Someone pastes a client’s contract into ChatGPT to get a summary. Someone else uploads a spreadsheet with staff salaries. A third types out a medical question they haven’t mentioned to their family. None of them think of it as publishing, but every one of those keystrokes lands on a server in another country.
That isn’t a scandal. It’s just how the product works, and the details are more mundane, and more interesting, than the panic threads suggest. Here’s what actually happens to your data, which switches change it, and where the real risks sit.
What ChatGPT stores, and for how long
Every conversation you start is saved to your chat history, tied to your account, and kept until you delete it. Delete a chat and OpenAI says it’s removed from its systems within 30 days, though copies can linger longer in backups and logs. Prompts and responses are also held for up to 30 days for abuse monitoring, where human reviewers sometimes read them. That’s standard practice across large platforms, not a ChatGPT peculiarity.
Two things catch people out. First, the share button. Hit it and you get a public URL. Anyone with the link can read the entire thread, including anything identifying in it, and it can turn up in search results once someone posts it somewhere. Second, the March 2023 bug. A caching flaw let some users see the titles of other people’s conversations, and roughly 1.2% of Plus subscribers had payment details exposed for a few hours. It was fixed quickly, and it’s a fair reminder that no company gets this perfect.
Training: the setting to check first
In many regions, your conversations can be used to improve the models by default. That’s the trade-off the product is built on, and for plenty of people it’s fine. It stops being fine the moment you paste in something you don’t own or something under an NDA.
Turning it off
In the web app, open Settings, go to Data controls, and switch off “Improve the model for everyone.” Mobile has the same option in the same place. The change applies to new conversations. Anything already swept into a training run won’t be pulled back out.
Temporary Chat
There’s also Temporary Chat, which starts a thread that won’t appear in your history, won’t be used for training, and won’t be remembered. It’s the right tool for a one-off sensitive question. It isn’t a cloak of invisibility: standard abuse monitoring still applies, and the text is still transmitted and processed. What it removes is the paper trail on your account.
Memory is the part people underestimate
ChatGPT can carry facts between conversations: your job, your kids’ ages, the fact that you’re training for a half marathon in October. Genuinely useful, and it quietly turns a string of unrelated questions into a profile. What ChatGPT memory actually remembers about you is worth reading in full, but the short version: you can review every stored memory, delete individual entries, or switch the feature off entirely under Settings, then Personalization.
Memory is not the same as training. Turning off training doesn’t clear memories, and clearing memories doesn’t stop your chats being stored. Separate switches, and you may well want them in different positions.
Voice mode and uploaded files
Talk to ChatGPT and it processes audio, and depending on your plan and region, clips may be retained for a period for safety review. Transcripts land in your history exactly like typed text. If you use it to think out loud about a work problem on a commute, assume the transcript exists afterwards. There’s a fuller breakdown of where speech input is strong and where it mangles things in this look at ChatGPT voice mode.
Uploads deserve their own warning. A PDF, a screenshot, an audio file: whatever is inside gets processed and, in most cases, stored alongside the conversation. Screenshots are the worst offender, because people forget about the browser tabs and notification banners in the background.
The settings worth changing today
- Data controls: turn off model training if you use ChatGPT for anything work-adjacent.
- Chat history: delete old threads in bulk rather than one by one, starting with anything containing client names.
- Memory: open the memory list and actually read it. Delete what’s stale or simply wrong.
- Shared links: find chats you’ve shared publicly and revoke the ones nobody needs anymore.
- Account security: a unique password plus an authenticator app. Most account takeovers are boring credential stuffing, not clever hacking.
If you’re still finding your feet with the tool, a beginner’s guide to what to try first beats clicking around menus with no plan.
Work accounts are a different product
Team, Enterprise and Edu plans are contractually distinct from the free and Plus tiers. Business data isn’t used for training by default, admins get controls and audit logs, and retention terms sit in a data processing agreement. If your employer gives you a workspace account, use it for work. Pasting the same material into a personal Plus account is where most real-world leaks originate.
Then there’s the plugin problem. Third-party apps connected to ChatGPT, custom GPTs built by strangers, connectors pulling from your Drive: each one is another company handling your data under its own policy. Installing a “summarise my inbox” GPT means trusting whoever built it, and many are built by hobbyists with no privacy page at all.
Getting useful answers without oversharing
You rarely need the real data. Swap names for roles, exact figures for ranges, companies for “a mid-sized retailer.” A model doesn’t need your client’s actual name to help you rewrite a difficult email. The eight-step chatbot how-to covers prompt structure properly, and one of its better lessons is that specificity comes from context, not identifiers.
Keep a simple test: if you’d hesitate to email it to a colleague, run it through Temporary Chat, or strip it out first. Prompt injection is the other live risk. Any text you paste can contain hidden instructions aimed at the model, which is why “summarise this webpage” or “read this email and draft a reply” occasionally produces something strange.
Where the rules are heading
Regulators have been circling. Italy briefly blocked ChatGPT in 2023 over GDPR concerns, the European Data Protection Board assembled a task force, and several national authorities have opened inquiries into how training data and user prompts are handled. OpenAI now shows consent prompts for training in some jurisdictions. Defaults vary by country and shift without much fanfare, so check your own Data controls rather than trusting a guide, including this one.
When privacy is the deciding factor
Some people need something structurally different: on-device processing, self-hosted models, or vendors with a stricter line on retention. ChatGPT alternatives picked by what you actually need maps out the realistic options. The honest trade-off is that privacy-first tools usually give up some capability or polish. Deciding which one you genuinely need is the entire decision.
The habit that beats any toggle
Settings are a one-time job. What actually keeps you out of trouble is a two-second pause before you paste. Ask what’s in the text, who else it describes, and whether the answer would change if the names were different. Nine times out of ten you can rewrite the prompt in ten seconds and get exactly the same quality back. The tenth time, open Temporary Chat and close it afterwards.
Do that consistently and the rest is housekeeping: training switched off, memory pruned twice a year, old threads cleared, shared links revoked. The tool is only as private as the last thing you typed into it.