← All posts

Is ChatGPT private? What OpenAI keeps, who can read it, and what never to paste

Is ChatGPT private? Not by default. What OpenAI stores on each plan, who can read your chats, the settings to change, and what never to paste — checked October 2026.

You have a client's email thread in one window and ChatGPT in the other. The reply has to be firm, polite and done in five minutes, so you paste the whole thread in and ask for a draft. The client's name, the deal terms and the annoyed paragraph about your invoice all go into the box.

Is ChatGPT private? Not the way a file on your own laptop is. On personal plans (Free, Go, Plus, Pro), OpenAI keeps your chats until you delete them, may train on them unless you switch that off, and lets a limited group of staff and contractors read content for abuse, support, legal and training reasons. Business, Enterprise, Edu and the API don't train on your data by default — but on ChatGPT's business plans, your employer's admins can read what you type.

I build TypeTab, an on-device autocomplete for the Mac, so "keep it on your machine" is my bias. Everything below about ChatGPT privacy comes from OpenAI's own help center and policies, one federal court order, and a few named outside sources, all checked October 2026.

Is ChatGPT private on your plan? The defaults side by side

Your plan decides almost everything: whether OpenAI trains on your chats, who can open them, and who controls deletion.

Personal: Free $0 · Go $8 · Plus $20 · Pro from $100/mo Business Enterprise · Edu · Healthcare API
Trained on by default? Yes, opt-out available No No No
OpenAI access Authorized staff and service providers: abuse, support, legal, model improvement (unless you opted out) Employees for support, abuse, legal; contractors for abuse review Employees only for incidents, recovery with your permission, or legal duty Employees; contractors for abuse review
Your organization can read it? No — but if you signed up with a work email, OpenAI may tell your employer the account exists Yes: admins can view, export and delete Yes: audit log via the Compliance API Your company owns the inputs and outputs
After you delete Removed within 30 days, except chats already de-identified or held for legal or security reasons Admin sets retention; deleted within 30 days, with exceptions (when law requires longer, or to protect the service or others from harm) Admin sets retention; deleted within 30 days, with exceptions (when legally required to keep it) Logs kept up to 30 days; zero retention available on eligible endpoints
Ads Possible on Free and Go; none on Plus or Pro None None on Enterprise or Edu —

Sources: ChatGPT pricing (checked October 2026), OpenAI consumer data FAQ (2026), OpenAI enterprise privacy (2026), OpenAI privacy policy (2026), OpenAI ads help (2026).

One correction, because a top Google result for this question gets it backwards: it says ChatGPT Business admins can't read your conversations. OpenAI's own page says Business workspace admins "can view, access, export, and delete end user conversations in the workspace" (OpenAI enterprise privacy, 2026).

Decision rule: if your employer pays for the seat, assume your employer can read the chat.

Does ChatGPT store your data, and for how long?

Yes. Saved chats stay until you delete them; deleted chats and Temporary Chats are purged within 30 days — unless a chat was already de-identified for training or a legal or security obligation overrides that.

  • Regular and archived chats stay in your account until you delete them. Archiving doesn't shorten anything (OpenAI retention help, 2026).
  • Deleted chats vanish from your sidebar at once and are scheduled for permanent deletion within 30 days. Two exceptions: chats already de-identified for training, and chats OpenAI must keep for security or legal reasons. Support can't restore them.
  • Temporary Chats never appear in history, but OpenAI can hold a copy for as long as 30 days for safety checks (OpenAI Temporary Chat help, 2026).
  • Saved memories are stored apart from chats. Deleting a chat doesn't necessarily delete a memory made from it, and logs of deleted memories can stay for up to 30 days (OpenAI memory help, 2026).
  • Files saved to Library survive the deletion of the chat that used them (OpenAI retention help, 2026).
  • Text sent through a custom GPT's action to an outside service follows that service's policy, which may keep it longer than 30 days (OpenAI Temporary Chat help, 2026).

The "legal obligation" exception isn't theoretical. For several months in 2025, a court order in The New York Times v. OpenAI forced OpenAI to keep consumer ChatGPT and API content indefinitely, deleted chats included. That obligation ended on September 26, 2025, though OpenAI says it still holds a limited set of April–September 2025 data under legal hold (OpenAI, 2025).

