These data are shared by ChatGPT, Gemini, Claude & Co. with third parties

A new study by IMDEA Networks Institute has investigated when which chatbot shares which data about its users with third parties. The result: all of them share data – sometimes even access data to entire chat histories of customers is included.

As part of the study, which is currently still an unreviewed preprint, the scientists investigated nine AI providers: ChatGPT, Claude, Gemini, Grok, DeepSeek, Perplexity, Le Chat, Meta AI, and MS Copilot. Some third parties to whom they share data are advertising trackers; others are external service providers – many are both simultaneously.

Even chat content is sold

In addition to a whole series of metadata, according to the study, five of the nine AI providers examined share permalinks to chat histories with data trackers when accessed via the browser. This means that, if there are no access restrictions, as with many AI providers, these could then access the conversations that users have with their AI. And thus all content that a customer discusses with an AI. In 40 percent of these cases, data sharing can be prevented by rejecting non-essential cookies, according to the results of the investigation. In 60 percent, this is not possible.

To find out whether these chats are actually accessed by third-party providers via the permalinks, the researchers placed Canary URLs, a kind of digital “trap,” in their investigated AI chats. In several AI providers, these triggered the moment the chat was shared with third parties, according to the study. With Grok, the Canary alert was also triggered several times afterward, according to the researchers, which presumably means that external parties accessed the AI chat several times.

In addition to permalinks, 33.3 percent of the AI providers examined probably share summaries with their advertising partners about what the individual AI chats were about. These automatically generated descriptions sometimes contained deep insights into users' lives. The study cites examples such as “What are the early-stage symptoms of Parkinson’s disease?” or “My salary is $85,000. How much can my mortgage be in New York City?” User location is also shared with third parties in individual cases.

For many AI providers, all such information is shared with email addresses and permanent user identifiers, according to the study – even in incognito mode in browsers. This means the data is usually no longer anonymous. Because such IDs are assigned to internet users by data brokers and subsequently identify them uniquely across platforms. For example, LiveRamp works with its so-called RampID for its collaboration with OpenAI – a pseudonymous identifier that OpenAI uses to identify ChatGPT users for targeted advertising. Unless appropriate protective measures are taken.

Hardly any opt-out options

Even if “non-essential cookies” are rejected, 80.8 percent of third-party trackers remained active, according to the study. If users accept all cookies, some AI providers probably deliberately activate additional trackers. A subscription also makes little difference: free and premium customers are tracked by largely identical third-party providers, according to the scientists.

Overall, the researchers found 44 tracking or advertising third parties that receive data from AI providers during their investigation. However, they emphasize that these can vary from test run to test run. Trackers from Google are the most widespread among all AI providers, followed by well-known data trackers such as Sentry, Meta, Datadog, and Intercom. Companies that are not primarily associated with their data collection practices are also included –, such as TikTok, Apple, and X.

That not only AI services but also other connected products share data on a large scale is shown by a current study on how ordinary passenger cars sell their users' data to third parties.

(rie)