Hidden traces in AI incognito mode - is our digital confessor really private?

More and more people are using AI for therapeutic conversations, as a confessor, or to share fantasies and personal secrets. GROK and CHATGPT offer incognito chat modes that allow private conversations without using chat elements for AI training. They also promise not to save conversation history or store our chats in account history. Theoretically, this ensures 100% anonymity and trace erasure after closing the chat window. So we close the window and take a big sigh of relief… it’s deleted, no one will know.

Or will they?

There’s a small problem that few people pay attention to(*). AI automatically generates titles for every conversation, even in incognito mode. These titles end up in the browser and remain in the visited pages history.

In my opinion, this is a problem because if we’re having an incognito conversation, why should AI generate a contextual name for it? If we’re using a normal browser, we might have titles like “how to cheat on my husband” or “I have suicidal thoughts” in our browsing history - because if I ask AI about this in incognito mode, that’s exactly how it will name the conversation and pass that parameter to the browser.

I assume that most ordinary people don’t additionally use the browser’s incognito mode. If they have paid AI accounts, they use the normal browser mode for convenience.

To test this, I started several conversations with Grok in incognito mode. Here’s what I got:

“Finding secret data on a flash drive - Grok”

“Secret attack plans: what to do? - Grok”

Here’s the crux of the matter. People have paid AI accounts, so they use incognito mode on the AI side, but out of convenience and ignorance, they don’t enable incognito mode in the browser. And that’s when these AI-generated titles end up in the browser history.

This is exactly the opposite of what we expect from private mode. Instead of minimizing data processing, AI does exactly the opposite - it deepens analysis to create a “useful” title that gets pushed to the browser. And the browser saves it in history.

Now imagine that instead of tests, you’re having real, intimate conversations. If someone has confidential information and analyzes with AI in incognito-chat mode what to potentially do about this problem, then closes the window. We’ll still see a trace in the browser history that, through search contextualization and a very informative title, can give us a lot of knowledge about what that person was doing or what they deny doing.

Pretty bad, right? Who has the energy for double incognito and paranoia requiring special procedures when content consumption and analysis takes 1-3 seconds?

The problem is that most people have no idea about this loophole. They buy a subscription, turn on incognito mode, and think they’re safe. Meanwhile, they leave traces in the browser that can be more compromising than the conversation content itself. Especially if our AI-confessor is the holder of sensitive knowledge.

(*) excluding privacy paranoids who constitute 0,1-0,5% of internet users

• Ai
• Chatgpt
• Grok
• AI Series