What the study measured
A Stony Brook University research team audited 42 AI chat platforms and found that 21 of them transmitted conversation titles to third-party trackers. Nine of the 42 also transmitted the user’s verbatim prompt. The paper, “Penny for Your Prompts: A Measurement Study of Contextual Leakage in AI Chat Platforms,” was submitted on 27 October 2026 (arXiv:2610.37467) and is accepted for the IEEE Symposium on Security and Privacy in May 2027. It has not yet been published.
The audited sample covers 25 general-purpose chatbots and 17 task-specific platforms for education, health and legal questions. The measurements ran from 6 to 20 May 2026 from a single vantage point in the eastern United States and used free web tiers. That scope shapes every number below.
“Leak” has a specific meaning here. The authors count a leak when conversation context (the full prompt or the auto-generated conversation title) or personal information such as the user’s email reaches a domain outside the platform’s own organization. They exclude infrastructure such as content delivery networks, consent-management tools and payment processors. A leak in this paper is transmission to a third-party tracker: an analytics, advertising or behavioural-tracking script. It is not a breach, a hack, a public dump or a sale.
The team measured in two phases. First, session capture: the researchers interacted manually with each platform in a browser instrumented with Puppeteer and followed the same sequence of interactions each time, creating an account, sending a prompt, starting a second chat, sending another prompt and moving between chats. Every outgoing request was logged and searched for the seeded prompt or the recorded title. Second, interaction decomposition: VisibleV8, an instrumented Chromium build that logs JavaScript API calls at the V8 engine level, was run across five isolated actions, including prompt submission, navigation to an existing chat, deletion, scrolling and an unrelated click. That phase links each outbound flow to the user action that caused it, which network-level analysis alone cannot show.
Why a conversation title is more revealing than it sounds
Most chat products name a conversation automatically, usually from the user’s first prompt. A title such as “Immigration status options for my brother” is a short summary of what the user asked, and once it exists the platform writes, stores and sometimes sends it like ordinary page metadata.
Of the 25 general-purpose platforms, 13 transmitted a conversation title in the audited period: Character.AI, ChatGPT, ChatOn, DeepSeek, Genspark, Grok, Kimi, Manus, PolyBuzz, QuillBot, Writesonic, You.com and ZenoChat. Four of them also transmitted verbatim prompts: You.com, Writesonic, Kimi and ZenoChat. The other 12 general-purpose platforms were not observed transmitting a title to a tracker in this sample and period. That absence is not a clean bill of health. It means the researchers did not observe it, not that it does not happen.
You.com and Genspark were the heaviest title transmitters, at 62 and 40 title-bearing requests per prompt interaction to Google. You.com also produced roughly 92 prompt-bearing requests to Google per prompt submission, and Writesonic distributed the raw prompt to three independent tracker endpoints.
The recipients are concentrated. Across the 25 general-purpose platforms, an average prompt interaction generated about 540 outbound requests, 118 of them (22%) to known tracker domains, and the researchers identified 153 distinct tracker domains. Google received 41.8% of tracker traffic, Microsoft 12.2%, ByteDance 8.6% and Meta 8.3%. Google alone accumulated 230 context-bearing requests across eight platforms.
Email makes the picture sharper. Eleven of the 13 general-purpose title transmitters also sent the user’s email address to a tracker, and for eight of them a single company received both the email and the conversation context. In six of those eight cases the study directly observed a linking identifier, either inside the same request or across separate requests that shared a token. Where that happens, one company can connect an email address to the topics a person asks about rather than holding two unconnected records. Kimi and PolyBuzz are the two exceptions: the same company received both signals, but the researchers found no shared token and describe linkage there as likely to rely on IP address or browser fingerprint instead. The study also found that accounts registered with Gmail produced a higher rate of tracker-bound requests than custom-address accounts, 24% against 19%, but with one session per account type the authors say confirming that gap would need a dedicated future measurement.
What this does not mean
- It is not evidence of a breach, a hack, a public dump or a sale. The paper does not claim that any recipient read, used or exploited the data, and it reaches no conclusion about intent or harm.
- Twenty-one of 42 audited platforms is not 21 of the market. The sample covers free web tiers during a two-week window from one measurement location, with a single session per configuration, and the team performed no re-check.
- A platform missing from the title list is not safe. It simply was not observed transmitting a title in this sample.
- The study says nothing about mobile apps, paid tiers, other regions or consent states other than the defaults it used.
- The paper is accepted for a conference but not yet published, so the version measured may still change before the final proceedings.
What you can check or do
Because the title usually comes from the first prompt, whatever you type into that opening message can become short, readable metadata that leaves the platform through an analytics or advertising script. On the audited platforms this is a documented transmission that the study measured directly, not something inferred from the mere presence of a tracker.
The paper’s recommendations target the platforms, not the user. It proposes auditing trackers beyond page load and prompt submission, because interaction-driven tags fire after those points; writing less specific or less revealing conversation titles; and using analytics tools that do not collect content-bearing page attributes. These are design proposals, and the researchers did not test any of them.
The study also did not test user-side mitigations. It does not measure whether tracker blockers, private browsing, consent choices or account settings change what is transmitted, so no such advice should be read into the results. Sessions in the study accepted cookie and consent prompts in their default state, so consent behaviour is untested too.
The practical weight lands on what you paste. For work, the relevant question is whether the first message you send to a chatbot can contain text restricted by an employer policy, a client agreement or a confidentiality rule. The study does not show anyone reading that text. It does show that on many audited platforms the title built from it is transmitted to third-party trackers, which is enough reason to treat an auto-generated conversation title as something that can leave the app.
