Before Uploading Personal Files to ChatGPT, Check These Privacy Settings

On Free, Plus, and Pro accounts, ChatGPT can use your conversations (including the files you attach) for model training, because the “Improve the model for everyone” setting ships switched on. Before you upload anything personal, open Settings → Data Controls and turn that toggle off, then check five more things: your saved memories, your connected apps, the thumbs-up feedback habit, the new locked-chats option, and Marketing Privacy. And no setting makes it safe to upload Social Security numbers, medical records, passwords, or client-confidential files. Here’s every control, drawn from OpenAI’s current documentation.

Locked chats, a Marketing Privacy setting, and a September privacy policy refresh are new this year. Older guides miss all three.

What ChatGPT does with your files by default

If you do nothing, your uploads are treated as ordinary conversation content. On personal workspaces with ChatGPT Free, Plus, or Pro, OpenAI’s data controls page states that data sharing is enabled by default, so chats and attached files can feed future model training. Turning the toggle off doesn’t delete or hide your saved chats: it only stops new conversations from being used. Archiving changes nothing either.

Data controls is the settings hub governing training, history, export, and deletion, per account, not per chat, so one flip covers everything afterward.

There’s a loophole worth flagging before anything else. OpenAI’s help page says that if you choose to provide feedback, the entire conversation associated with that feedback may be used to train OpenAI’s models, even if you’ve already opted out on Free, Plus, or Pro. That includes the file you uploaded in that chat. On Business, Enterprise, and Edu workspaces, feedback doesn’t permit training, and admins can switch it off entirely.

Even a Temporary Chat isn’t ephemeral: OpenAI’s retention page says the company may keep a copy for up to 30 days “for safety purposes”: excluded from training, but not instantly destroyed.

One reminder from the same docs: a settings toggle is a policy promise, not a technical wall. OpenAI’s help pages themselves note that locking a chat isn’t encryption.

The privacy checklist to run before you upload anything

Run these in order. Steps 1 and 2 take under a minute and do most of the work.

1. Turn off “Improve the model for everyone.” Settings → Data Controls → toggle off. This stops new chats and their attachments from being used for training on Free, Plus, and Pro. OpenAI’s page names those three plans as the default-on personal workspaces; Go isn’t named in that sentence, though third-party guides report Go behaves the same way, so check your own toggle either way.

2. Submit the formal opt-out at the Privacy Portal. Sign in at privacy.openai.com, choose “Make a Privacy Request,” and select “Do not train on my content.” OpenAI’s help center lists this as an equal opt-out path, independent of the in-app toggle.

3. Review what ChatGPT remembers about you. Settings → Personalization → Memory. Delete individual memories, or use “Delete and turn off memory” to wipe the remembered summary and switch the feature off. Two catches: deleting a chat doesn’t delete a memory created from it, and re-enabling memory later may create new ones from chats still in your history.

4. Skip thumbs feedback on sensitive chats. A thumbs-up can pull the whole conversation (file included) into training even after you’ve opted out. Don’t rate personal chats.

5. Audit your connected apps. Settings → Connected apps. Disconnect anything you don’t actively use, and check the OAuth scope an app requests before connecting it (the Gmail one is broader than you’d expect: more below). Disconnecting stops future access but doesn’t delete existing chats or memories.

6. Use a Temporary chat for one-off sensitive questions. Start a new chat and select “Temporary” before typing. It stays out of your history, creates no memories, and trains nothing. Don’t save it afterward: saving converts it into a regular chat.

7. Consider locking sensitive chats. OpenAI rolled out locked chats in October 2026: open the ••• menu on a chat, choose “Lock chat,” and set a 4-digit PIN (Face ID or Touch ID on supported phones). Locked chats are hidden from your sidebar, search, and previews, but locking follows your existing training setting. It protects against shoulder-surfing, not against OpenAI.

8. Check Marketing Privacy. Ads arrived in January 2026, and Settings → Data Controls → Marketing Privacy lets you turn off ad personalization and clear the data already used for it. Several outlets reported marketing cookies defaulted on for free users, so this one is worth an explicit look.

Free accounts should also know the upload caps: 3 file uploads per day, 80 files per 3 hours across plans, files up to 512 MB, and text documents capped at 2 million tokens, per OpenAI’s file upload page (November 2026). CSVs sit at roughly 50 MB and images at 20 MB.

What you should never upload to ChatGPT

Settings reduce risk; they don’t make consumer ChatGPT a vault. This table is a judgment call.

Never upload	Why	Do this instead
Social Security numbers, tax IDs, passport scans	Identity theft fuel; you can’t un-train a model or un-leak a breach	Redact first, or keep the task local
Medical records, lab results	HIPAA doesn’t cover consumer AI apps: experts flagged this when ChatGPT Health launched in 2026	Use your provider’s portal; ChatGPT Health’s controls if you must, knowing the gap
Passwords, API keys, 2FA recovery codes	Any exposure is an immediate security incident	Rotate immediately if one slipped in; use a password manager
Client-confidential or NDA-covered files	You may be breaching the contract, not just risking privacy	Ask clients, or use your firm’s approved workspace
Signed contracts under legal privilege	Potential privilege waiver	Consult your attorney: this crosses into legal-advice territory
Children’s school or medical data	Kids can’t consent, and the data links back to them	Keep it out entirely
Financial account statements	Account numbers plus personal details enable fraud	Use your bank’s tools; strip identifying details first

The pattern: if exposure would cost you money, health privacy, or a professional relationship, consumer ChatGPT isn’t the right tool for it.

