Deleted AI Chats May Not Erase All Data, Georgia Tech Study Finds

Overview

A new study from the Georgia Institute of Technology reveals that deleting a conversation with an AI assistant on an Android phone may not erase all traces of that interaction. Researchers have developed a forensic tool capable of recovering deleted conversations and even rebuilding them into active AI sessions.

The findings highlight a significant privacy gap for users who rely on local AI assistants—those that run directly on their devices rather than sending data to cloud servers. Even after a user deletes a chat, fragments of the conversation can remain in the phone’s working memory.

This research introduces ORISA, a novel tool designed to extract and reconstruct these lingering digital footprints. While the technology offers new possibilities for digital investigators, it raises serious questions about the permanence of deleted data in an era of increasingly personalized AI.

The ORISA Tool and Methodology

The core of the research is a tool called ORISA, which stands out for its ability to recover not just raw data, but functional conversational context. Unlike traditional forensics that might recover deleted files, ORISA searches the volatile memory of Android devices for pieces of deleted AI conversations.

The system works by identifying artifacts left behind in a phone’s working memory while a local AI model is running. These artifacts persist even after the user actively deletes the conversation thread through the app interface.

Once recovered, the tool can reassemble these fragments into a working AI session. This means the deleted context is not merely archived as static text; it can be rehosted and interacted with as if the conversation had never been erased.

The study involved extensive testing across multiple configurations to ensure the robustness of the findings. Researchers tested ORISA with 10 different Android apps and three distinct AI models, resulting in 60 unique test configurations.

Across all scenarios, the tool successfully recovered conversation information that remained in the phone’s memory. The experiments demonstrated that recovery is feasible even when users believe they have completely removed their data history.

Experimental Results and Data Recovery

To demonstrate the practical implications of their work, the research team conducted a controlled experiment using a fictional AI assistant named “Project Note Keeper.” This assistant was built using the ChatterUI Android app and fed 100 fictional facts about a made-up project.

After the AI processed these facts, the researchers deleted the entire conversation. They then used ORISA to scan the device’s memory for any remaining traces of the interaction.

The results were striking. ORISA successfully recovered information from the deleted session and reconstructed it into a new, active working session. When asked to list the project notes it remembered, the restored assistant returned 26 of the original facts.

Notably, the original active session could only access 14 of those facts because many had fallen outside the assistant's immediate active memory window. This indicates that ORISA can recover information the AI had already stopped actively using or highlighting.

This capability suggests that deleted conversations may contain far more recoverable data than users might expect. The tool does not rely on the AI still "remembering" the information in its current operational context but rather on the residual data left in the device's memory architecture.

“It can help establish what information someone shared with an AI assistant and what assistance they received,” said Ken Xu, lead author and Ph.D. student at Georgia Tech.

Privacy Implications for Local AI Users

The ability to recover deleted AI chats has profound implications for privacy, particularly for users of local AI assistants. Many users turn to locally running models specifically to avoid sending sensitive data to cloud-based services, assuming this provides a higher degree of security and anonymity.

However, the study shows that local processing does not guarantee local deletion. Information from conversations can remain in the phone’s working memory even after the user deletes the chat history through the app’s interface.

As the phone’s memory is overwritten by other activity—such as installing new apps, taking photos, or running other processes—the recoverability of this data decreases. Nevertheless, the window of opportunity for forensic recovery remains significant for a considerable period after deletion.

For digital investigators, this presents a new avenue for evidence gathering. ORISA could potentially help establish the timeline and content of communications between a user and an AI assistant, which may be relevant in legal or corporate investigations.

Conversely, for the general public, the findings serve as a stark reminder that "delete" buttons on AI interfaces may not function as they do for other types of files. The assumption of digital impermanence is challenged by the persistent nature of memory artifacts in modern mobile devices.

Research Context and Future Directions

The research was presented at the 2026 IEEE Symposium on Security and Privacy, one of the premier conferences in the field of cybersecurity. The paper, titled Recovering and Rehosting Mobile Local LLM Conversations and Contexts via Memory Forensics, underscores the growing intersection of AI usability and digital forensics.

The study was conducted at Georgia Tech’s Cyber Forensics Innovation Laboratory, a collaborative effort between the School of Cybersecurity and Privacy and the School of Electrical and Computer Engineering.

The author team includes Ph.D. students Ken Xu, David Oygenblik, Runze Zhang, and Mingxuan Yao; Research Scientist Muhammad Ibrahim; and Associate Professor Brendan Saltaformaggio. Their work contributes to a broader understanding of how emerging AI technologies interact with existing security and privacy frameworks.

As local AI assistants become more prevalent on mobile devices, the need for updated forensic tools and clearer privacy standards becomes increasingly urgent. The researchers suggest that this work could inspire further investigation into memory management practices in AI applications and potentially lead to more secure deletion protocols.

The study also aligns with ongoing discussions about the ethical and legal responsibilities of AI developers regarding data retention and deletion. It highlights the importance of transparent privacy policies, especially for features marketed as private or local.

While ORISA is currently a research tool, its existence signals a shift in how digital evidence is perceived. In a future where AI is deeply integrated into personal devices, the boundaries between user-deleted data and recoverable history may continue to blur, necessitating new approaches to digital privacy and security.