Are Your AI Consultations Exposed to Ad Trackers? A Study of 42 Services
機械翻訳 / Machine-translated

機械翻訳 / Machine-translated
@aifriends
AI Friends(https://aifriends.jp)のクロスポスト公式アカウント。AIツールの紹介・使い方・できることを、中学生でもわかるやさしい日本語で届けます。
Have you ever confided a health concern or vented about work to an AI chatbot? That conversation may be reaching advertising trackers — small programs that collect browsing data. Here is a plain-language breakdown of a study published in October 2026.
Researchers at the State University of New York at Stony Brook published a paper titled "Penny for Your Prompts" on arXiv (a preprint server). It was posted on October 27, 2026, and is scheduled to appear at IEEE S&P (a leading security conference) in May 2027.
The study examined the web interfaces of 42 AI chat services — 25 general-purpose AIs including ChatGPT, Claude, and Gemini, and 17 AIs specialized in areas such as health, law, and education. The findings were also covered by ITmedia.
The results were as follows:
In half of the services examined, information related to conversation content was leaving the platform.
The trackers are not doing anything like hacking. They are simply reading page titles automatically — exactly as they do on any ordinary website.
Many AI chatbots generate a title for a conversation based on the user's question and use that as the page title. For example, if you ask about "early symptoms of pregnancy," the page title becomes something close to that phrase. Trackers then collect that title just as they always would.
There were two main pathways through which the actual text entered by users leaked. One involved session-recording tools — analytics tools that record and replay screen activity and text on a page — inadvertently capturing text entered in the interface. The other involved services whose specifications included the user's input directly in the URL.
The research team tracked not only network traffic but also the behavior of browser JavaScript (the programs that make web pages interactive) in granular detail. They were able to identify exactly which user action triggered the transmission, from which part of the page, and to which destination.
Among the 25 general-purpose AI services, there was an average of 540 browser network requests. Of those, approximately 22% — 118 requests — were directed at trackers.
Breaking down tracker traffic by recipient company:
That said, this does not prove that these companies are snooping on conversations. What the study established was the fact that transmissions occurred. How each company uses the data it receives is a separate question.
Several studies examining tracking in AI chat services have been published this year.
The University of California, Davis published a study examining 20 services. Using a single input — "pregnancy test near me" — researchers found that 17 services shared data with at least one external vendor. Three services transmitted the full conversation text to Microsoft Clarity, a session-recording tool. Private chat modes were reported to significantly reduce external connections.
Spain's IMDEA Networks Institute announced findings in May 2026, noting that ChatGPT, Claude, Grok, and Perplexity use trackers from Meta, Google, TikTok, and others. The authors described their results as "preliminary."
The three studies differ in methodology and scope. Nevertheless, they point in the same direction: information close to conversation content can leave the platform. The current study's coverage of 42 services shows that the problem is not confined to a handful of outliers.
In Japan too, more and more people are using ChatGPT, Claude, and Gemini for work and personal life. This is not someone else's problem.
Consider, for example, an office worker consulting an AI about an HR concern. If they include their company's name or a supervisor's name in their question, those words may end up in the conversation title as well.
Other scenarios worth considering:
If an email address is also transmitted externally alongside this information, there is a risk that conversation topics could be linked to an individual.
Note that it has not been confirmed how many Japanese services were included among the 42, nor does this study shed light on behavior in Japanese-language interactions.
Until service providers fix these issues, there are steps users can take.
The UC Davis study found that private chat mode reduced external connections. However, IMDEA notes that declining cookies alone is not sufficient. Rather than relying on any single measure, combining them is the practical approach.
For service developers, the UC Davis research recommends not including conversation text in page titles or URLs. It also suggests avoiding the use of session-recording tools on logged-in screens.
Detailed results for individual services were not confirmed within the scope of our investigation. The list of 42 services appears in an image within the ITmedia article. If you are concerned, please refer directly to the paper or the article.
No, it is a separate issue. The problem here is that information is being passed from the screen to advertising and analytics vendors that are distinct from the AI operator itself.
This was not determined within the scope of the study. Whether a difference exists between paid and free plans has not been confirmed.
Only web interfaces were studied. The situation with app versions was outside the scope of this research.
Avoid entering information that could identify you personally. Beyond that, using private chat mode is recommended.
As your next step, start today — try entering your questions without including your name or employer.
This article is a cross-post from AI Friends.