Then on January 5, 2026, Judge Sidney Stein affirmed an order requiring OpenAI to hand the plaintiffs a sample of 20 million de-identified ChatGPT logs (SDNY order, 2026). OpenAI says the sample was drawn at random from consumer conversations between December 2022 and November 2024, and that Business, Enterprise, Edu and API customers weren't in it (OpenAI, 2025).

The court's reasoning matters more than the number. The judge found the privacy interest in ChatGPT chats weaker than in wiretapped phone calls, because these are conversations "which users voluntarily disclosed to OpenAI and which OpenAI retains in the normal course of its business."

Practical rule: treat anything typed into a consumer chat as a record another company holds — and one a court can reach.

Does ChatGPT train on your data, and can you opt out?

On personal plans, yes by default; one switch turns it off for new chats. Content from Business, Enterprise, Edu, ChatGPT for Healthcare and the API isn't used for training unless the organization opts in (OpenAI consumer data FAQ, 2026).

  • It applies to new conversations. Old chats stay in your history; turning training off doesn't delete them (OpenAI data controls FAQ, 2026).
  • The thumbs trap: rate a response with thumbs up or down and "the entire conversation associated with that feedback may be used to train" OpenAI's models, opt-out or not (OpenAI model-improvement help, 2026).
  • Codex has a separate "Include environments" setting for training on your code environments. The main ChatGPT switch doesn't change it (OpenAI data controls FAQ, 2026).
  • OpenAI says it takes "steps to reduce the amount of personal information in training datasets." Reduce, not remove.

If you pasted something sensitive before opting out, the switch won't pull it back. Deleting the chat — or the whole account — is the lever you have left, and a chat already de-identified for training is outside even that.

Who can actually see your ChatGPT conversations?

More people than the chat window suggests: OpenAI staff and contractors under defined conditions, your employer on business plans, anyone holding a shared link, and courts or governments with legal process.

  • OpenAI personnel and service providers, for four reasons: investigating abuse or security incidents, support you asked for, legal matters, and model improvement unless you opted out. OpenAI says it logs every access. Its own advice is blunt: "Please do not enter sensitive information that you would not want reviewed or used" (OpenAI consumer data FAQ, 2026).
  • Your employer. On a business account, admins "may access and control your OpenAI account, including being able to access your Content." And if you signed up for a personal account with a work email, OpenAI may tell your employer the account exists (OpenAI privacy policy, 2026).
  • Anyone with a shared link. A link from a personal account opens for anyone who has it, you can't set an expiry date, and it can be forwarded (OpenAI shared links help, 2026). In July 2025 a Google site search turned up nearly 4,500 shared chats; OpenAI then removed the opt-in feature that let them be indexed (Fast Company, 2025).
  • Courts and government authorities, when the law requires it — see the New York Times case above.
  • Not advertisers. Ads started testing on Free and Go in the US on February 9, 2026. OpenAI's line: "Advertisers do not have access to your chats, chat history, memories, or personal details." With ad personalization on, though, your past chats and memories can shape which ads you see (OpenAI ads help, 2026).
  • Strangers, through a bug. In March 2023 a library bug showed some users other people's chat titles. The same bug may have made payment details — name, email, last four card digits — visible for 1.2% of Plus subscribers active in a nine-hour window, though OpenAI believes very few were actually seen (OpenAI, 2023).

Is ChatGPT safe for confidential information? Paste, redact first, or never paste

On a personal plan, only after redaction — and some data belongs in no chatbot at all. The rule outside OpenAI is consistent. The UK's National Cyber Security Centre recommends "not to include sensitive information in queries to public LLMs" (NCSC, 2023). Harvard bars confidential data — non-public research, finance, HR, student records, medical — from publicly available AI tools and allows it only in tools its security office has approved (Harvard HUIT, 2026). Samsung banned generative AI for staff in 2023 after employees uploaded sensitive code (Bloomberg, 2023).

