If you untick a box that keeps your AI chats out of a company's training data, you'd expect it to stay unticked. A Hacker News thread that pulled in 134 points in about two hours is a good reminder that it doesn't always work that way — and that's why so many teams are suddenly searching for how to protect their data from OpenAI training. A developer wrote that their company is "pretty paranoid about leaking IP to the frontier labs through prompts," so everyone had been careful to switch off data sharing in Codex, Claude Code and similar tools. Then, one day, the setting in Codex was back on. Nobody remembered opting in again.
What happened
The post landed on Hacker News with 46 comments and climbing. The author wasn't reporting a hack or a breach. They were reporting a checkbox that had quietly flipped back to the default they'd deliberately turned off.
The tricky part is that nothing about the tool changes when this happens. Codex keeps working. Your prompts still run. There's no banner, no email, no notification. The only way to catch it is to go look — which is exactly what most people never do after the first time they set it.
The two links doing the rounds
The original poster shared where to check: chatgpt.com/settings/data-controls, and look for the toggle labelled "Improve the model for everyone." Later edits added a second place to switch sharing off: OpenAI's privacy portal. The poster also linked an earlier thread where someone had flagged the same thing, which suggests this isn't a one-off.
Worth saying plainly: this is a user report, not an OpenAI announcement. No official explanation has been posted in the thread.
Plain English: what "training on your chats" actually means
Every time you ask an AI a question, you're handing it text. Model training is the process of using text like that to make the next version of the model smarter. Think of it as notes passed in class: if someone collects enough of them, they can write next year's textbook from what students actually asked.
Most AI companies give you a switch to say "don't use my notes." The catch is that the switch is a setting on an account, not a physical lock. It can be changed — by you, by a bug, by a policy update, by a rollout to a new interface. If you never look back, you'll never know which way it's pointing.
The comment that explains why workarounds backfire
One commenter half-joked about stuffing a taboo sentence into every prompt so labs wouldn't want to train on it. A reply from user AlanYx poured cold water on the idea: get flagged often enough and your account gets quietly tagged, and the contractual confidentiality restrictions that normally protect you stop applying. Their evidence was a report called "Detecting and countering misuse of AI: September 2026" from Anthropic. They argued the same pattern likely holds at OpenAI, because OpenAI models hosted on Azure run documented "asynchronous" filters that can flag risk across conversations. They also noted that recent Microsoft contracts removed the specific retention window for flagged content. As they put it, this is one plausible route by which something like a Navier-Stokes-related leak could happen even with training switched off.
That's a commenter's read, not a confirmed fact. But it's a useful warning: clever tricks to dodge training are worse than simply switching it off.
What it means for you
You probably don't work at a frontier lab. You still paste things into AI tools you wouldn't want published. Here's what this thread actually changes for ordinary users.
At home
You've probably used a chatbot to draft a complaint to your landlord, describe a medical symptom, or untangle a family argument. That text is personal even when it feels throwaway. A toggle that flips itself back on means last year's private questions could end up as raw material. Checking once a month is cheap insurance.
At work
This is where the thread started. If your job involves code, contracts, customer lists or unreleased plans, a prompt containing them is a copy of company data leaving the building. Most people set the toggle once during onboarding and never touch it again. The poster's whole team had done exactly that — and still had to catch it manually. If you want to see what free tools can do before you decide what's safe to paste, there's a browsable set at MyKreaTool.
Running a business
If clients trust you with their information, an auto-flipped switch is a compliance headache, not just an annoyance. You can't promise "we never share your data" on the strength of a checkbox you haven't looked at since Q1. The practical fix is boring: one named person checks the settings for every AI tool your team uses, on a schedule, and writes down the result.
Studying
Students paste essays, lecture notes and research questions into chatbots constantly. Most of that is harmless. Some of it — a draft thesis chapter, an unpublished dataset, an interview transcript — isn't yours to hand over. The same two-minute check applies, and it costs nothing.
Creative projects
Writers and designers worry about this the most, and rightly so. An unfinished manuscript or a brand concept is valuable precisely because nobody has seen it yet. The awkward truth is that AI tools are genuinely useful for brainstorming — so the goal isn't to stop using them, it's to know which switch is on.
Side income
Freelancers often work for several clients on one personal account. One mis-flagged prompt can leak pieces of three projects at once. Turning off training doesn't make you anonymous, but it does keep your prompts out of the training pipeline, which is the specific thing this thread is about.
How to try it right now
This is a five-minute job, and the free route comes first.
1. Sign in to your free ChatGPT account and open chatgpt.com/settings/data-controls.
2. Find the row labelled "Improve the model for everyone." That's the exact wording from the thread — not "share data," not "analytics." Read it carefully.
3. If it's on, switch it off. Then reload the page and confirm it stayed off. That reload is the step people skip.
4. Check the second control. Open OpenAI's privacy portal and review what's set there. The original poster added this route specifically because the first setting alone may not cover everything.
5. Repeat for every AI tool you use. Codex, Claude Code, whatever's in your browser. Settings don't carry across products, and they don't carry across accounts.
6. Do it on work accounts too. If your company has a team plan, the person who owns the billing should be the one checking — not five people guessing.
7. If your data is genuinely sensitive, a checkbox isn't enough. One commenter's blunt take: if a company really cared, they'd run their own inference instead of trusting "a couple of checkboxes." The reply was equally blunt — self-hosting means roughly $30k a month for a server plus the job of training a 2-trillion-parameter model. A middle path several commenters recommended: gated models inside Azure, GCP or AWS, which can't dial out unless you grant permission.
Upsides and what changes
Flipping the switch doesn't make the model dumber or lock features away — it's a data-use setting, not a performance dial. What changes is what happens to your text after you hit send.
The real upside is the habit. This thread got 134 points in two hours because a lot of people recognised themselves in it. Teams are now auditing settings they'd forgotten existed, and businesses are asking sharper questions about retention before they sign. That pressure is why zero data retention options exist at all, and why enterprise buyers can now demand terms that individual users can't.
Limitations
Be honest about what a toggle can and can't do. It only covers what the company says it covers, in the products it says it covers, and you have no way to independently verify it — you're trusting a status light. Commenters in the thread argued that flagged content may be retained regardless of your setting, and that zero data retention is an enterprise feature rather than a default. Nobody in the thread produced an official explanation for why the Codex setting came back on, so treat the whole episode as a strong prompt to check rather than proof of anything darker. And if your obligations are contractual or regulatory, a switched-off checkbox is not a compliance program — you need the paperwork and the retention terms behind it.
Conclusion: Your One Action for Today
Open chatgpt.com/settings/data-controls, find "Improve the model for everyone," and check which way it's pointing. If it's on, turn it off, reload the page, and confirm it stuck. Then do the same for every other AI tool you use for work — today, while you're thinking about it, not next quarter.



Comments 0