Most ChatGPT users I talk to think their delete button works.
You finish a conversation. Maybe it was about your relationship, your finances, a health concern, a question you'd never ask out loud. You delete the chat. You move on.
Here's what actually happens to that conversation.
In May 2025, a federal magistrate judge ordered OpenAI to preserve every ChatGPT conversation, including the ones users had explicitly deleted. The order came out of the New York Times' copyright lawsuit against OpenAI, where the plaintiffs argued that user conversations might contain evidence of how the model was trained on copyrighted material.
The court agreed. OpenAI had to suspend its standard 30-day deletion policy. Conversations users had marked "temporary." Conversations users had cleared from their history. All of it had to stay, indefinitely, on OpenAI's servers.
In January 2026, a separate ruling went further. U.S. District Judge Sidney Stein upheld an order requiring OpenAI to produce a sample of 20 million chat logs to the plaintiffs. The users whose conversations were included received no notification. They had no opportunity to object. Most of them, even now, have no idea their conversations are part of that sample.
A portion of the preservation order was eventually narrowed in September 2025. But every conversation logged before that cutoff is still archived. If you used ChatGPT in 2024 or in the first nine months of 2025 — that's most ChatGPT users on earth — your "deleted" conversations are still on a server somewhere, waiting to be searched if a court asks.
This isn't a privacy violation in the legal sense. OpenAI didn't break a rule. The rules just don't protect you the way you think they do.
What "delete" actually means
When you talk to a lawyer, what you say is protected by attorney-client privilege. When you talk to a doctor, you're protected by HIPAA. When you talk to a therapist, by therapist-patient confidentiality. These protections exist because society decided some conversations are worth shielding from disclosure even when courts come asking.
When you talk to ChatGPT, you have none of those protections.
Sam Altman, OpenAI's CEO, has said this out loud. In a podcast appearance last year, he acknowledged that people are using ChatGPT "like a therapist, a lawyer, a priest" — without realizing those conversations carry no legal privilege whatsoever. He called for new laws to create something like AI-user privilege.
The laws don't exist. As of right now, in May 2026, anything you've ever typed into ChatGPT is potentially discoverable in litigation. Your delete button is a UI affordance, not a legal one. It removes the chat from your visible history. It does not remove it from OpenAI's servers when a court orders preservation.
Why this is structural, not a one-time problem
The thing that bothers me about this story isn't the court order. Courts will do what courts do. The thing that bothers me is the architecture that allowed the problem in the first place.
ChatGPT was built so that OpenAI retains every conversation by default. The retention isn't incidental; it's how the product works. The conversations are needed for moderation. They're useful for model improvement. They're valuable for analytics. They're functionally part of OpenAI's business — which means they exist, in a database OpenAI controls, even when you've clicked delete.
Once those conversations exist somewhere a company can access them, three things become true:
- The company can be subpoenaed for them.
- The company can change its retention policy.
- The company can be hacked, leaked, or have its data exposed in a breach.
All three have happened, to ChatGPT specifically, within the past 18 months. The court order I described above is the subpoena scenario. OpenAI has updated its memory and data retention policies multiple times in the past year. And in 2025, a single security incident exposed roughly 300 million AI chat messages across providers.
You can opt out of training. You can use temporary chat. You can delete your history. None of it changes the underlying architecture. Your conversations exist on a server you don't control, in a system whose business model depends on retaining them, governed by policies that can change, in a legal landscape where any of it can be ordered into evidence at any time.
What "actually private" looks like
There's a different architectural model, and it's worth describing in plain terms because most people have never seen it explained.
A genuinely private AI doesn't retain conversations on a server you don't control. The conversations are encrypted in a way that even the company itself can't read. There is no admin panel where someone at the company can search your history. There is no log file that contains the words you typed.
When the company gets subpoenaed, there's nothing to hand over — not because they're refusing, but because there's nothing to hand over.
That's a fundamentally different threat model. With the major AI products, you're trusting the company's promise that they won't misuse your data, and trusting that no court order will override the promise, and trusting that they won't change the policy, and trusting that no one will hack their servers. With a structurally private product, you don't have to trust any of those things, because the architecture itself prevents the failure modes.
This is what we believe AI should look like. Your conversations are encrypted end-to-end. We can't read them. We can't sell them. We can't be subpoenaed for them, because there's nothing on our servers that matches what you typed. The protection is structural, not promissory.
What to actually do
If you've used ChatGPT for personal things — and most people have — there are a few practical moves to make right now:
Stop typing the personal stuff into a system you don't control. The simplest privacy decision is the one upstream of all the settings: don't put it there in the first place.
Use a tool with a different architecture for the parts of your thinking that are actually personal. The decisions you wouldn't email to a stranger. The questions you'd never ask in front of HR. The reflections you'd want shielded if a court ever came asking.
If you keep using ChatGPT, the practical guide to using AI better covers the settings most people don't know about — including how to opt out of training and what temporary chat actually does.
Try Blob free. No credit card. The architecture is the product.
Your delete button should mean something. Build the version of your AI life where it does.