Researchers who tested AI assistants found that some shared chat links, titles or conversation text with advertising and analytics firms. The findings put a practical point on AI privacy: a conversation can travel further than the person who typed it expects.
Watch Desk analysis
What happened
A report by news.lavx.hu, citing Notebookcheck’s coverage of the research, says researchers at Madrid’s IMDEA Networks Institute tested nine AI assistants in browsers and eight Android apps in May 2026. All nine web services contacted at least one advertising or analytics company. The researchers counted 44 recipient firms; six web versions and three apps sent some conversation-related data to third parties.
The report says the material included shared-chat links, generated titles, screenshots, prompts and messages. On Grok shared pages, it says, screenshots and the latest prompt were sent to Google, Meta and TikTok. Perplexity’s Android app reportedly sent typed text to the testing service Wingify. The researchers said they had no proof that recipients read the conversations.
Why it matters
A shared link can expose more than its title: the report says anyone with the link could open the conversation on some services. And rejecting cookies was not a complete answer. In the study, 80.8% of trackers remained active after rejection, according to the report.
There are practical steps for users. The researchers’ recommendations include setting Perplexity threads to private and deleting unwanted shared links in ChatGPT, Claude, Gemini and Grok. Deleting a link can stop new visits, but it cannot recall copies or screenshots already made.
Our read
This is a useful warning about the gap between “I shared a chat” and “I know where its contents may go”. Treat shared links as public-facing, check your sharing settings, and avoid putting sensitive details into a conversation you plan to share. Cookie buttons deserve less credit than their confident little labels suggest.
What to watch
- Whether the full study explains the tests and data flows in enough detail for other researchers to reproduce them.
- Whether the named services change their sharing defaults or disclosures.
- Whether data-protection authorities take further action after receiving the researchers’ findings.
Discussion spark: Should AI services be allowed to send conversation-related data to analytics firms by default, or should users have to opt in first?
Sources and evidence
- ChatGPT, Claude and Grok chats reach trackers, study finds – news.lavx.hu (19 October 2026, 08:18 UTC)
Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.