Are ChatGPT Conversations Private? The OpenAI Court Order and What Writers Must Know in 2026

There is a conversation I had with ChatGPT last summer that I genuinely do not want anyone else to read. It was not anything scandalous. It was a long, rambling, embarrassingly personal brainstorm about a novel I was trying to write — one of those sessions where you tell the model things about your characters that you would not say out loud, because the characters are clearly you with the serial numbers filed off. I deleted that chat the next morning. Or I thought I did.

In June 2025 a friend forwarded me a Malwarebytes article with a subject line that ruined my afternoon. A federal judge had ordered OpenAI to preserve every ChatGPT conversation log, retroactively, including the ones users had already deleted. The order applied to the entire window from roughly mid-April 2025 onward. My embarrassing brainstorm sat squarely inside it.

I am writing this in late May 2026, almost a year after that ruling, with a more complete picture than anyone had at the time. The short version is that the conventional wisdom most writers operate on — “I deleted the chat, so it's gone” — has not been true for over twelve months. If you are using ChatGPT to brainstorm, draft, or edit creative work, this is the privacy story you need to actually understand, not skim.

How a Copyright Case Became a Privacy Story

The case started as a straightforward copyright dispute. The New York Times alleged that OpenAI had trained on its articles without permission and that ChatGPT could be coaxed into reproducing those articles verbatim. To prove the second part of that claim, the plaintiffs needed evidence. So they asked the court for it.

Here is the timeline stripped of legal jargon:

• May 13, 2025: Magistrate Judge Wang issues the preservation order. OpenAI must keep every conversation log, including deleted ones, until further notice.
• June 2025: OpenAI publicly calls the order a “privacy nightmare.” CEO Sam Altman says it sends the wrong signal to users worldwide. The company appeals.
• Late September 2025: The preservation order is lifted. OpenAI resumes its standard 30-day deletion policy for new conversations. But everything generated between mid-April and late September remains in storage.
• November 12, 2025: Judge Sidney Stein upholds an order forcing OpenAI to hand over 20 million de-identified conversation logs to the plaintiffs as part of discovery.
• January 2026: The first wave of those logs is delivered to the New York Times legal team. Sampling, de-identification, and analysis begin.
• Today (May 2026): The case is still active. Any conversation from the preserved window is still on a server, still potentially discoverable, and still completely outside the user's control.

None of this is conspiratorial. All of it is documented in court filings. The reason it matters for writers is that the precedent has been set: if there is a credible legal interest in your conversations, they can be preserved and produced. ChatGPT logs are now a category of legal evidence.

The Numbers Most Writers Have Not Seen

A 2026 survey of US AI users captured how people actually feel about this once they hear it. The headline numbers, all from independent polling:

• 51% of users said they would be “much more likely” to consult a human lawyer or therapist instead of ChatGPT if they knew their conversations could be subpoenaed.
• 76% of users said the government should regulate AI companies to provide a form of legal privilege for user conversations, similar to attorney–client privilege.
• Roughly 20 million de-identified logs were ordered into discovery in November 2025 — a small slice of the total preserved volume.
• Zero existing US laws grant ChatGPT users any form of legal privilege over their conversations.

That last number is the one to sit with. When you talk to a doctor, a lawyer, or a therapist, the law assumes a privacy interest and protects it. When you talk to an AI model, the law assumes you typed into a website and nothing more. If you would not paste something into a public Google Doc, the conservative reading right now is that you probably should not paste it into ChatGPT either.

Why Writers Are Caught in the Net Hardest

I have been talking to working writers about this for months — novelists, journalists, screenwriters, ghostwriters — and a pattern keeps showing up. Writers do not use ChatGPT the way the average person does. We use it like a private notebook. We dump entire drafts in for restructuring. We brainstorm character flaws nobody else knows about. We feed it confidential client work to summarize. We type out conversations we would never put in an email.

