Chatbots let you delete a conversation, but does that also erase what the AI figured out about you?

This explores whether the delete buttons in real chatbots also wipe what the AI has inferred about you, not just what you typed, and whether that matches what users believe happens. The short answer is that the corpus doesn't settle the technical question. No study here audits a production chatbot to check whether its inferences actually disappear after deletion. What the corpus does show is a sharp gap between how much users want deletion and how much they trust it.

The clearest evidence comes from a 354-person study in which deletion controls beat every other privacy option on both preference and perceived effectiveness. Even so, participants doubted that deleting a chat really erases the data or the conclusions drawn from it Do deletion controls actually remove AI chatbot data as promised?. The finding most readers won't expect is the cause of that doubt. Users weren't confused about how deletion works. They doubted the companies offering it. The researchers call this an "assurance gap," and better explanations won't close it, because the issue is credibility, not understanding.

A second study shows why inferences matter so much. Most of its 40 participants never noticed that their LLM's memory feature held private information it had inferred rather than been told. Once shown, every one of them wanted editing and deletion controls Do users notice when LLMs infer and retain private information?. Taken together, the two studies say something uncomfortable. People mostly picture deletion as erasing what they said, while the real privacy exposure often sits in what the system concluded. Users can't ask for inferences to be deleted if they don't know those inferences exist.

One response is to stop data from being collected in the first place rather than relying on deletion afterwards. In a 188-person study, AI-suggested edits to chatbot transcripts removed much more personal information than people managed by editing on their own, with no loss in data quality Can AI suggestions help people remove their own personal information?. That matters because chatbots draw out disclosure: people tell them more intimate things precisely because no human is there to judge Do chatbots help people disclose more intimate secrets?. The same features that make chatbots good confidants also increase the amount of inferred material a delete button would need to reach.

The open question, then, is not whether people want deletion. They clearly do. It's whether anyone can verify what deletion actually removes. Until a study traces inferences through memory, logs and training pipelines, the corpus can only show that users' scepticism is reasonable, not whether it's correct.