News

Deleted AI chats can leave traces in Android memory, Georgia Tech tests show

A deleted conversation with an AI assistant that runs on an Android phone may not disappear from the device completely. Researchers at Georgia Tech built an experimental forensic tool called ORISA that, in controlled tests, recovered remnants of deleted local AI conversations and used them to rebuild a working session, according to the university’s research news page and subsequent reporting.

The qualification matters as much as the finding. ORISA is described in the reporting as a forensic tool for investigators, not an app that phone owners can install to retrieve their own chats. It needs access to device data or a copy of the handset’s memory and storage. Nothing in the available reporting describes a breach, an attack, a product vulnerability or a public release of the tool.

A controlled test, and a partial result

In the demonstration described by Georgia Tech, the team created a fictional assistant called Project Note Keeper with the ChatterUI Android app. They gave it 100 fictional facts about a project, then deleted the conversation. By that point the original session could actively use only 14 of the facts, because they were no longer within its active memory. ORISA recovered enough material to create a new working session, which returned 26 of the facts when asked.

That result is easy to overstate. It showed that the researchers could recover information the assistant had already stopped using, which is more than the live session could reach at the moment of deletion. It also showed substantial loss: 26 of the 100 facts came back in the reported result. Digital Trends described the outcome as partial recovery, and the reporting gives no per-app or per-model figures.

The researchers also tested ORISA across 60 configurations involving 10 Android apps and three AI models, reporting that the tool recovered conversation information that remained in phone memory. The specific apps and models are not named, so the finding cannot be mapped to particular products or every local AI app.

Why deleting is not the same as erasing

Local AI assistants process conversations on the handset rather than sending them to a cloud service. That keeps data away from remote servers, but it does not make the phone’s working memory a clean slate. While a local model is running, it holds intermediate data. ZDNET reported that the researchers describe ORISA as using batch tokens and key-value caches preserved in process memory to reconstruct token sequences and restore session-level attention states.

The app can remove a conversation from view without necessarily clearing every remnant that the model used while it was active. ORISA searches those remnants and stitches together enough context to restart a session. PhoneArena noted that forensic recovery from memory is a long-established technique; the newer point is that local AI chats leave traces of this kind on the device.

Recovery is also time-sensitive. Georgia Tech and ZDNET both report that the information becomes harder to retrieve as other activity overwrites the phone’s memory. That is a description of forensic difficulty, not a deletion setting that users can rely on to erase data.

Who can actually use it

The access requirements are the strongest limit on the finding. Android Authority and Digital Trends describe a tool that needs an examination of the device or a copy of its contents, not a remote connection. That places it in the hands of an investigator examining a device, not an ordinary user hoping to recover a chat.

Scope is limited in another direction. Android Authority reported that the technique does not establish that conversations from cloud-based chatbots are similarly recoverable. The research concerns local processing and its remnants in phone memory. It does not compare the privacy of local and cloud AI, and the available reporting does not quantify a benefit for either approach.

What the evidence does not show

The underlying paper, Recovering and Rehosting Mobile Local LLM Conversations and Contexts via Memory Forensics, was presented at the 2026 IEEE Symposium on Security and Privacy, according to Georgia Tech. The technical details rest on Georgia Tech’s summary and subsequent reporting, not a review of the paper. No independent replication or field study appears in the available material. The available evidence does not show that ORISA is publicly available, commercially sold, actively used by investigators or used by criminals.

ZDNET raised a hypothetical in which attackers who recreated the technique or obtained the tool could gather deleted information from chats and recreate queries with it as context. The publication presented that as a theoretical scenario, not an observed campaign. The research reports no threat actor, no breach and no exploited vulnerability.

Ken Xu, identified by Georgia Tech as a PhD student and lead author, is reported as saying the tool can help establish what information someone shared with an AI assistant and what assistance they received. That is a forensic framing: the value lies in reconstructing what happened during an examination, not in routine surveillance of phone users.

What it means for local AI users

The practical privacy point is narrower than full recoverability, but it goes beyond one app’s delete button. If a conversation was processed on the device, deleting it through the interface may leave enough in working memory for a forensic examination to recover some of its content, at least in the configurations tested and before overwriting. Whether that matters depends on who obtains the device and when.

Local processing still changes the exposure path, because the conversation is not being sent to a remote service as part of the session. The reporting does not establish that cloud chatbots are equally recoverable by this method, nor does it establish that local AI is safer overall. The two arrangements retain and expose data in different ways, and the available reporting does not rank them.

ZDNET’s long-standing advice is not to share sensitive information with chatbots, even in a local session. That is editorial guidance rather than a tested control. The research supports a more specific expectation: deleting a local conversation removes it from the interface, but in the tested configurations it was not a guarantee that every trace had gone.

The open questions are concrete. Which apps and models were tested, and do they behave the same way outside a controlled demonstration? Can other researchers reproduce the 26 of 100 result? What would a deletion process need to do to clear the working memory forensic tools examine? Until those are answered, the defensible conclusion is that app-level deletion and reliable erasure are different promises, and only the first has been demonstrated in this research.

An empty conversation list is a statement about the interface, not about the device, and the controlled test here turns that gap into a question worth sitting with: what does a deletion actually remove, and what evidence would show it? Cybersecurity fundamentals give that question a workable shape, covering how data persists across a system after it leaves view, why a staged demonstration differs from a field result, and how to read a partial recovery without rounding it up into full erasure. With those habits, the next forensic claim becomes easier to weigh for its conditions and its limits rather than its most striking number. If that appeals, The Hack Academy’s online cybersecurity training programme lists cybersecurity fundamentals among its subject areas, with self-paced learning and practical exercises alongside the theory.

Leave a Reply

Your email address will not be published. Required fields are marked *