"Delete" on your AI chats may just mean "hide", and app makers owe you honesty

You tap the trash icon, the conversation vanishes, and you move on. Done, right? Not necessarily. For AI chats, "delete" is starting to look like a polite word for "stop showing me this."

Here is what the reporting says. As Hipertextual reports, researchers have shown that chats deleted from AI assistants on Android can remain recoverable. Other outlets identify the tool as ORISA, built by a team at the Georgia Institute of Technology. It targets assistants that run on the phone itself and, according to the coverage, restored deleted conversations as a new session.

One detail stands out. Ecosistema Startup says ORISA recovered 26 data points from a conversation the user had already deleted. Before deletion, the assistant itself could reach only 14 of them, because the rest had fallen out of its context window. Another report points to the phone's working memory as where those traces live. We only have summaries of the research, not the paper, so details are thin.

"Private" is not the same as "gone"

Running an assistant locally sounds like the privacy-friendly option, and in one sense it is. Your words don't travel to a server. But as Digital Trends puts it, that data stays in the phone's working memory while the app runs, and deleting a chat through the interface doesn't necessarily wipe it from memory.

We think this is the real lesson. People have learned to think of privacy as a question of where data goes, cloud or device. This research is a reminder that a second question matters just as much: how long it stays, and who can get at it. A local chat is safer from a company's servers. It isn't automatically safe from someone with your unlocked phone, a forensic kit or a very bad day.

To be fair, the scenario looks narrow. These are research findings about specific setups, not proof that every chatbot app leaks everything. Don't panic. But the finding seems plausible to us, because it matches how storage often behaves: deleting frequently removes the pointer, not the contents.

Companies have taught us a vocabulary that flatters them

The cloud side isn't obviously cleaner. Per the coverage, companies like OpenAI and Anthropic keep chats on their servers for legal reasons and remove them after a while. That sounds like a reasonable policy to us. It just isn't what a user pictures when they hit delete.

So we have two quiet gaps: the button says one thing, the system may do another. Our view is that app makers should stop using one word for three different actions: hiding it from your list, erasing it from the device, and purging it from the servers. A short, honest line under the button ("removed from your history now; copies may remain for up to X days") would cost almost nothing. We'd call it table stakes for a product people use to think out loud.

There is also a trust angle. Another story this week, reported by Infobae, says OpenAI will use Daybreak Blue sessions on personal accounts for training from November 29, even for users who opted out of model improvement. That is a different issue, and it affects a narrow group, but in our view it feeds the same suspicion: users keep discovering that the controls mean less than they assumed.

Treat AI chats like your banking app

What should you do? Not stop using assistants. Change the mental category. People paste in medical worries, work documents, drafts of difficult messages and sometimes passwords. That is sensitive data, so handle it like sensitive data.

• Don't put in anything you would be devastated to see recovered.
• Lock the phone properly and keep it updated.
• Prefer assistants that spell out retention in plain language.
• Don't treat "delete" as a security measure. It is housekeeping.

We'd also say that anyone selling a local assistant as the private option should explain what happens in memory and storage once you hit delete. If they can't, the privacy claim is half a sentence.

The verdict: the trash icon tidies your screen; it doesn't promise erasure. Until app makers say clearly what it does, assume a deleted chat is a hidden chat, and write accordingly.