Chatbots IA: On 42 audited platforms, 21 transmit the title of conversations to third party trackers, and nine messages themselves

Researchers have instrumentalized the browser to track each data sent by chatbot web interfaces. Discussion titles, complete prompts and e-mail addresses go to advertising and analytical scripts.

Chatbot interfaces receive confidences every day about their users' health, rights or privacy. But like any site, they often include third-party scripts: audience analysis tools, advertising networks, session recorders. A team led by Paschalis Bekos, with Nick Nikiforakis and Michalis Polychronakis, has measured for the first time on a large scale what these scripts retrieve from the content of conversations.

Instrumentation at the JavaScript engine level

Rather than simply intercepting network traffic, the authors modified the browser to link each data stream to the precise user action that triggered it and how the content was read on the page. They audited 42 platforms, both general and specialized in health, law and education.

As a result, half of them (21) leak the titles of conversations to trackers, and nine transmit the text of prompts as is. Many also send the user's e-mail address, which makes it possible to link the content of a discussion to a real identity. For the authors, the old problems of privacy on the web take on a new seriousness with these tools where people open up more than elsewhere.