Deleted AI Chats May Leave Recoverable Traces

Summary

Georgia Tech researchers developed ORISA, a tool that can recover fragments of deleted conversations from local AI assistants running on Android phones and rebuild the information into a working AI session. The study shows that deleting a conversation does not necessarily remove every trace when the assistant keeps information in the phone’s working memory. In one experiment, the researchers created a fictional assistant called Project Note Keeper in the ChatterUI Android app, supplied it with 100 fictional project facts, and then deleted the conversation. ORISA reconstructed a session that returned 26 facts when asked to list the notes, while the original session could actively access only 14 because other information had left its active memory. The result indicates that memory forensics may recover information an assistant was no longer actively using. The team tested ORISA across 10 Android apps and three AI models in 60 configurations, recovering information that remained in memory. Recovery becomes more difficult as later phone activity overwrites that memory. The work could aid digital investigators, but it also raises privacy concerns for users of local AI assistants. The research was presented at the 2026 IEEE Symposium on Security and Privacy.