of minors, but it would potentially harm the privacy of everyone.

SB 868/HB 743 radically expands the scope of Florida’s social media law HB 3, which passed last year and itself has not yet been fully implemented as it currently faces lawsuits challenging its constitutionality. The state was immediately sued after the law’s passage, with challengers arguing the law is an unconstitutional restriction of protected free speech. That lawsuit is ongoing—and it should be a warning sign. Florida should stop coming up with bad ideas that can’t be implemented.

Weakening encryption to the point of being useless is not an option. Minors, as well as those around them, deserve the right to speak privately without law enforcement listening in. Florida lawmakers must reject this bill. Instead of playing politics with kids’ privacy, they should focus on real, workable protections—like improving consumer privacy laws to protect young people and adults alike, and improving digital literacy in schools.

Reposted from the EFF’s Deeplinks blog.