AI Privacy & Data Security Checklist (Personal and Teams)

Updated: 2026-08-12. Privacy toggles, enterprise terms, and retention policies follow each product’s settings and legal pages on the day you check.

Why this guide

The costliest AI mistake is rarely weak prose—it is sending data that should never leave your boundary. This checklist covers data classes, never-paste items, personal/team gates, mirror risks, and retention basics. Official references: ChatGPT, Claude, Gemini, OpenAI Help, OpenAI Privacy Policy. For non-sensitive practice only: ChatGPT domestic entry.

Outcomes

• Know what is practice-safe, redact-first, or forbidden
• Personal and team checklists you can tick
• Realistic risks of mirrors, “top-ups,” and shady extensions
• A 15-minute response if something sensitive was pasted

Data classes (label before paste)

Level	Examples	Default
L0 Public	Published posts, public docs, textbook drills	OK to chat; still verify facts
L1 Internal non-sensitive	Redacted weekly notes, public product names	OK; avoid private chat screenshots
L2 Personal sensitive	ID numbers, address, health, salary, biometrics	No raw paste; use synthetic data
L3 Secrets	Passwords, API keys, cookies, OTP, private keys	Never; rotate if exposed
L4 Business confidential	Unreleased finance, client lists, private source, M&A	No consumer chat; approved channels only
L5 Regulated	Medical/finance customer detail, minors’ data, government secrets	Forbidden unless approved systems

Rule: if unsure, treat as one level higher.

Never paste

• Passwords, recovery codes, seed phrases
• Full government IDs, bank numbers
• Client name + contact + contract value together
• Production connection strings, .env, private keys
• Full legal/medical case files
• Others’ private messages or performance reviews without authorization

Safe pattern:

Advise using ONLY the redacted material below.
Do not guess masked identities.
Do not ask me for real names, accounts, or keys.
Material:
Customer A (retail), contract value in [range], dispute type [clause class]…

Personal checklist

• Strong password + MFA if offered
• Review training / data-control settings in the current UI
• Delete sensitive threads when the product allows
• Avoid unknown browser extensions on the official site
• Log in only via chatgpt.com or official stores
• Domestic/third-party entries: L0–L1 only
• Fact-check before publishing (fact-checking guide)

10-second pre-send: scan for email/phone/employee IDs/exact amounts → mask → ask “would I mind this in a training screenshot?”

Team checklist

• Written ban-list for external AI
• Approved tool list only; no personal mirrors
• SSO, audit, seat offboarding
• Training/retention configured per contract
• Prompt library in company wiki—not random personal drives
• 30-minute onboarding + signed red lines
• Vendor/plugin review
• Quarterly spot-checks for L2+ raw data

One-pager template:

1. L2+ must not enter unapproved AI products.
2. Only company-approved domains/tools.
3. External content needs fact + sensitivity checks.
4. Accidental paste: delete thread if possible → rotate secrets → notify {security contact}.
5. Violations follow {policy clause}.

Third parties, mirrors, and “top-ups”

Risk	Impact	Mitigation
Fake login pages	Account theft	Verify domain; bookmark official
Shadow mirrors storing chats	Resale / leaks	L0 practice only; never production
Shared/top-up accounts	Bans, no invoices	Official payment only
“AI” browser extensions	Read pages/clipboard	Least privilege; IT-managed
API keys in frontends	Bill shock	Server-side keys + hard limits

If you must use third-party access, read China access and mirror sites: non-sensitive, disposable, no opaque fees.

Retention, training, and “delete = gone?”

1. Consumer ≠ enterprise terms.
2. Disabling training ≠ zero server logs (abuse/security retention may remain).
3. Deleting chats reduces reuse risk; it may not unwind past training use.
4. API/workspace options can differ—put retention in the contract.
5. Know who can export chats and for how long.

Check Settings → Privacy / Data controls against OpenAI Help (or the vendor’s help center).

Redaction prompts

Review the text for PII/secrets/business confidential items.
Table: snippet | L0–L5 | suggested replacement.
Do not rewrite the whole document.
Text: {paste}

Create a clearly fictional support ticket (name, phone, order id).
Prefix with “FICTIONAL DATA”.

15-minute incident response

1. Stop sending the same secret
2. Delete the thread / revoke share links
3. Rotate passwords, API keys, tokens
4. Scope impact (customers? screenshots?)
5. Report (personal stakeholders / team security process)
6. Retro: which level judgment failed

Scenario matrix

Scenario	OK use	Avoid
Public textbook study	Explain, quiz, critique	Upload paid PDFs to unknown sites
Public blog drafting	Outline/edit	Unreleased financials as source
Support macros	Synthetic tickets	Real tickets with addresses/phones
Coding help	Function-level snippets	Whole configs with secrets
Medical/legal structure	Public statutes as structure	Case files in consumer apps

FAQ

Will my chats train the model?

Depends on account type, toggles, and contracts. Trust settings pages and OpenAI’s privacy policy (or the vendor’s)—not hearsay.

Can I “lightly redact” company secrets?

If a customer or deal remains identifiable, it is not redacted. L4/L5 stay out of consumer tools unless security/legal say otherwise.

Are third-party mirrors safe?

Unknown mirrors can phish, steal sessions, and retain logs. L0 practice only; paid work on official sites. Treat domestic entries as third parties—read their terms.

Personal ChatGPT on a work laptop?

Follow company policy. Many require work data only in the enterprise tenant.

Extra risk from plugins/connectors?

Yes—mail, drive, repos. Least privilege; review authorizations regularly.

Related reading

• Getting started
• Choose an AI assistant
• China access guide
• AI fact-checking

Next reading

• FAQ
• Mirror sites
• Prompt tips

Action path

Today: pin the never-paste list; enable MFA; review privacy toggles. This week: classify common materials L0–L5; teams draft the one-pager and name a security contact. Ongoing: 10-second scan before send; use the 15-minute playbook on mistakes. Authoritative detail: OpenAI Help and product settings.