A research team at Stony Brook University in the U.S. published a paper on arXiv on October 27, 2026, titled "Penny for Your Prompts: A Measurement Study of Contextual Leakage in AI Chat Platforms." It examines how what users type into AI chat services reaches outside advertising and analytics trackers. Of the 42 services studied, 21 sent conversation titles to trackers, according to the paper. The work is set to be presented at the IEEE Symposium on Security and Privacy (S&P) in May 2027.

Half of the 42 Services Sent Conversation Titles to Third Parties

The study covered 42 services in total: 25 general-purpose AI chat services and 17 task-specific services focused on health, legal and education topics. On May 6 and May 20, 2026, the researchers used each service from a vantage point in the eastern U.S. and recorded the network requests the browser sent out.

In 21 services, conversation titles were sent to third-party trackers. In nine, the prompts users typed were passed along verbatim, the paper says. And among the 13 general-purpose services whose conversation content leaked, 11 also sent the user's email address to trackers.

The paper lists which services were affected. Because the companies' responses could not be confirmed at the time of writing, this article does not name individual services.

540 Requests per Exchange, More Than a Fifth Going to Trackers

According to the paper, a single prompt exchange generates an average of 540 outbound requests, of which 118 (22%) go to known tracker domains. Across the whole study, the team found 153 distinct tracker domains, with about 13 used per session on average.

Tracker traffic is concentrated among a few companies. Google accounts for more than 40%, followed by Microsoft (12.2%), ByteDance (8.6%) and Meta (8.3%). On the general-purpose services, these four companies received more than 70% of tracker traffic, the researchers say.

ItemResult
Services studied42 (25 general-purpose, 17 task-specific)
Services leaking conversation titles to trackers21
Services leaking verbatim prompts9
General-purpose services also leaking email11 of the 13 with leaked conversations
Requests per exchange540 on average (118 to trackers)

The Leak Runs Through the Page Title

The main path the paper identifies is the page title shown in the browser. Many AI chat services treat each conversation as a separate page and automatically give it a title based on the user's prompt. Tracker scripts read that title (document.title or the title element) for analytics, so a summary of the conversation ends up being sent outside.

For email addresses, tracker code attached to the registration form pulled out the value the user entered, according to the paper. In both cases, data collection that trackers routinely perform on ordinary websites overlaps with the way AI chat services are built.

Health and Legal Advice Services Were Also Covered

The study included task-specific services used for health, legal and education questions. On these services, prompts are likely to contain things users would rather keep private, such as symptoms of an illness or a legal dispute. Since titles are generated from prompts, the possibility that those summaries reach outside parties is something users cannot ignore.

Note that the measurements reflect conditions as of May 2026. Whether each service has since changed its tracker setup could not be confirmed at the time of writing.

Our Take: Conversation Titles Deserve the Same Care as Personal Data

What stands out to us is that while AI chat providers pay attention to how conversation content is handled, an inconspicuous path like the page title was overlooked. A title is only a short line at the top of the screen, but because it summarizes what the user asked, it should carry information close to the conversation itself.

That said, removing trackers would leave services unable to measure usage, so it is hard to imagine every provider dropping them right away. Replacing titles with generic text or removing analytics scripts from chat screens seem like realistic steps.

Still, users can do something too. Basic precautions, such as not putting names, medical conditions or other sensitive information into prompts and using the browser's tracking protection, are worth revisiting now.

Summary

A Stony Brook University study found that, of 42 AI chat services, 21 sent conversation titles and nine sent the prompts themselves to outside trackers. The findings directly concern anyone who regularly types personal questions into AI chat services. The paper is scheduled to be presented at IEEE S&P in May 2027, and how each service responds will be the next thing to watch.

Source: Penny for Your Prompts: A Measurement Study of Contextual Leakage in AI Chat Platforms (arXiv, paper by Stony Brook University researchers)