Do AI chatbots send your chats to trackers? Two 2026 studies, explained
Some AI chat websites sent conversation information to trackers under the studies' test conditions. Your chats are never on our servers, because we never receive them. Outlier's local chat runs on your Mac, with no chat web page or trackers in the chat; the dated lsof capture below shows the engine check.
Two studies asked related questions
A tracker is an outside service used by a website to record activity. A conversation title can reveal a topic even when the full message is absent. These studies looked at information reaching trackers through AI chat websites. They used different methods and reported different results.
| Study and date | Scope and method | Finding |
|---|---|---|
| Bekos, Tsouvalas, Nikiforakis and Polychronakis, Stony Brook University. Submitted 27 October 2026. | Audited the web interfaces of 42 general-purpose and domain-specific chat platforms, including health, legal and education services. | 21 leaked conversation titles to third-party trackers. 9 leaked verbatim user prompts. A substantial number sent user emails to trackers. |
| IMDEA Networks Institute. Published 6 May 2026. | A separate study of trackers used by AI chat services. The report names ChatGPT, Claude, Grok and Perplexity. We do not have a verified total service count. | Reported trackers from Meta, Google, TikTok and others, with potential transmission of conversation information such as titles, permalinks and metadata. |
Sources: Penny for Your Prompts: A Measurement Study of Contextual Leakage in AI Chat Platforms, submitted 27 October 2026, and IMDEA Networks Institute, 6 May 2026. Both sources read 2026-11-20 by our reviewer.
How to read the difference
The Stony Brook counts describe observed leakage under the study's test conditions. The IMDEA report describes tracker use and potential transmission of conversation information. Finding a tracker and observing a verbatim prompt reach it are different findings. The numbers from one study should not be attached to the other.
Neither result makes a lasting promise about any named service. A result under the study's test conditions covers that test. Different methods and results do not support calling a service clean or declaring every cloud chat unsafe. For a particular app, check the date, the feature in use and what information was actually observed leaving.
Settings you can review now
- Use your browser's tracker blocking options. They may reduce requests to outside trackers. They do not change the fact that a hosted model receives the prompt you send it.
- Consider signed-out use where the service offers it. This can avoid attaching a signed-in account to the session, but it does not establish that tracking stops.
- Read the app's own privacy settings. Check controls for history, training and connected features separately. A training setting alone does not answer which web requests a page makes.
These are choices to review, not guarantees that a browser session has no sharing. For work that should stay off a chat website, a local model offers another way to do the task.
A local chat and a marketing website are separate
A local app with no chat web page and no trackers in its chat gives a website tracker no chat page to read. Outlier has no analytics SDK in the desktop app. Our marketing site uses Plausible: it counts page views, not chats. The site's scripts and the app's network uses are listed in our privacy policy.
Existing home page measurement: Captured 2026-11-04 on Outlier 1.11.992 while Core was answering, web search off. A past capture, not a live reading.
It was sampled 22 times across a 24s answer. Sockets leaving this Mac: 0. Every socket the engine held used 127.0.0.1. See the original lsof capture.
You can inspect the engine while it answers with lsof -a -p $(pgrep -x outlier-cli) -i -n -P. The -a makes the process and network filters apply together. This past capture cannot describe a session on your Mac today, and connections can occur between samples. See how to run and read the check.
Local mode means using a local model in Chat or Code mode with web search off; Pro can pick Research in Auto mode. Web search and cloud models send information to outside services when you turn them on, and Pro's Research also sends search queries even with web search off. Download Outlier if that local workflow fits your work.