Deleted AI Conversations Can Be Recovered From Your Phone Memory According to New Research

Deleting a conversation with an artificial intelligence assistant on your smartphone might not be as permanent as it seems. New research from Georgia Tech has uncovered a significant privacy vulnerability, revealing that sensitive data from supposedly wiped AI interactions can linger within a device’s local memory, leaving it open to recovery.

Recovering Forgotten Digital Conversations

The team developed a specialized tool dubbed ORISA, designed to scan the internal memory of Android devices to pinpoint remnants of deleted AI sessions. Beyond merely identifying these fragments, the tool can effectively reconstruct that information, potentially allowing a deleted session to be fully restored and queried.

This discovery centers on the architecture of local AI models. Unlike cloud-based services that process requests on remote servers, local assistants function directly on the handset. While this approach is often marketed as a privacy-friendly alternative because data never leaves the device, the researchers found that these apps often fail to properly sanitize their working memory. Even after a user manually deletes a chat, the information remains stored in the phone’s temporary memory, waiting to be overwritten by new processes.

Stress-Testing Local AI Security

To evaluate the efficacy of their recovery method, the researchers conducted extensive testing across 10 distinct Android applications and three different AI models. Across 60 unique configurations, ORISA consistently succeeded in extracting conversation fragments that users believed had been discarded.

In a controlled experiment, the team utilized a custom AI assistant named Project Note Keeper, feeding it 100 specific facts before deleting the interaction history. Remarkably, ORISA not only recovered the deleted data but managed to reconstruct a functional session containing 26 of those original facts. Interestingly, this surpassed the recall of the original, non-deleted session, which could only access 14 facts due to limitations in its active memory management.

“It can help establish what information someone shared with an AI assistant and what assistance they received,” explained Ken Xu, a PhD student at Georgia Tech and lead author of the study.

Implications for Privacy and Digital Forensics

The findings, presented at the 2026 IEEE Symposium on Security and Privacy, highlight a nuanced challenge for both user privacy and the field of digital forensics. For investigators, this capability offers a powerful new avenue to reconstruct digital activity that was previously thought to be scrubbed. For the average user, however, it serves as a stark reminder that local storage is not synonymous with absolute data destruction.

The researchers noted that the likelihood of recovering this data decreases over time as the device performs other tasks and gradually overwrites the memory space where the old conversation resided. Nonetheless, the existence of such a tool underscores the need for more robust data-handling protocols in the development of mobile-based artificial intelligence applications.