The three rules
- Assume anything you type could be seen by someone other than the AI. Even on enterprise plans where data isn't used for training, support staff, contractors, or law enforcement can sometimes access logs. Type accordingly.
- Never paste secrets. Passwords, API keys, social security numbers, credit cards, bank account info, full health records. Not even "just to check something."
- Verify everything that matters. AI sometimes makes things up. For legal, medical, financial, or anything client-facing, double-check.
What's safe to share
- Public information
- Your own ideas, drafts, and notes
- Generic templates and outlines
- Marketing copy, social posts, blog drafts
- Code that doesn't contain secrets or credentials
- Most "how do I" questions about software, business, life
What to be careful with
- Client names and details. Use placeholders ("Client A") if you can. Always check your plan's data handling rules before sharing real client info.
- Internal financials. P&Ls, salary info, investor decks. Fine on Enterprise plans with the right contracts in place. Risky on personal accounts.
- Anything HR or legal-privileged. Use enterprise plans with the right data handling agreement, or don't share at all.
- Health information. HIPAA covers this. You need a Business Associate Agreement (BAA) with the AI provider. Personal plans don't have one.
- Customer PII. Names, addresses, emails. Anonymize when possible. Use enterprise plans for actual customer data.
What never to share
- Passwords, security questions, MFA codes
- API keys, secret tokens, private keys
- Social security numbers, government IDs, passport numbers
- Full credit card or bank account numbers
- Anything covered by attorney-client privilege without the right enterprise setup
- Anything you wouldn't want screenshotted and posted publicly
The newspaper test
Before pasting something into AI, ask: "If this exact paste showed up in a newspaper article tomorrow, would I be fine?" If no, don't paste it without first checking your plan's data terms.
Settings to flip right now
ChatGPT
- Settings → Data Controls → "Improve the model for everyone." Turn this OFF if you don't want your chats used for training.
- Settings → Data Controls → "Chat history." If you turn this off, your chats won't be saved long term. Useful if you handle sensitive stuff.
- Settings → Personalization → "Memory." Review what ChatGPT remembers about you. Delete anything sensitive.
Claude
- By default, Anthropic does not use your chats for training. You don't need to flip anything for that.
- Settings → Privacy → "Help improve Claude." If this is on, turn it off to be extra cautious.
- Memory. If you've enabled memory, review what Claude remembers periodically and delete sensitive entries.
Gemini
- myactivity.google.com → Gemini Apps Activity. Turn it off if you don't want your chats saved to your Google account.
- Note: Google's privacy policy is more permissive than Claude or ChatGPT. Be extra careful with what you share.
Plan tiers and what they actually mean for privacy
| Plan | Used for training? | Data handling agreement? | Audit logs? |
|---|---|---|---|
| ChatGPT Free | Yes by default (can opt out) | No | No |
| ChatGPT Plus / Pro | No | No formal DPA | Limited |
| ChatGPT Team | No | Standard DPA | Yes |
| ChatGPT Enterprise | No | Full DPA, BAA available | Yes |
| Claude Free / Pro / Max | No by default | No formal DPA | Limited |
| Claude Team | No | Standard DPA | Yes |
| Claude Enterprise | No | Full DPA, BAA available | Yes |
"DPA" means Data Processing Agreement. "BAA" means Business Associate Agreement (for HIPAA). If you handle health data, legal client data, or financial customer data, you need these in writing. The Team and Enterprise tiers provide them.
Special advice if you handle client data
- Law firms. Use Claude Enterprise or ChatGPT Enterprise. Anything less and you have a problem with attorney-client privilege if the AI provider's data handling is broader than your firm allows.
- Healthcare. You need a signed BAA. Only Enterprise tiers offer one, and you have to explicitly request it.
- Financial services. Same deal. Enterprise plus your own internal AI usage policy.
- Anyone with NDAs. Check your NDA. Many say "you may not share confidential info with third parties." AI counts as a third party. Get permission or anonymize.
The intermediate take: red flags to watch for
- Browser extensions that "supercharge ChatGPT." Many are spyware. Stick to the official apps.
- Random "free ChatGPT alternatives" you've never heard of. Most are reselling someone else's API and logging everything. Stick to Anthropic, OpenAI, Google, and major established players.
- Email or text messages "from OpenAI" or "Anthropic." These companies almost never reach out by email. If you get one, it's probably phishing.
- AI tools asking for full Google or Microsoft account access. Read what they're requesting. "Read all your email" and "send email as you" are huge permissions. Don't grant unless you trust the tool.
The advanced take: AI-specific attacks to know about
- Prompt injection. A bad actor hides instructions inside a document, webpage, or email. When AI processes it, the hidden instructions take over. Defense: don't let AI take actions based on untrusted content without your review.
- Data exfiltration. Combined with prompt injection, AI can be tricked into emailing your files to an attacker. Defense: keep AI sandboxed away from sensitive data and outbound communication.
- Memory poisoning. If memory is on, a sneaky chat can plant false "facts" that show up in later chats. Defense: review what AI remembers, periodically. Delete anything weird.
- Connector chains. AI agents that connect to multiple tools can be tricked into combining permissions in unexpected ways. See topic 15 for the full breakdown.
None of these should scare you off AI. They should just inform how you set things up. Sensible defaults plus enterprise plans for sensitive work covers 95% of the risk.