Short answer: Never paste text that isn't yours to share into a consumer AI chat. That means client data, passwords, and other people's files or transcripts. Then check one training setting in ChatGPT, Claude and Gemini. Turning training off stops future training, but it doesn't erase what the vendor stores.
Never paste what isn't yours to share
The risk is rarely the AI itself. It is you moving someone else's confidential data onto a server they never agreed to. A vendor's privacy policy doesn't override your NDA or your client contract.
Keep these out of any consumer chat box:
- Client or customer personal data. Names, emails, order histories, health details. Anything that identifies a real person.
- Passwords, API keys and tokens. Pasted secrets can end up in transcripts, shared links or agents. Pasted text can even become an instruction, as our guide to prompt injection explains.
- Unreleased financials or embargoed news. If it isn't public, it doesn't belong in someone else's cloud.
- Legal, medical or HR material that isn't yours. Professional confidentiality rules apply, and no toggle satisfies them.
- Whole documents when you need one clause. Redact first, then paste only the part you need.
Your own drafts, notes and secret-free code are fine.

Yours: paste it. Names clients: redact. Whole client file: business plan. Someone else's transcript: ask first. Passwords, keys or tokens: never paste.
File connectors widen the reach. A connected Drive, Slack or Notion can expose more than the one file you named. A hidden instruction in a document can then act on it, as our guide to the real risks of AI agents explains.
Consumer vs business plans: business chats don't train by default
Business tiers from OpenAI, Anthropic and Google don't train on your content by default. Consumer apps are where the training settings matter.
Claude (Anthropic)
- Anthropic changed its consumer terms on August 28, 2025, for Free, Pro and Max plans.
- Existing users had until October 8, 2025, to choose whether chats train its models.
- If you allow training, Anthropic keeps data for five years. If not, the existing 30-day retention applies.
- The control sits in Privacy Settings. Anthropic's pages call it the model training setting, or Model Improvement. Check your own label.
- Commercial products such as Claude for Work and the API aren't used for training by default.
Gemini (Google)
- With Keep Activity on, Google uses your chats to improve its services, including training generative AI models.
- With it off, future chats aren't used for training unless you send feedback.
- Google's Gemini Apps privacy page asks you not to enter confidential information you wouldn't want a reviewer to see.
- On a work account, Google's Workspace privacy commitments say content isn't used for training outside your domain without permission.
ChatGPT (OpenAI)
- Press coverage quoting OpenAI's help pages says ChatGPT Team was renamed ChatGPT Business on August 29, 2025. The same coverage says business data is excluded from training by default.
- The consumer default and the toggle's exact label aren't confirmed here. Open your settings and look.
| Consumer chats used for training? | Off switch | Business plan | |
|---|---|---|---|
| ChatGPT | Not confirmed; check your settings | In settings; use Temporary Chat meanwhile | Excluded by default, per coverage |
| Claude | Only if you allow it | Privacy Settings, model training | Not used by default |
| Gemini | Yes, while Keep Activity is on | Turn Keep Activity off | Not used for training outside your company's domain |
How opting out works: it stops training, not storage
Opting out keeps future chats out of training, but the vendor still stores them for a while.

- Gemini: temporary chats and chats with Keep Activity off stay with your account for 72 hours. Reviewed chats are kept for up to three years. With Keep Activity on, auto-delete defaults to 18 months.
- Claude: with training off, Anthropic's August 2025 post describes 30-day retention. With it on, five years.
- ChatGPT: coverage of OpenAI's help pages says Temporary Chat may be kept for up to 30 days for safety. It isn't used to improve models.
Feedback is the carve-out people miss. Rate a Claude reply and Anthropic may store the whole conversation for up to five years, no longer tied to your account. On Gemini, feedback can still be used for training. So don't rate the reply that holds client data.
If your rule is that the text exists nowhere you don't control, no consumer chat meets it. Use a private, local AI setup or a redacted prompt. Memory is a separate setting too, covered in our ChatGPT memory guide.
Meeting bots record people who never agreed
A notetaker records everyone on the call, but only the person who invited it saw a settings page. Courts are now testing this:
- Otter.ai: UC Today reports that on August 13, 2026, a federal judge let wiretap and privacy claims proceed. Some other claims were dismissed.
- Granola: PPC Land reports a proposed class action filed July 30, 2026, in Northern California. These are allegations only.
- The Granola complaint says the tool's optional notices must be turned on separately.
Nothing is decided, and this isn't legal advice. The practical rules:
- Announce the notetaker in the invite and out loud at the start.
- Ask before recording client calls, and take no for an answer.
- Don't paste other people's transcripts into a consumer chat.
- Check the notetaker's own training setting. The people whose voices it records usually have no account, so no setting to find.
For calls that should never leave your machine, see our guide to private meeting transcription.
Try it today: one contract sentence, three settings
If the text isn't clearly yours to share, redact it, move it to a business plan, or ask the client first.
Put one sentence in your client contracts:
Where AI tools are used on client materials, the contractor will use plans that don't train on customer content by default, won't enter client personal data into consumer AI services, and will get consent before recording any call.
Then spend five minutes on settings:
- ChatGPT: open settings and find the model-training control. For a throwaway chat, use Temporary Chat.
- Claude: open Privacy Settings and set the model training setting to match your intent. Anthropic's consumer terms update explains the choice.
- Gemini: decide whether Keep Activity stays on. If it does, set auto-delete. It defaults to 18 months, with shorter and longer options.
- Your notetaker: turn on every disclosure option it has, and find its training setting.
Choosing one assistant? Our Claude vs ChatGPT comparison covers more than privacy. For EU clients, Goodwin reports EU AI Act transparency duties took effect August 2, 2026. Your duty of confidence to clients never depended on a toggle.




