Quick answer: Don’t paste secrets into consumer chat. For everyday work, turn off model-improvement settings on personal accounts, use temporary/incognito chats for sensitive one-offs, and keep real confidential work in tools your company actually contracts for. Settings reduce training risk — they don’t make a free chatbot a vault.
Skip this if you only ask public trivia. Keep reading if you’ve ever pasted a customer list, contract, or medical note “just this once.”
The 60-second hygiene checklist
- Never paste: passwords, API keys, full IDs/SSNs, payroll, unpublished strategy, patient/student records, anything under NDA.
- ChatGPT (personal): Settings → Data Controls → turn off Improve the model for everyone. Use Temporary chat for one-offs you don’t want in history.
- Claude (consumer): Settings → Privacy → turn off model-improvement / “Help improve…” Use Incognito chat (ghost icon) for temporary conversations.
- Gemini Apps: Gemini Apps Activity → control Keep Activity / what can be used to improve Google AI. Prefer temporary chats for sensitive one-offs.
- Still: assume a human reviewer could see some content under safety processes — vendors say this explicitly in places. If it would be a disaster on a projector, don’t paste it.
What the switches actually do
| App | Everyday setting | One-off mode | Honest limit |
|---|---|---|---|
| ChatGPT | Off: “Improve the model for everyone” (future chats aren’t used to train; history can still exist) | Temporary chat — stays out of history and isn’t used to improve models while temporary | Saving a temporary chat converts it to a regular chat under your account settings |
| Claude | Privacy → turn off model improvement for future chats/coding sessions | Incognito chat — not saved to history/memory; not used for training | Retained ~30 days for safety; can’t convert Incognito into a saved chat |
| Gemini | Gemini Apps Activity / Keep Activity controls storage and improvement for future chats | Temporary chats — not used to train Google’s AI models | Even with Keep Activity off, Google still processes chats to respond and for safety/abuse defenses |
Labels move in product UIs. If you can’t find a toggle, search that vendor’s help center for the phrases above — don’t guess from a YouTube thumbnail.
A safer prompt pattern (when you must use AI)
Strip identifiers. Replace names with roles. Keep the structure, lose the secrets.
Outcome: Clearer email a manager can act on.
Audience: Busy manager.
Constraints: Under 120 words. No real names — use Role A / Company X.
Context:
[paste redacted draft]Same skeleton as my prompting guide — just redacted.
Work vs personal accounts
Business / Enterprise / Edu workspaces often default to not training on your workspace content (OpenAI documents this for ChatGPT Business/Enterprise/Edu). That is not a free pass for regulated data — follow your employer’s policy. Personal free/pro chats are a different product with different knobs.
Common “it’s fine” mistakes
- “I’ll delete the chat after.” Deletion isn’t the same as “never used for training,” and retention windows exist for safety.
- “I turned the toggle off once last year.” Check again after app redesigns — menus move.
- “It’s only an email draft.” Drafts contain names, deals, and medical/HR context more often than people admit.
- “Temporary mode means anonymous.” It means different history/training rules on a company server — not invisibility.
FAQ
If I turn training off, is my data deleted from old models?
No. Anthropic is explicit: turning the setting off stops future use; data already in started runs / trained models isn’t yanked out. Treat the past as spilled milk.
Is Temporary / Incognito “anonymous”?
No. It’s a product mode with different history and training rules — plus short safety retention. You’re still talking to a company’s servers.
What’s the simplest rule if I remember only one thing?
If you wouldn’t paste it into a random web form, don’t paste it into a consumer chatbot.
What I’d do today
- Flip the model-improvement toggles on every personal account you use.
- Practice one Temporary/Incognito chat on a non-sensitive task so the muscle memory exists.
- Move anything confidential to the channel your workplace already approved.
Sources
- OpenAI — Temporary chat in ChatGPT
- OpenAI — Privacy Center in ChatGPT (model improvement controls)
- Anthropic — Incognito chats
- Anthropic — Model improvement privacy settings
- Google — Gemini Apps Privacy Hub
Updated Oct 1, 2026. Settings labels move — confirm in-product.
Read next (Privacy → Hub) → Start here — pick your next job
Want the weekly note? 1 SI/AI change + 1 workflow
Leave a Reply