Your ChatGPT Privacy Toggle May Have Reset. You Won't Get a Warning.

If you or anyone on your team turned off "Improve the model for everyone" in ChatGPT, go check that setting today. On September 10, users on Hacker News (416 points, 169 comments) reported that the toggle reverts to on after app updates — quietly, with no dialog and no email. OpenAI hasn't addressed the behavior or said how many accounts were affected.

That's the immediate, do-it-now item. The bigger lesson is less convenient: a privacy setting is not a permanent decision. It's a state that has to survive every update, every re-login, every policy change, and you're the one who has to keep checking it, because the product won't tell you it slipped.

The same week gave the concept a second, stranger data point. Simon Willison flagged a controversy around the Navier-Stokes Millennium Prize problem: two mathematicians who'd been working the problem inside ChatGPT found that OpenAI's own agents solved a related piece of it days later, and started asking whether their session data had shaped that outcome. Willison's question is the one worth sitting with: if you use an AI tool to think through a hard, valuable problem, what happens to that thinking afterward — even if you never touched a training toggle at all? "Opted out of training" and "not retained anywhere" are two different guarantees, and most people who've flipped the switch assume they got the second one.

What to do differently Monday morning: stop treating the privacy toggle as a one-time setup step. Put a recurring check on your calendar — after any app update, not just once a quarter — and treat "does this tool learn from what I put in it" as a question you re-answer, not one you answered in onboarding. For any tool touching client work, competitive analysis, or anything you'd rather a competitor not read, know which of the two guarantees you actually have before you paste anything in.

Take a marketing manager at a 40-person SaaS company who uses ChatGPT to draft competitive positioning against a rival's new pricing page. She'd opted out of training months ago specifically because the analysis references unreleased roadmap details. If that toggle silently reset last week, her unreleased pricing strategy is now, by the product's own default, eligible to shape a model that her rival's marketing team can also prompt. She didn't do anything wrong — she trusted a setting to hold.

The counterpoint comes from a solo consultant handling client financials, who assumed a personal ChatGPT account was fine for a first-pass read of a client's cap table because "I'm not saving anything." That's a different failure than the toggle reset — it's using a personal account at all for work with confidentiality obligations, where the settings a consultant controls are looser than what an enterprise tier would guarantee by contract. Two different mistakes, same root cause: assuming the tool's default behavior matches what the situation actually requires, instead of checking.

Neither of these professionals needs to distrust AI tools wholesale. They need a five-minute habit: after any update, check the setting; before any sensitive session, know which guarantee you're actually operating under; and treat "I turned it off once" as the start of the job, not the end of it.

The toggle reset is fixable in thirty seconds once you know to look. The habit of looking is the part that has to be yours.