This tool shows how deleted AI chats can be recovered from your phone

Georgia Tech ORISA Recovers 26 Deleted AI Chat Facts From Android Phones

Meta description: Georgia Tech researchers built ORISA to recover deleted AI chats from Android memory, showing how local sessions may leave evidence longer than expected.

Tags: ORISA, Georgia Tech, local AI, Android forensics, deleted chats

Featured image: abstract smartphone memory circuit pattern with faded chat bubbles, no people, no logos, no text.

Researchers at Georgia Tech have created an experimental tool called ORISA that can recover deleted AI chats from Android phones. The result matters for two reasons: it may help cybersecurity investigations recover erased evidence, and it shows that perceived local AI security has limits. Privacy-conscious users often keep AI chats on-device, but the study suggests deleted chats may leave traces in working memory longer than intended.

Background: why local AI was treated as private

Many privacy-conscious AI users prefer to run assistants locally on their devices because their data is not sent out to the cloud, where it could be exposed to prying eyes or data breaches. Instead, the chat session uses the phone's memory. The researchers said a downside of local AI is its lack of centralized logging and enforcement. That creates a gap for investigations: bad actors can fine-tune models for crimes, for example, and then delete the evidence afterward, making it hard for forensic memory techniques to recover useful traces.

Development: how ORISA rebuilt a deleted session

To test the idea, the team used ChatterUI, an Android app, to create a fake note-keeping AI assistant. They gave the assistant 100 fictional pieces of information about a project, then deleted that information. ORISA searched the device memory for remnants that could be used as context to restart the session. The researchers wrote that ORISA leverages the batch tokens and key-value caches preserved in process memory to reconstruct token sequences and restore session-level attention states.

In the experiment, ORISA found enough usable information to create a new working session with 26 fictional facts from the original set of 100. That was more than the 14 facts the original assistant session had access to, because the rest had fallen outside its active memory, the researchers said. Ken Xu, a doctoral student and lead author of the study, said the method can help establish what information someone shared with an AI assistant and what assistance they received.

Current stage: what the recovery result means now

The researchers described the approach as potentially promising for getting other types of erased information off phones during investigations. But it also means a user's on-device AI chats, which are supposed to be private, may have a longer shelf life than intended. The tool was developed to fill a cybersecurity gap, so the most immediate relevance is to investigators trying to stop cybercrime rather than to personal snooping. Still, the finding changes how on-device AI privacy should be understood. If someone could recreate a similar technology, or use a tool like ORISA, deleted chat information could theoretically be gathered and used as context to reconstruct queries. The researchers noted that recovery becomes harder as time goes on because the phone's memory gets overwritten by newer activity. The study notes that sensitive information shared with chatbots can be recoverable from device memory, even in local sessions.

For now, ORISA remains an experimental recovery method, not a widely available consumer feature. What to watch next is how time affects recoverability: as newer activity overwrites phone memory, the window for recovery narrows.