On September 28, 2026, findings from a study by researchers at IMDEA Networks and Universidad Carlos III de Madrid were reported: direct conversation data had reached third-party trackers in six web assistant versions and three mobile apps. The same study distinguished those cases from three assistants that sent device information and chat metadata.
The researchers tested nine assistants using health-related prompts and documents containing medical test results, and varied privacy settings. The study was reportedly accepted at PETS, the Privacy Enhancing Technologies Symposium.
Which assistants and data types were named?
The reported results separate assistants that sent direct conversation data from those associated with device information and chat metadata. The web and mobile lists overlap, so they are distinct test groupings rather than nine additional services.
| Surface or group | Assistants named | Reported data type |
| Web versions | ChatGPT, Gemini, Claude, Grok, Perplexity and Mistral | Direct conversation data sent through advertising trackers |
| Mobile apps | ChatGPT, Claude and Perplexity | Direct conversation data sent through advertising trackers |
| Other assistants in the nine-assistant study | DeepSeek, Microsoft Copilot and Meta AI | Device information and chat metadata |
The reported categories of conversation-related data included chat titles, questions, uploaded documents, screenshots and, in some cases, shared-chat links. The findings do not assign every category to every service. Chat metadata describes information about a conversation; it is different from the conversation’s actual messages.
Grok links and tracker access
For Grok’s web version, the researchers reportedly found conversation URLs being sent automatically to seven external advertising and analytics networks. Meta, TikTok, Google Ads and Twitter Analytics were among the networks named. The requests could also carry cookies and user identifiers.
In tests of shared Grok chats, the last user question was reportedly sent to Meta and TikTok, and TikTok received a screenshot of a chat. Follow-up tests recorded up to 70 automatic tracker accesses to Grok conversations from 14 countries.
These findings concern data transmission and, in some cases, tracker access to shared links. They do not establish that a person read every chat or that a particular conversation triggered an ad.
What happened when privacy settings changed?
The study’s reported results said that as many as 80.8% of third-party trackers remained active under privacy-protective settings, including rejecting cookies or using incognito mode.
A separate IMDEA Networks disclosure dated May 6, 2026, described an earlier analysis of the web versions of ChatGPT, Claude, Grok and Perplexity. That analysis examined network traffic under different combinations of sign-in status, cookie consent, account tier and privacy mode. Rejecting nonessential cookies blocked some observed flows on Claude and Grok, while other flows persisted.
Company responses and regulatory review
Google said Gemini has no ads and described Google Analytics as a tool for monitoring technical performance, detecting outages and maintaining reliability. OpenAI said its analytics tools supported traffic analysis, diagnostics and performance monitoring, and said the data was not used for advertising or shared or sold to advertisers. Microsoft said it found no violation of applicable law or its privacy statement.
The Spanish Data Protection Agency reportedly referred the findings to the European Data Protection Board for examination.
