Most people type into a chatbot the way they'd talk to a pharmacist across the counter. A question goes in, an answer comes back, and the exchange feels like it happened between two parties. That feeling is a big part of why people ask chatbots about symptoms, money and relationships they'd never put into a search bar. Few people picture the chat window as an ordinary web page, with its own crowd of advertising and analytics code running in the background.
On some platforms, that crowd is listening more closely than most users would guess. For anyone who has asked a chatbot something sensitive, a preprint by Muhammad Jazlan, Ethan Wang, Yash Vekaria and Zubair Shafiq at the University of California, Davis, makes uncomfortable reading. Of the 20 popular chatbots they examined, 17 passed information to at least one outside company during a single chat. On three of them, the outside company received the actual words of the conversation.
The paper in 200 words
Jazlan, Wang, Vekaria and Shafiq set out to measure web tracking on the websites of 20 popular AI chatbots, chosen using Similarweb traffic rankings. They created a fresh test account on each service, not tied to any real person, and used Chrome, which offered no built-in tracking protection at the time. On each chatbot they sent one fixed prompt, "pregnancy test near me", chosen because it pairs a health question with an implicit location request. They recorded all network traffic from page load until the response finished. They then decoded and decompressed that traffic and searched it for two kinds of information. Content meant the prompt, words from it, prompt-derived chat titles, response text, and each conversation's unique link or identifier. Identity meant the account's name, email, internal user ID, site cookies, and IP address or browser details placed in a message. They searched for exact text and for common encoded and hashed versions (scrambled with a standard formula). They then sorted each recipient into first party, platform party (same owner, different domain) or third party. The seven chatbots offering a private or temporary mode were tested a second time in that mode. Finally, they compared what they observed with each service's privacy policy.
This paper was posted to arXiv on 30 April 2026 and has not yet been peer reviewed. Its results come from one session per chatbot in one browser, so they describe a snapshot of each site at the time of testing rather than a permanent property of the service.
The conversation leaves by several routes
Four of the chatbots embed Microsoft Clarity, a session replay tool (software that records what appears on a page so a company can play back a visit later, a bit like a security camera pointed at the screen). Because the chat page displays the conversation, the recording captures the conversation too. On Genspark, SeaArt and ChatOn, Clarity received the exact prompt in readable text, plus pieces of the reply. On Genspark, that reply read "Here are a few pregnancy test options near you." On ChatOn, it named CVS, Walgreens and Rite Aid as places that carry home tests.
Session replay is the most vivid leak, though probably not the most common one. Genspark tucked the full prompt into a Google Maps link to draw a map. That hands the words to Google and, depending on the browser, may tie them to the user's Google account. Five chatbots, including Claude and Gemini, put the prompt or a keyword like "pregnancy" into the page title (the text on the browser tab). Analytics and advertising code reads page titles by default, so on Kimi the title reached Google's advertising system.
The broadest exposure looks the dullest. Fifteen of the 20 chatbots sent the conversation's link or unique ID to at least one third party, reaching 29 destinations in total. SeaArt sent it to 12, including Facebook, TikTok and Yandex. An ID reveals no text by itself... but it lets whoever receives it attach other information to that particular chat. If the link is shareable, it points straight at the conversation.
behavior change 101
Your name often arrives before your question
On Claude and Mistral, the customer support chat bubble (a widget made by a company called Intercom) sends a message to Intercom as soon as the page loads. That message carries the user's email, name, internal account ID and a user hash, a scrambled code tied to the account. Nobody has to click the support bubble for this to happen. The user hasn't typed a question yet.
Analytics tools pick up identity too. On Character.AI, an error monitoring tool received the user's email and was set up to record the user's IP address. An experimentation tool on the same site received the email, account ID, IP address and browser details. The authors point out that typical error monitoring needs neither the email nor the IP address.
Advertisers get a version of identity built to travel between sites. Perplexity sent a hashed email to an advertising company called Singular, and SeaArt sent one to TikTok's tracking code. Hashing sounds like anonymising, but the same email always produces the same hash. Any company receiving it can probably match a person's activity across every site that reports the same code. Meta's advertising cookie showed up on six chatbots, including Character.AI and SeaArt.
