QWE AI Academy The #1 mistake after Tell HN: OpenAI keeps re-enabling the ‘allow training’ setting
The biggest mistake right now is flipping Improve the model for everyone off once, closing the tab, and assuming you’re done. That Tell HN thread (jacquesm, hundreds of points, large comment pile) is loud for a reason: people who wrote down the disable date later found the toggle on again.
I did the lazy version too. Thread hits → open ChatGPT → expect last quarter’s off setting to still be off. It wasn’t. What works instead: a short dual-path ritual you can repeat, plus controls that don’t live only in one UI panel.
What the training control actually controls
Personal Free / Plus / Pro: chats can train models until you opt out. That may have changed after a product update – re-check the live UI – but OpenAI’s own help article still frames it this way. You either turn off Improve the model for everyone under Settings → Data Controls, or you send Do not train on my content through the privacy portal. How your data is used says either path covers ChatGPT and regular Codex tasks. Only new conversations. Stuff already in a training run does not get magically removed.
History? Still there in the sidebar. Temporary Chat is the separate option: no history, no memories, not used for training, gone after about 30 days (abuse review can still apply) – per the Data Controls FAQ. Business, Enterprise, Edu, and API start off-training by default (Enterprise privacy); API feedback can still be an opt-in path, so don’t treat “API” as a free pass if you click feedback flows.
Pro tip: Skip thumbs up/down on anything sensitive. Same official “how your data is used” page: feedback may pull the whole related conversation into training even when the main opt-out is on.
Step-by-step: lock it down and verify it stuck
UI first. Portal second. Then prove they still match.
- Sign in at chatgpt.com. Logged-out chats ignore account toggles.
- Profile icon → Settings → Data Controls.
- Turn Improve the model for everyone off. Close the panel, open it again, confirm it stayed off.
- Mobile: sidebar → profile → Data Controls. Should already match; the setting syncs on the account, not per device.
- At the privacy portal, request Do not train on my content. Docs say portal opt-outs show up in account data settings; after the HN re-enable reports, a lot of people still do both on purpose.
- Codex users: open Codex Settings and check full-environment training there. ChatGPT Data Controls + portal do not move that switch.
- Note the date. Put a monthly 30-second Data Controls re-check on a calendar.
One sensitive prompt and you don’t want to think about globals? Start Temporary Chat from the model picker.
Common pitfalls that keep data flowing
The catch is the toggle is not a one-time contract. HN comments keep citing flips after updates, resubscribes, or TOS acceptance. Some EU reports say theirs stayed off for months. Others saw it return. “I set it last quarter” means nothing until you look.
- Feedback override: thumbs can still train that thread (see pro tip – don’t re-litigate it everywhere).
- Codex split brain: full-environment training is a different control.
- Logged-out sessions: no account, no opt-out.
- Safety retention: training off ≠ “OpenAI stores nothing.” Limited abuse/safety monitoring still exists; Temporary Chat’s ~30-day delete window is the documented special case.
Community habit worth stealing: after you flip it, re-open Data Controls from another device. If the switch lies, you’ll catch it in seconds.
How this stacks against the real alternatives
| Path | Training default | What you must do | Best when |
|---|---|---|---|
| ChatGPT Free/Plus/Pro | On unless opted out | Toggle + portal + re-check | Everyday personal use |
| Temporary Chat | Never trains | Pick it per session | One sensitive prompt |
| API / Business / Enterprise | Off by default | Avoid feedback opt-in paths | Work or code with secrets |
| Local / other hosts | Varies | Read that product’s policy | Any cloud training risk is unacceptable |
Keep getting bitten by re-enables? Move the sensitive work to API or a business workspace where the default already matches. Consumer toggle is handy. It is not something you set once and forget.
Weird question worth sitting with: if a privacy control can silently drift, how many other “I already handled that” toggles in your stack have the same failure mode?
FAQ
Does turning the setting off delete past chats from training?
No. New chats only. Past training runs stay put.
I turned it off and submitted the privacy form – why check again?
Because people in that live HN thread watched the UI come back on after a clean disable. Picture this: you opt out in January, ignore Data Controls until a model launch in the spring, then paste client text into a normal chat. Thirty seconds of re-checking beats that surprise. Docs still treat toggle and portal as valid ChatGPT/Codex-task paths and say older form/support opt-outs continue to be honored – practice still says open the panel after plan changes or big UI updates.
Is the privacy portal stronger than the in-app toggle?
OpenAI’s wording: either is enough, and a portal opt-out should appear in account data settings. Community threads push both when trust is thin. That still leaves three hard limits people blur together. (1) Neither path rewinds history already used for training. (2) Neither is the Codex full-environment switch – you set that in Codex Settings. (3) Neither is a promise that logged-out use, thumbs feedback, or safety monitoring behave like a full “never store / never train” mode. When the stakes clear a consumer toggle, use Temporary Chat for the one-off or move the work to Business/API defaults.
Open Data Controls now. Confirm the toggle. Hit the portal if you never did. Calendar the re-check. Done.