Does ChatGPT Free's opt-out actually stop training on your prompts?

Yes, it works. Turn off "Improve the model for everyone" in ChatGPT's Data Controls settings and your new conversations stop feeding training. OpenAI confirms this directly in its own Data Controls FAQ, last updated three days before this article was written. But the toggle answers a narrower question than most lawyers think they're asking, and the gap between what it does and what people assume it does is exactly where firms get into trouble.

What the toggle actually stops

Training is on by default for ChatGPT Free, Plus, and Pro personal accounts. Go to Settings, then Data Controls, then switch off "Improve the model for everyone," and every conversation after that point is excluded from training runs. OpenAI states this applies account-wide and syncs across web and mobile immediately, no per-device setup required. That part is genuinely useful and worth doing regardless of what else this article says.

Here's the limit nobody puts in the headline: the opt-out is forward-looking only. If you sent a client's deposition summary through ChatGPT last month with training left on, that conversation may already sit inside a completed training run. Flipping the switch today does not claw it back. There is no retroactive undo button, and OpenAI's policy language never claims there is one.

Opting out of training is not the same as deleting anything

This is where the confusion actually costs firms something. Turning off training does not delete your chat history. It doesn't even hide it. Per OpenAI's own Data Controls FAQ, "your conversations will still appear in your chat history but won't be used to train ChatGPT." The data still exists on OpenAI's servers, still shows up when you log in, and is still subject to whatever retention and safety-review policies apply regardless of your training setting.

If you actually want a conversation gone, that's a separate action: delete the specific chat, or use Temporary Chat mode, which OpenAI says is deleted from its systems within 30 days and never touches training or memory in the first place. Three different settings, three different outcomes. Treating them as one setting is the mistake.

The part that should worry a lawyer more than training ever did

In May 2025, a federal court presiding over The New York Times' copyright suit against OpenAI ordered the company to preserve all ChatGPT output logs that would otherwise be deleted, including logs generated by users who had explicitly requested deletion. District Judge Sidney Stein affirmed the order that June, rejecting OpenAI's argument that user privacy interests should override the preservation obligation. For roughly four months, a user's deletion request and a user's training opt-out meant nothing against a litigation hold that a court had imposed on the company itself.

That specific order ended. OpenAI's preservation obligations under it concluded on September 26, 2025, and the company returned to its standard practice of deleting conversations and Temporary Chats within 30 days of a user's deletion request. But the order happened, it was litigated up to a district judge, and it resolved in favor of preservation over privacy for four months. Any lawyer weighing whether opt-out settings are "enough" protection for client material should sit with that fact for a second: a privacy toggle is a company policy. A court order is not bound by that policy at all. Nothing in a user's account settings can override a preservation order OpenAI itself is legally compelled to follow.

Where the free and paid tiers genuinely diverge

For ChatGPT Team, Business, and Enterprise, OpenAI states plainly that it does not train on inputs or outputs by default, full stop, no toggle required. That default-off posture is a contractual term of the business agreement, not a setting a single user can accidentally leave on. Enterprise and Team accounts also come with the option of a proper data processing agreement, something Free and Plus accounts don't offer at all. If your firm is routing any client-identifying material through ChatGPT with any regularity, the jump from Free or Plus to Team is the actual fix, not a deeper dig through the Free settings menu.

We've written before about what a defensible AI setup looks like for a solo or small firm in our confidentiality checklist, and about the broader question of whether ChatGPT is safe for legal work at all in our earlier piece on ChatGPT and legal work. Both assume you understand the difference this article covers: training, deletion, and litigation exposure are three separate questions, and a lawyer who only answers the first one hasn't actually assessed the risk.

What to actually do with this

• Turn off "Improve the model for everyone" in Settings, Data Controls, on every account anyone at your firm uses. It's free, it's immediate, and there's no reason to leave it on.
• Don't treat that toggle as a substitute for keeping client-identifying details out of ChatGPT Free or Plus entirely. It stops training. It does not stop retention, and it does not stop a court from ordering OpenAI to preserve the data anyway.
• Use Temporary Chat for anything you want gone within 30 days and never touching memory or training in the first place.
• If client matters are routinely going into ChatGPT, move the firm to Team or Enterprise and get the DPA in writing. A settings toggle is not a substitute for a contract.

The opt-out is worth using. It's just not the confidentiality control most lawyers think it is, and the NYT preservation order is the clearest evidence available that a company's privacy settings and a court's authority over that company's data are two entirely different things.

For the full framework on what counts as reasonable precautions with client data and AI, get The Lawyer's AI Starter Kit, free. To see which tools actually offer contractual data protection out of the box, check the best AI tools for lawyers.

FAQ