What you're pasting Personal plan, training off Your employer's Business / Enterprise / Edu workspace
Your own draft with no names or numbers Paste Paste
A personal health, money or relationship question Your call: use an Unpersonalized Temporary Chat Don't: admins can read work chats
A client's or colleague's email thread Redact first: names, companies, amounts, dates Paste if your company's AI policy allows it
A contract clause or anything under NDA Redact first, or skip Paste only if policy and the contract allow it
Customer list, CRM export, any spreadsheet of personal data Never Only in a workspace your company approved; a DPA is available for Business and Enterprise
Proprietary source code Never Per policy; no training by default, and check the Codex environment switch
Patient health information (PHI) Never Never on Business, which has no BAA. Only in a BAA-covered product, such as ChatGPT for Healthcare or Enterprise with Regulated Workspace, and only with covered features
Passwords, API keys, account numbers, ID scans Never Never
Privileged legal advice Never Ask your counsel first

Table sources: OpenAI enterprise privacy (2026), OpenAI BAA help (2026), OpenAI HIPAA-eligible products (2026), OpenAI shared links help (2026), plus the NCSC and Harvard rules above. None of this is legal advice; for regulated data, your company's policy and counsel outrank any table.

How to redact. Swap every identifier for a bracketed label, get the draft, then swap them back:

Before: "Dana, Acme will pay $48,500 by March 3 if Bolt Logistics drops clause 7.2."

After: "[Name], [Client] will pay [Amount] by [Date] if [Vendor] drops [Clause]."

One failure mode: if the story itself identifies the client — "the only bakery chain in Lisbon that just lost a franchise lawsuit" — swapping names doesn't help. Cut the context, not just the labels.

How to make ChatGPT more private: the privacy settings, checked October 2026

Turn off training, use Unpersonalized Temporary Chats for one-offs, audit memory and delete old shared links; on Free or Go, also limit ad signals. Steps below are from OpenAI's help center as of October 2026, for a personal account; menu locations can differ by platform, and step 1 lists the web and mobile paths OpenAI documents. In a Business or Enterprise workspace, admins control several of these; ask them.

  1. Turn off training. Web: account menu → Settings → Data controls → Improve the model for everyone → off → Done. iOS or Android: sidebar → profile icon → Data controls → off. Signed out, the choice is saved per browser. Alternative: select "Do not train on my content" at privacy.openai.com (OpenAI data controls FAQ, 2026). Result: new chats aren't used for training; old chats stay in history. Undo: flip it back on.
  2. Use Temporary Chat for one-offs. Open a new chat → select Temporary → choose Unpersonalized → send. Pick before the first message; you can't change it later (OpenAI Temporary Chat help, 2026). Result: the chat stays out of history, creates no memories and isn't used for training; a copy may be kept up to 30 days. A Personalized temporary chat can still read your existing memories, which is why Unpersonalized is the private choice. Undo: nothing to undo; the next regular chat works as before.
  3. Audit memory. Settings → Personalization → Memory. Read the memory summary, delete what shouldn't be there, or pick Delete and turn off memory from the ••• menu. If you see Reference chat history, turning it off schedules what was learned from past chats for deletion within 30 days (OpenAI memory help, 2026). Result: ChatGPT stops using the memories you deleted. OpenAI notes the same detail can also live in other chats, files or custom instructions, and deletion can take time to propagate. Undo: turn Memory back on — it may build new memories from chats still in your history.
  4. Delete old shared links. Settings → Data controls → Shared links → Manage → More actions → Delete all shared links (OpenAI shared links help, 2026). Result: the links stop working. Copies that recipients already saved to their own accounts stay there. Undo: none — share the chat again to make a new link.
  5. Free or Go: limit ad signals. Settings → Ad controls → turn off Personalize ads and Past chats and memory; Delete ads data clears it within 30 days (OpenAI ads help, 2026). Result: ads still appear, picked without your past chats and memory. Temporary Chats show no ads. Undo: switch the toggles back on.
  6. Clean up a paste you regret. Delete the chat from your ChatGPT history, then delete any memory and any Library file created from it. Result: gone from your account now, and from OpenAI's systems within 30 days, unless it was already de-identified or is held for legal or security reasons. Undo: none — deleted chats can't be restored, even by Support.

