We call them symptoms, problems at work, a draft of a resignation letter or a fight with a partner. AI assistants have earned a position of confidence and that’s why it’s important Know who else is joining the conversation.
A team of Spanish researchers went to see it.
Author, of IMDEA Network and Carlos III University of MadridAnalyzed from Spain in May 2026, nine participants: ChatGPT, Claude, Grok, DeepSeek, Gemini, Perplexity, Microsoft Copilot, Mistral and Meta AI.
They tested the web version and the Android appAnd their study, presented at the Privacy Enhancing Technologies Symposium (PETS), concluded something that doesn’t fit the trustworthy picture: Nine communicate with at least one third-party advertising or tracking service.
What comes from your chat and where does it go?
Six out of nine web versions (ChatGPT, Gemini, Claude, Grok, Perplexity and Mistral) and three out of eight Android apps Disclose conversation addresses, titles, messages or screenshots to third partiesOften alongside permanent user identifiers. To better understand this: an organization that is not the assistant’s organization may know that this conversation exists and sometimes what it is about.
Not everyone filters the same. Five web clients sent the address of the conversation and three, the title given by the assistant itself to summarize it. Gemini, for example, sends that header to Google Analytics. Confusion sent prompts to WingFi. And Grok is the most interesting case: According to the study, when sharing conversations, TikTok collects screenshots, and Meta and TikTok users’ last messages, as well as addresses and titles. Your chat links are accessible by default (with an option to disable it).
Total, author count 124 third-party domains Of which they attribute 44 organizations, 34 of them are advertising or tracking (Google Ads, Meta, TikTok or Twitter Analytics, among others). They make it clear that they do not want to assess whether there is legal non-compliance and that their figures are a floor measured at a specific point in time.
And it’s not the first time that a chat with an assistant ends up where it shouldn’t: in December 2025, it became known about browser extensions that copy conversations with ChatGPT, Claude or Gemini. The difference is here you don’t have to install anything.
Reject cookies: 80.8% are still active
The first thing anyone does is “reject all” and it does: 77.3% of web identifier transmissions to third parties measured by researchers Occurred only after accepting non-required cookies. What it doesn’t do is turn off trackers: According to the authors, after they were rejected, 80.8% of third-party trackers were still active.
To avoid engaging with accounts: The study itself paints another picture in another category, with four of the nine free services having trackers that continue to collect data with opt-out cookies.
Paying doesn’t fix it– Free and subscription accounts feature almost identical tracking infrastructure And there are assistants (Confusion, Claude and Grok) that allow you to use them without choosing anything in the cookie notice. The authors recommend rejecting unnecessary cookies, disabling chat sharing where possible, and using browsers like Brave, which block ads, analytics, and trackers by default, although they acknowledge that these are partial measures.
“We don’t have ads on Gemini”: Company response
Google, OpenAI and Microsoft, When asked about the research, they rejected it. Google says the report misunderstands how its products work: “We don’t have ads on Gemini”. OpenAI responded that it uses analytics tools to understand traffic and diagnose errors, and that the data is not used for advertising or shared with advertisers. Microsoft assures that, after examining the complaints, it found no violations of applicable law
Caution is in order, according to research: The authors notified the data protection authorities of the European Union and the United Kingdom on April 13, and xAI via Grok links on April 17.
D Spanish Data Protection Agency (AEPD) He cited the work on 27 May and promoted its presentation at the plenary session of the European Authority on 8 and 9 June. OpenAI updated ChatGPT’s privacy policy on August 15 to clearly mention third-party trackers, though the authors say they can’t confirm it was caused by their work. And as of September 10, Grok still had accessible links and no official response.
Confusion, for its part, stopped sending conversation addresses to trackers like Meta on April 3, possibly, the authors said, in response to a class-action lawsuit filed in the United States on March 31, which alleges that its prompts reached Google and Meta’s ad platforms.
ChatGPT already has ads in Spain: is it the same?
No. But it is understood that they are mixing: The study begins by assuming that providers like OpenAI adopt an advertising-based business model, and that traditional tracking extends to conversational AI. And ChatGPT ads arrive in Spain on August 24, in 31 European markets, for free users and those on Go plans.
According to OpenAI, Conversations are kept private from advertisersAnd advertisers in the European Economic Area receive only aggregate data, such as views or clicks. These are two different channels: the ad you see within chat, and the page or app with third-party trackers, which are measured in research.
Personalized advertising is not “primarily” available in Europe. It is worth paying attention to that adverb.
Picture | GPT image with editing