How each ChatGPT plan handles your data

The plan you’re on matters more than any single setting. Here’s the comparison, drawn from OpenAI’s data controls, retention, and pricing pages (November 2026):

Plan	Training on your data	Retention	Key privacy controls
Free	On by default (opt-out)	Chats until deleted; deletion completes within ~30 days	Training toggle, Privacy Portal request, Temporary chat, self-service export & deletion
Go	Not named as default-on; check your toggle	Same as Free	Same as Free; export included
Plus	On by default (opt-out)	Same as Free	Same as Free; paid model access
Pro	On by default (opt-out)	Same as Free	Same as Free
Business	Off by default: never trained on	Per workspace/admin policy; deletion within ~30 days	Admin-owned settings, feedback can’t trigger training, locked chats; self-service export replaced by owner-managed export; starts at 2 users
Enterprise / Edu	Off by default: never trained on	Admin-configured retention; some transient files expire in 48 hours	Business controls plus data residency options and admin tooling; no self-service account deletion

Two notes on that table. First, the “~30 days” figure is a deletion deadline, not storage time. OpenAI schedules permanent removal within 30 days of deletion, and legal obligations can extend it. Second, OpenAI’s pricing page confirms Business starts at two users, billed per user per month: it does not show dollar figures.

Deleting chats, exporting data, and the truth about Temporary chat

Deleting a chat is a scheduled event, not a switch. It vanishes from your view immediately and OpenAI schedules permanent deletion from its systems within 30 days. You can’t restore it through ChatGPT or Support, and legal obligations can hold data longer. From September 26, 2025, OpenAI said it was no longer under the court order from the New York Times litigation that had forced indefinite retention of some consumer content: standard deletion windows are back.

Files live separate lives from chats. If you’ve saved a file to your Library or a Space, deleting the chat that used it doesn’t remove the file: delete it there yourself. Files uploaded as a custom GPT’s knowledge persist until you delete the GPT, and project files last until deleted, with the same 30-day removal window afterward.

Export gives you everything, on most plans. Self-service export works on Free, Go, Plus, and Pro via Settings or the Privacy Portal; on Business, Enterprise, and Healthcare workspaces you go through the workspace owner. Self-service account deletion is likewise limited to personal plans, and it’s permanent.

Temporary chat is the closest thing to incognito. It stays out of your history, doesn’t create or update memories, and isn’t used for training while it remains temporary: you can start it unpersonalized so even existing memories stay out. The 30-day safety-retention window is the honest caveat, and saving the chat ends the protection entirely.

Connected apps: the Gmail connector wants more than read access

Don’t treat connectors as harmless. OpenAI’s Google app FAQ says the Gmail connector requests the gmail.modify OAuth scope: the one that covers reading mail, marking as read or unread, archiving, applying labels, and drafting. Google’s own scope documentation puts it well above gmail.readonly. Connect your inbox and ChatGPT can act on it, not just look at it.

The Calendar connector asks for calendar.events plus a Meet-related read scope, and Drive connectors request several read-only scopes, with a full Drive scope reserved for actions that create, move, or delete files. On data use, OpenAI says it does not train its generalized models on data pulled directly from connected Google apps, even when “Improve the model for everyone” is on. That’s a genuine carve-out.

Connect the minimum and disconnect the rest: disconnection stops future access but doesn’t delete existing chats or memories built from the connection.

How ChatGPT compares with Claude and Gemini on privacy

If your threat model is “I don’t want my conversations training anything, ever, by default,” the consumer default is the deciding factor. Anthropic’s Claude is opt-in: consumer users are asked before their data is used for training, so the default is off. ChatGPT and Google Gemini both use an opt-out model: training on until you flip the switch. Multiple 2026 comparisons call Claude the strongest default for privacy without touching settings.

On business tiers, the field converges: OpenAI, Anthropic, and Google all exclude workspace and API data from training by default. So the real decision for a small business isn’t “which AI respects me”: it’s whether your team is on a business workspace before anyone pastes a client file into a personal chat. OpenAI’s enterprise privacy page states its no-training default, encryption at rest and in transit, and data-residency options for Enterprise.