Snags people hit, with fixes:

  • "My thumbs up/down buttons disappeared." A user on OpenAI's forum reproduced this in September 2026: the buttons vanish whenever training is off, and OpenAI Support replied that it shouldn't happen (OpenAI Community, 2026). Fix: leave training off. Turning it back on just to rate a reply defeats step 1.
  • "There's no Save option in my Temporary Chat." Since October 1, 2026, users report the save option missing on web and the Mac app, although the help center still documents it (OpenAI Community, 2026). Fix: copy out anything you want to keep before you leave the chat.
  • "I deleted the chat, but ChatGPT still knows." Usually a saved memory, which lives apart from the chat — though other chats, Library files or custom instructions can hold the same detail. Fix: check Settings → Personalization → Memory first, as in step 3.

End state: training off, sensitive one-offs in Unpersonalized Temporary Chats, a memory you've read, no forgotten shared links.

What to do instead of pasting

Pick by job: company data belongs in your employer's workspace, and text that can't leave your machine belongs in a tool that runs on it.

  • You need ChatGPT's reasoning on company data: use your employer's Business or Enterprise workspace. No training by default, a DPA, admin-set retention — that's what those plans are for.
  • You need a chatbot, but the text can't leave your machine: run a model locally. My guide to running local LLMs on your Mac covers which models fit how much memory. Pro: nothing is uploaded. Con: smaller models and some setup.
  • You'll paste anyway: paste the redacted version and let the prompt carry the rest — role, reader, format, length. How to write better AI prompts walks through those patterns.
  • You only need the reply written: write it yourself, faster.

That last case is bigger than it looks. Writing is the most common work use of ChatGPT — 42% of work-related messages — and about two-thirds of writing requests ask it to rework text the user supplied (OpenAI/NBER working paper, 2025). Much of that is text you could write yourself if the words came quicker.

That job is what TypeTab does. It suggests your next words as faint ghost text in almost any Mac app; Tab takes a word, Ctrl+E takes the line, and you keep typing to ignore it. Every prediction and every training run happens on your Mac, keystrokes are never uploaded, and there's no account and no telemetry. The app goes online for two things: downloading the model on first launch, and a one-time license check that sends your key, never your text. The details are in how TypeTab learns your writing style on your Mac.

What TypeTab won't do

Honest limits, since I'm the vendor:

  • It isn't a chatbot. It won't summarize a 30-page contract, answer questions or draft a reply from a prompt. It finishes sentences you started.
  • It's Mac-only: macOS 14 or later, Apple Silicon, 16 GB of memory recommended. No iPhone, Windows or Linux.

If your team needs ChatGPT-grade reasoning over confidential documents, a Business or Enterprise seat with a signed DPA is the honest answer — not a workaround.

Your ChatGPT conversations are as private as your plan and the settings you changed. The text you never pasted is the only text with no retention policy.

FAQ

Is ChatGPT safe? For security, mostly yes: OpenAI encrypts data in transit (TLS 1.2) and at rest (AES-256) (ChatGPT pricing FAQ, checked October 2026) and offers multi-factor authentication (OpenAI consumer privacy, 2026). Privacy is the weaker side. On personal plans your chats are stored until you delete them, can train models unless you opt out, and can be read by authorized OpenAI staff and contractors for abuse, support and legal reasons.

Does ChatGPT store your data? Yes. Chats stay in your account until you delete them; deleted chats and Temporary Chats are removed from OpenAI's systems within 30 days unless a chat was already de-identified for training or a legal or security obligation requires longer. Memories and Library files are stored separately and need their own deletion.

Is ChatGPT safe for confidential information? Not on a personal plan. Redact client emails and contracts before pasting, and never paste customer data, source code, passwords or patient information. Business, Enterprise and Edu don't train on your data by default, but your company's admins can read those chats, so follow its AI policy.

Is ChatGPT's Temporary Chat private? More private, not invisible. A Temporary Chat stays out of your history, creates no memories, isn't used for training and shows no ads, but OpenAI can hold a copy for as long as 30 days for safety checks. Choose Unpersonalized before the first message so it doesn't read your existing memories either.

Can my employer see my ChatGPT history? On a ChatGPT Business, Enterprise or Edu account, yes: Business admins can view, export and delete conversations, and Enterprise admins can pull an audit log through the Compliance API. On a personal account they can't read your chats through OpenAI, but if you signed up with a work email, OpenAI may tell your employer the account exists.

Try TypeTab free
Try TypeTab free $49 lifetime · no account
Download