Disclosure up front: I make Outbound Veil, a Mac menu-bar app that checks what you type for personal information, and it appears in step 4. The comparison of tools, with several free ones, is in a separate guide. I read the provider pages below on October 11, 2026. They change, so check the pages for your plan. This is not legal advice.
The five checks
- Know which account you are using. A personal ChatGPT or Claude account and a workspace your employer runs follow different rules.
- Check the training setting for that account. It is a toggle, and turning it off does less than people expect.
- Remove identifiers before you send, not after. Replace them with placeholders and keep the key somewhere else.
- Have something besides your own eyes check the text. Names and numbers hide in free-text notes.
- If the data must never leave your machine, do not send it. A local model, or no AI at all, is the answer.
1 and 2: what the providers say about a personal account
ChatGPT. The setting is called Improve the model for everyone, under Settings, then Data controls. OpenAI says that when it is off, your new conversations are not used to train its models, and that turning it off does not delete or hide saved chats. A temporary chat does not appear in your history, does not create or update memories and is not used to improve models, and OpenAI says it may be retained for up to 30 days for safety purposes. If you give a thumbs up or down, OpenAI says the whole conversation may be used to train its models, even if you opted out. For Business, Enterprise, Edu and ChatGPT for Healthcare workspaces, OpenAI says it does not use content to train its models by default. Source: Data controls in ChatGPT.
Claude. The setting is Help Improve our AI models, under Settings, then Privacy. Anthropic says that when it is off, new chats and coding sessions are not used for future model training, and that data already used for training is not removed; flagged conversations may still be used to improve its internal trust and safety models. When it is on, Anthropic says it may keep de-identified chats for up to 5 years in its training pipelines, for new or resumed chats. It says Incognito chats are not used to improve Claude even when the setting is on. I could not find the default position of the toggle on those pages, so open it and look. Sources: how to change the setting (updated August 3, 2026) and how long data is stored (July 1, 2026).
One point holds for both: not used for training is not the same as not kept. Each keeps conversations for a while, and a training setting only changes what happens to new ones.
3: remove identifiers first
Check the text for the obvious ones: names, street addresses, phone numbers, email addresses, government, tax, bank and card numbers, dates of birth and record numbers. HIPAA's de-identification rule lists 18 kinds of identifiers in 45 CFR 164.514(b), and the list is a good prompt even if you are not covered by HIPAA. Then look for the quiet ones: a rare job in a small town, a named employer, a third party mentioned in passing.
Replace each identifier with a consistent placeholder (Client A, Client B, early March) so the note still reads sensibly, and keep the list of what each placeholder stands for in a file that never goes near the chatbot. Do this before you paste. Editing a chat afterwards does not take the original back.
4: a check as you type
A tired person misses things in their own notes. A checker that runs where you type catches some of them before you press send. Outbound Veil runs on your Mac, with the checking done locally and nothing you type stored or sent. It marks names, addresses, phone numbers, email addresses, Social Security, card, passport, driver's license, tax and bank numbers, dates of birth, medical record numbers and IBANs, and a Redact all command replaces them with placeholders.
What I have tested is narrow. On October 3, 2026, on one Mac running macOS 26 with builds made before release 1.0.0, it found and redacted a made-up note in the ChatGPT prompt box and in the claude.ai prompt box in Chrome. I have not tested Safari or other browsers, and I have not run those tests again on later releases. The compatibility page records every app and field, including the ones that do not work. It does not read the body of a new message in Apple Mail. It handles Latin-script names only, it also flags public figures' names, and it will miss some things. It is $29.99 once after a 7-day trial; the free options are in the comparison guide.
5: when the data must not leave your machine
If a policy, a consent form or common sense says the text cannot go to a third party, no setting fixes that. A model that runs on your own Mac keeps the text there, and so does doing the task by hand. The comparison guide covers that option next to the redaction tools.
If you already pasted something you should not have
Delete the chat. OpenAI says it is removed from your view at once and scheduled for permanent deletion within 30 days, with exceptions for data already de-identified or held for security or legal reasons (Deleting and archiving chats). Anthropic says deleted conversations leave your history immediately and its back-end storage within 30 days. Neither undoes training that has already happened. Then tell whoever sets the rules for that data, such as your privacy officer, your IRB or your employer, and follow their process. The deadline for that conversation is theirs, not the provider's.
Frequently asked
If I turn off training, does ChatGPT stop keeping my chat?
No. OpenAI's page says turning off Improve the model for everyone does not delete or hide saved chats; they stay in your history. A temporary chat is the option that stays out of history, and OpenAI says it may be retained for up to 30 days for safety.
Is deleting a chat enough?
It removes it from your view at once. OpenAI says it then schedules the chat for permanent deletion within 30 days, with exceptions for data already de-identified or kept for security or legal reasons. Anthropic says deleted chats are removed from your history immediately and from its back-end storage within 30 days. Deleting does not undo training that has already happened.
Does a redaction tool make this compliant?
No. A checker lowers the chance of a mistake. It does not make a workflow compliant with HIPAA, FERPA, your IRB protocol or your employer's policy, and every tool misses some things. Ask whoever sets the rules where you work.
Which is safer, ChatGPT or Claude?
I did not compare them on safety, and I cannot tell you. Both describe a training control and a 30-day deletion window for personal accounts, and OpenAI says it treats business workspaces differently. I did not verify how Anthropic treats its business plans. Read the pages linked above for the plan you actually use.
Related guides
- Best tools to redact personal information before you use ChatGPT and other AI chats (2026): the tools compared.
- Where Outbound Veil works: every app and field tested, and the ones that do not work.
- Mac privacy permissions explained: what Accessibility and Input Monitoring allow.
- Outbound Veil, all products and all guides.