Study: ad trackers see what you discuss with every second AI chatbot
Researchers at Stony Brook University examined 42 AI chat services, including ChatGPT, Claude and Gemini. On 21 of them, auto-generated chat titles ended up with third-party trackers, and on 9 the trackers even got users' prompts word for word. Nobody hacked anything. The trackers simply read the page title, just as they do on any other website.

What it's about
When you tell an AI chatbot about your symptoms, a fight with your landlord or something personal that's worrying you, you probably assume the conversation stays between you and the provider. A new study shows that often isn't true. Researchers at the State University of New York at Stony Brook (USA) measured how conversation content and personal data flow out of AI chats to outside companies through embedded trackers. Trackers are small third-party programs that websites embed for advertising and usage analytics.
The paper is titled “Penny for Your Prompts: A Measurement Study of Contextual Leakage in AI Chat Platforms.” Paschalis Bekos is the lead author, and the work has been accepted to IEEE S&P 2027, one of the top IT security conferences. A preprint is available on arXiv. The Japanese tech outlet ITmedia AI+ summarized the findings.
What was examined
The team tested 42 services in total:
- 25 general-purpose chatbots, including ChatGPT, Claude and Gemini
- 17 specialized services for health, law and education
The starting point: under the hood, many AI chats are ordinary websites. Like online stores or news sites, they embed third-party scripts for advertising and behavioral analytics. Users often don't realize this, because a chat feels more like a private conversation than a visit to a website.
The researchers didn't just record what data the browser sent out. They also closely monitored the browser's JavaScript engine, the part that runs a website's code. That let them trace the whole chain: which user action triggers a leak, which part of the page the information comes from, and where it ends up.
The numbers
The key findings:
| Finding | Result |
|---|---|
| Services passing chat titles to trackers | 21 of 42 |
| Of those: services passing prompts word for word | 9 |
| Leaking general-purpose chatbots that also send the email address | 11 of 13 |
| Leaking specialized services that also send the email address | 8 of 8 |
| Average browser requests per general-purpose chatbot | 540 |
| Of those, sent to trackers | 118 (approx. 22%) |
The chat title is the short heading many chatbots generate from your first question, such as “Rash after antibiotics” or “Contesting an eviction for owner move-in.” It sounds harmless, but it captures the core of what someone wanted to know.
The email address is even more sensitive. Almost every leaking service also sent the sign-up address to third parties. That could make it possible to link a conversation topic to a specific person. The specialized services for health, law and education did this without exception, even though they're exactly where you'd expect the most sensitive questions.
Who gets the data
For the general-purpose chatbots, the researchers worked out which companies owned the tracker requests. Google is far ahead:
| Company | Share of tracker traffic |
|---|---|
| 41.8% | |
| Microsoft | 12.2% |
| ByteDance | 8.6% |
| Meta | 8.3% |
The ITmedia report doesn't say which service sent which data to which tracker. So you can't tell from this source whether ChatGPT, Claude or Gemini are among the 21 leaking services. The names only describe the group that was studied.
Why it happens without any hack
The analysis shows the trackers aren't doing anything unusual. They don't attack anything or bypass any protections. They do what they do on millions of other websites: automatically read the page title, the text in the so-called title tag. It's what you see in the browser tab, and analytics tools normally use it to identify which page someone is viewing.
Many AI chatbots, however, write the auto-generated conversation summary into that very page title. When you send a prompt or switch to another chat, the tracker kicks in. It reads the title as an ordinary page name and sends it off, without “knowing” there's an intimate question inside. So this isn't the fault of individual ad companies. Two harmless-looking building blocks just don't fit together well.
The researchers found two other routes by which prompts leak word for word:
- Session replay: These tools record what a user does on a page, a bit like a screen recording for usability analysis. Along the way, they also capture the text on the page, including the prompt you typed.
- Prompts in the URL: Some services put the question into the web address. Since trackers almost always send the visited URL, the prompt goes along with it.
What you can do
According to ITmedia, the study doesn't give recommendations for users. But some general advice applies. Tracker blockers or privacy-focused browsers can suppress many of these scripts. Using the official apps instead of the web interface can also make a difference, though apps may have their own analytics tools. It's best not to enter very sensitive information, such as diagnoses, names or addresses, into online services whose data flows you don't know. If you need full confidentiality, you can run a language model locally on your own computer.
The bigger picture
The study matters because it exposes a problem hardly anyone is watching. The debate over privacy in AI chats usually focuses on whether providers use conversations for training. This is something else: ordinary web technology passes conversation content to third parties that have nothing to do with the AI service itself, above all to big advertising and analytics firms like Google, Microsoft, ByteDance and Meta.
This is especially serious for health and legal services. Depending on the country, strict rules govern how that data is handled, and in Europe it's likely to raise GDPR questions. According to the source, however, the study doesn't offer a legal assessment.
The good news: the problem should be easy to fix. Providers would only need to stop writing conversation summaries into the page title and prompts into the URL, or drop trackers from chat pages. It remains unclear which services are affected, whether providers were notified before publication, and whether they've already fixed things. The ITmedia report doesn't say, and no company statements are available. The measurements are also a snapshot in time. Websites change constantly, so the situation may look different today.
Sources
- ITmedia AI+ (JP) itmedia.co.jp
Ask the AI assistant
Questions about this article, installation or background? Our assistant knows this page and runs on our own machine, not in someone else's cloud.
The Assistant is an AI (Gemma 4) and can make mistakes – please check important details against the source. Questions are used only to generate an answer and are deleted after one hour. Please don't enter any personal data.
Comments
What do you think? Already tried it? Write it here, no sign-up needed.
Comments are public. Links aren't allowed, and we delete abusive posts. You can delete your own comment at any time in this browser. Learn more in the Privacy Policy.