What changed in OpenAI’s privacy setup in 2026

Four things moved this year that most walkthroughs still don’t mention:

• The US Privacy Policy was updated September 10, 2026, tying retention to data type and legal obligations. OpenAI also runs a separate privacy policy for the EEA, UK, and Switzerland, served through OpenAI Ireland Limited. Some EU-focused guides report the training toggle defaults to off there, but OpenAI’s help pages don’t state a regional default: check your own settings rather than assume.
• Locked chats launched in October 2026 (PIN or biometric protection, hidden from history and search) on every plan except Healthcare. Resetting a forgotten PIN permanently deletes all locked chats: that’s the recovery path, not an unlock.
• Ads and Marketing Privacy arrived in January 2026, with OpenAI saying it will always offer a way to not see ads. Critics flagged the cookie defaults; the opt-out lives in Data Controls.
• Model governance documents moved: an updated Model Spec published October 29, 2026, and a Usage Policies update effective November 20, 2026. Neither changes your settings, but both shape how OpenAI says it reviews and enforces behavior.

One 2026 launch deserves its own warning: ChatGPT Health, announced January 7, 2026 and US-wide from July, lets people upload medical records with health-specific controls. Security publications stress that HIPAA still doesn’t cover consumer uploads, so treat even that feature as convenience over confidentiality.

What to do

If you use ChatGPT casually: run steps 1-5 of the checklist once and you’ve covered the realistic risks. If you handle client work: get on a Business workspace (two-seat minimum) instead of relying on personal toggles, and keep client files inside it only after checking with whoever owns the contract. For something genuinely sensitive: use a Temporary chat with a redacted version of the document, or a locally-run tool, which is overkill for recipes and exactly right for tax returns.

FAQ

Does ChatGPT train on my uploaded files?

On Free, Plus, and Pro personal accounts, yes by default. Files you attach become part of the conversation, which can train OpenAI’s models unless you’ve turned off “Improve the model for everyone.” Temporary chats are excluded, and Business, Enterprise, and Edu workspaces never train on your content by default.

How do I stop ChatGPT from training on my data?

Open Settings → Data Controls and turn off “Improve the model for everyone.” For a formal opt-out, sign in at privacy.openai.com and submit the “Do not train on my content” request. Both affect new conversations; neither removes anything you already shared. Avoid thumbs feedback on chats you want kept out of training.

Does deleting a ChatGPT chat delete it from OpenAI’s servers?

Not immediately. The chat disappears from your view at once and is scheduled for permanent deletion from OpenAI’s systems within 30 days, and deletion is irreversible through the app or support. Security and legal obligations can extend that window. Files saved in your Library or Spaces stay even after the chat that used them is deleted.

Are Temporary chats in ChatGPT actually private?

Mostly, for a consumer tool. Temporary chats don’t appear in your history, don’t create or update memories, and aren’t used to train OpenAI’s models. The catch: OpenAI may keep a copy for up to 30 days for safety purposes, and saving a temporary chat converts it into a regular chat under your normal settings.

Is ChatGPT safe for medical records or tax documents?

Don’t put them in consumer ChatGPT. HIPAA does not cover what consumer AI apps do with health information you upload yourself, and tax documents carry Social Security numbers and financial data no setting can un-share. ChatGPT Health (US, 2026) adds health-specific controls, but researchers still flag the gaps. Compliant enterprise setups or local tools are safer.

Which ChatGPT plan is the most private?

ChatGPT Enterprise and Edu, followed by Business. All three exclude workspace content from model training by default, put admins in control of retention and feedback, and support workspace retention policies. On personal plans (Free, Go, Plus, Pro) you rely on your own toggles. Enterprise adds data residency options and admin tooling Business lacks.

Do thumbs up or thumbs down affect whether my chats are used for training?

Yes, and this surprises people. OpenAI’s documentation states that if you choose to provide feedback, the entire conversation associated with that feedback may be used to train OpenAI’s models, even if you’ve opted out of training on Free, Plus, or Pro. On Business, Enterprise, and Edu, feedback does not permit training, and admins can disable it.

What should I do if I already uploaded a sensitive file to ChatGPT?

Delete the chat containing it, then check whether the file was saved to your Library or Spaces and remove it there too: deleting the chat alone doesn’t delete saved files. Review saved memories for anything picked up from the conversation. You can’t recall data already used in training, so treat exposed credentials as compromised and rotate them.

Sources

• Data controls in ChatGPT: OpenAI Help Center
• How your data is used to improve model performance: OpenAI Help Center
• Chat and file retention in ChatGPT: OpenAI Help Center
• Uploading files and audio to ChatGPT: OpenAI Help Center
• Temporary chat in ChatGPT: OpenAI Help Center
• Memory in ChatGPT: OpenAI Help Center
• Using locked chats in ChatGPT. OpenAI Help Center
• Google app data controls FAQ. OpenAI Help Center
• Connected apps in ChatGPT: OpenAI Help Center
• ChatGPT pricing: OpenAI
• How we’re responding to The New York Times’ data demands: OpenAI
• Introducing ChatGPT Health. OpenAI
• US Privacy Policy: OpenAI
• EU Privacy Policy: OpenAI
• Enterprise privacy at OpenAI
• An updated Model Spec: OpenAI
• ChatGPT Plus vs Business vs Enterprise data security (2026): dope.security
• ChatGPT Health: Top Privacy, Security, Governance Concerns. BankInfoSecurity