If you are a working writer, your typical ChatGPT history from the last year probably contains some combination of:

• Unpublished chapters of a novel under contract.
• Personal essays that were not yet ready to share.
• Client work covered by NDAs (the most legally fragile case).
• Memoir material involving real, identifiable people.
• Pitches and book proposals not yet sold.
• Therapy-adjacent journaling you did because it felt private.

Every single one of those categories is now sitting on a server inside the preservation window. For the NDA work in particular, this is not just a personal embarrassment risk — there is a real contractual exposure if discovery ever pulls a sample that includes client text. We covered the broader scraping problem in our deep dive on whether OpenAI is reading your novel, but the preservation story is a different beast. Scraping is about training data. Preservation is about your specific conversation, still readable, sitting on a disk waiting for a subpoena.

What Actually Happens to Your Text in Each Tool

The honest answer to “is this private” depends entirely on the architecture of the tool you are using. Marketing copy is not architecture. Here is what the major AI writing tools actually do with your text, based on their own published documentation and ToS as of May 2026:

For a fuller breakdown of which tools have which retention defaults, see our audited list of AI writing tools that don't train on your work.

7 Things to Change in Your Writing Workflow This Week

This is the list I went through myself after I stopped being upset about the preservation order and started being practical. None of it requires giving up AI assistance. It just requires being deliberate about which AI and what you feed it.

1. Audit your ChatGPT history. Open your conversation list and skim the titles. Anything that mentions a real client, a real identifiable person, an unpublished book, or an NDA project, write down where it appeared and roughly when. You probably cannot remove it, but you should know it exists.
2. Turn off “Improve the model” in settings. In ChatGPT, go to Settings → Data Controls and disable model training. This stops future conversations from being added to training pipelines, even though it does not retroactively remove anything from the preservation window.
3. Stop dumping full manuscripts into AI tools. If you must use an AI assist for a sensitive draft, use short, anonymized excerpts. Strip out names, locations, and identifying details before pasting.
4. Move client work off consumer chatbots entirely. If you ghostwrite, do agency work, or write under NDAs, none of that text should be in a consumer ChatGPT account. The contractual exposure is real. Use either a fully offline tool or a zero-knowledge encrypted environment.
5. Pick one privacy-first tool and centralize there. The biggest risk is having drafts scattered across five different AI tools, two browser extensions, and a Google Doc. Consolidation makes a real difference. We compared the strongest 2026 options in our guide to the best free book writing apps and the best distraction-free writing apps.
6. Back up everything you actually care about. The preservation order is one kind of risk. A breach or a forgotten password is another. Make sure the work you care about exists in at least three places. Our writeup of the 3-2-1 backup rule for novelists walks through the routine.
7. Treat “zero-knowledge” as a contract term. Before you trust a tool with serious work, read the privacy policy. Look for the literal phrase “zero-knowledge” or “end-to-end encrypted” and a description of where the encryption key lives. If the provider holds the key, the provider can produce the data. If only you hold the key, they cannot. We explain the full model in our zero-knowledge encryption guide for writers.

The Honest Take

I do not think ChatGPT is evil. I still use it for some kinds of work — mostly throwaway research, structured reasoning that does not involve any of my own creative material, and code that I would happily put on GitHub anyway. The preservation order did not make ChatGPT useless. It made it specific. The tool is fine for some categories of writing and dangerous for others, and we now have a year of court records to draw the line confidently.

The thing that genuinely changed my workflow was realizing that “private brainstorming” is not a category ChatGPT can serve anymore. The infrastructure does not support it. If I want to think out loud about a character's addiction, or process a real conflict from my life through the safety of fiction, I need a tool whose provider cannot read what I wrote, by mathematical design. That is what zero-knowledge encryption means: not a promise, not a policy, but cryptography that makes the data unreadable to anyone but me.

The good news is that the tools to do this now exist and are genuinely usable. The bad news is that most writers have not realized they need them yet. If this article is the nudge, I will count that as a fair outcome.