How others are reacting
When the paper was shared on LinkedIn, much of the thread was shock and some of it was shrugging. The more useful replies pushed on what the study actually measured, and three of those arguments change how the findings read.
A data analytics specialist working on AI argued that the paper's categories need pulling apart. In that view, a support widget sending a name and email on page load is ordinary behaviour for subscription software, with the widget's vendor doing what it always does. Conversation text landing in a session replay tool is a different class of problem, and lumping the two together blurs where the real risk sits. A senior lecturer in organisational psychology aimed a similar objection at the title. Only three fairly small chatbots shared actual conversation content, the lecturer noted, and showing ads seems different from tracking conversations.
One of the authors replied that "tracking" carries its standard meaning here: third-party code on a page counts as tracking, whether or not the prompt leaves. That definition is defensible, but the objection still changes how the headline number should be read. The 17 of 20 figure counts any third-party contact, while plaintext conversation sharing happened on three services most readers have probably never used. The gap is narrower than the critics suggest, though, since 15 chatbots sent conversation IDs and Claude's page title, carrying a prompt keyword, reached Intercom. The headline number overstates the content leak, and the conversation IDs show the critics' distinction has limits.
A co-founder of a business automation firm pointed out that the method captures what third parties receive in one session. It can't capture what they work out by matching that data against profiles they already hold. If an ad network already recognises a browser from its technical quirks, a line about pregnancy test options becomes one more fact on an existing file. A cardiologist who also works in product management made the stakes concrete for medicine. Knowing what a doctor asked a chatbot could be enough to aim drug advertising at them, which might shape prescribing without the chatbot ever biasing an answer.
The paper can't test this from network traffic alone, and it doesn't claim to. The authors do flag two places where joining is possible: the Maps link that may tie a prompt to a Google account, and hashed emails that let advertisers recognise people across sites. The clinician's scenario remains a hypothesis, since the study tested consumer chatbots with an invented account rather than clinical tools. Read this way, the paper's counts describe a lower bound on what outside companies can learn.
A monitoring and evaluation specialist working on equitable AI noticed that Gemini and Meta AI showed the least third-party sharing. The commenter wondered whether Google and Meta simply have little reason to share data outward when they already hold so much of it. Another reader asked why Google and ChatGPT seemed absent from the chart, and an author pointed to a project website showing the full set of flows.
The paper supports the observation without settling the motive. Gemini, Meta AI and Duck.ai contacted no third parties, but Gemini sent its page title, which can carry a prompt keyword, to Google Analytics. The authors count that as platform party, since Google owns both ends, and ChatGPT likewise routed analytics data through its own domain. Duck.ai looks like the genuine outlier, stripping users' IP addresses before forwarding requests to model providers. A zero in the third-party column means the data stayed inside the company, which is a different promise from no tracking at all.
Taken together, the replies narrow the paper's headline and widen its implications. Fewer chatbots leak conversation text than the top-line number suggests, but each flow the study counts is probably worth more to its recipient than it looks in isolation.
Private mode helps, though nobody says so
Among the seven chatbots with a private or temporary mode, pairings between a chatbot and a third-party domain fell to 13 in private chats. Outside owners fell to three: Datadog, Mapbox and Google. The researchers saw no content or identity reaching any third party in any private session. That's a strong result, though it rests on a single private session per chatbot.
Oddly, the providers don't advertise this. The notices shown when a user opens a private chat talk about retention, model training and personalisation rather than tracking. Meanwhile, eight of the 20 privacy policies describe kinds of recipients without naming the third parties seen in the traffic. ChatOn, SeaArt and Genspark name some companies but leave out Microsoft Clarity, the one receiving conversation text. Claude and OpenRouter, by contrast, list a broad set of third parties and the types of data each receives.
The authors argue that most of these exposures come from design and configuration choices, so providers could fix them without ripping out analytics or ads. They suggest keeping prompt text out of page titles and stripping chat IDs from links reported to outside tools. They also suggest keeping session replay off signed-in conversation pages, or masking what it records. Until providers make those changes, the private chat button looks like the most useful privacy control on these sites, which is probably not quite what it was built for.



















