In June 2017, Google made a promise: the content of consumer Gmail inboxes "will not be used or scanned for any ads personalization". In May 2021, at its I/O conference, the company repeated that it never uses content from Gmail, Photos and Drive for ads. Its privacy policy carried a simple line: "We don't show you personalized ads based on your content from Drive, Gmail, or Photos."
On 16 October 2026, that line gained a continuation. It now reads: "…unless you choose to make data from these services available to other Google services that support personalized ads."
At first glance it's a minor legal edit. In fact it shows how AI assistants are changing the very bargain we make about privacy — and why the most valuable thing in our data is not the emails themselves, but the conclusions that can be drawn from them.
What exactly changed
The new policy applies to people who have connected Gmail, Calendar, Photos or Drive to Personal Intelligence, a feature of Google's AI Mode in Search and of Gemini. It lets the AI assistant answer questions using your personal data: find a ticket in your inbox, remind you of a calendar event, suggest something based on your photos.
The policy now says plainly that Search "may process some of the data that you allow Search to access to personalize content, including ads". According to Techlicious, Google stresses that ads won't be based "directly" on the contents of your inbox or photo library. But Search may use summaries, excerpts and inferences drawn from that data.
A few important details:
- It is not on by default. Connecting these services is a separate step that users take themselves. The feature is only available on personal accounts for people aged 18 and over.
- Ads in AI Mode are still being tested. Trade publications report that Google has confirmed ad tests in AI Mode for Personal Intelligence. That makes this early deployment, not established practice.
- Sensitive categories remain off-limits. Google still doesn't allow ads to be targeted by race, religion, sexual orientation or health.
- Data doesn't disappear instantly. If you disconnect the services, what's already stored in your search history can remain until you delete it.
Why the conclusions matter more than the emails
Older privacy debates focused on who reads your email. But for advertising, as for many other purposes, the emails themselves aren't really needed. What's valuable are the inferences: you're planning a move, expecting a baby, looking for a job, flying often, recently bought a car.
An AI assistant with access to your inbox draws conclusions like these naturally — otherwise it couldn't help you. The new policy opens a route by which some of those conclusions can reach advertising.
How much can be learned about a person from their digital traces was shown by researchers long before today's AI assistants.
What science knows about inferences from data
2013: Facebook likes. Researchers at Cambridge analysed data from about 58,000 volunteers who shared their "likes", demographic details and psychological test results. A simple algorithm, using the likes alone, correctly distinguished gay from straight men 88% of the time, and Democrats from Republicans 85% of the time. It also predicted personality traits, religious views, substance use and even whether a person's parents had separated. The paper appeared in the peer-reviewed journal PNAS.
2024: language models. A team at ETH Zurich tested what large language models can "guess" from ordinary Reddit posts. The models identified location, income, sex and other personal attributes with up to 85% accuracy on the first guess and up to 95% within three guesses. It was roughly 100 times cheaper and 240 times faster than having humans do the same. Anonymising the text offered little protection. The work was presented at ICLR, a leading machine learning conference.
Both studies have limits: the 2013 volunteers weren't representative of the whole population, and Reddit posts aren't the same as private email. But the general finding is robust and supported by different methods: indirect data can predict, with reasonable accuracy, things a person has never stated directly.
Inferences serve more than advertising
What follows is interpretation, not established fact about Google.
The inference "this person is probably looking for a new job" is valuable to more than advertisers. All personalisation is built on inferences like it: which news to show first, which products to recommend, which videos to suggest. The better a system "understands" a person, the more precisely it can tailor everything it shows them.
There's an interesting tension here, too. In the 2013 study, the algorithm was best at predicting exactly the characteristics Google forbids in ad targeting: orientation, political and religious views. A ban on targeting by sensitive categories limits how inferences are used. It doesn't stop a model from making them.
The view from Europe
In the European Union, two laws shape this situation. The General Data Protection Regulation (GDPR) governs the processing of personal data and profiling. The Digital Markets Act (DMA) bars the largest platforms — "gatekeepers", which include Google — from combining personal data across their services without the user's specific consent.
In Europe, the new exception currently has nothing to attach to. According to Google's own page, the Personal Intelligence beta is available globally except in the European Economic Area, Switzerland, the UK and Nigeria. The same regions were excluded when the feature reached Android Auto on 20 October. Google has not explained why.
While preparing this article, we checked the settings of a Google AI Pro account in Poland. Its list of linked services includes Search, Chrome, Android, Maps, Google Play, Translate, Lens, News and Shopping. Gmail and Google Photos are not among them.
So as of 25 October, this account cannot connect email and photos to the AI assistant — which means the new ad exception doesn't apply. Google doesn't say whether this is linked to European law.
Outside Europe the picture is different. Personal Intelligence started in January 2026 for a limited number of users in the US. In April Google opened it to Gemini subscribers worldwide, and in May it brought the feature to AI Mode in more countries.
What we still don't know
- Which inferences are stored, and for how long. The policy describes this only in general terms.
- What the ads will look like. Tests are ongoing; no public data on the results.
- Whether anyone will check these systems from outside. There is no independent audit yet of how inferences from email affect advertising.
- Where it will apply. Regional restrictions change and affect different products differently.
Three plausible scenarios
The new normal. Other companies with AI assistants adopt similar rules, and "help in exchange for inferences" becomes an ordinary bargain, just as "free email in exchange for ads" once did.
The law draws the line. Regulators, above all in the EU, require separate consent specifically for using AI inferences in advertising, or ban it. Different regions then get different versions of the products.
Users decide. Some people deliberately don't connect personal services to AI assistants, or switch to services that process data on the device itself. Privacy becomes a competitive advantage.
What you can do
If you use Personal Intelligence and don't want inferences from your data to affect ads, Techlicious suggests:
- in your Google Account, open Search Services Personalization and, under connected content apps, disconnect Google Workspace and Google Photos;
- in the same place, you can turn off personalised recommendations;
- in My Activity, find Search Services History and delete it — disconnecting doesn't erase it;
- note that turning off personalised ads in My Ad Center, according to the same publication, doesn't necessarily stop this data being used to personalise the AI.
Menu names may differ by interface language and region. In the EU, it's worth simply looking at the list of linked services in your Google Account: it shows which services share data with each other, and that's also where you can switch the sharing off.
And the main rule, whatever your settings: before giving an AI assistant access to your email or photos, ask yourself what conclusions could be drawn about you from them — and whether you're comfortable with those conclusions being stored somewhere.
Questions for discussion
- Which would you find more acceptable: a system that reads your emails but remembers nothing, or one that never sees your emails but knows the conclusions drawn from them?
- Do you think "a convenient assistant in exchange for ads based on your data" is a fair deal? What does your answer depend on?
- If a model can guess something you've never said, does it matter that it "only guessed"?
Connectez-vous ou Inscrivez-vous pour rejoindre la discussion.