2026-08-25 · 4 min read

What should you never paste into ChatGPT? The complete list

What happens to text you paste into ChatGPT?

The moment you paste, the content is sent to OpenAI's servers and processed there. On consumer plans (Free/Plus), conversations may be used to train models unless you disable it in Data Controls; on business plans (Team/Enterprise), data is not used for training by default.

For your organization that means one thing: pasting an internal document into an external tool is a transfer of data to a third party — with consequences under privacy law (GDPR, and in Israel the Privacy Protection Law after Amendment 13), customer NDAs, and your own security policy.

The list: 10 things that must never go into an AI tool

  1. Personal identifiers — names alongside ID/SSN numbers, phone numbers, addresses or emails. That's personal data under privacy law.
  2. Medical information — diagnoses, treatments, patient records. Special-category data with stricter obligations.
  3. Personal financial data — card numbers, bank accounts, salaries, debts.
  4. Customer and supplier data — client lists, commercial terms, usage data.
  5. Trade secrets — pricing, strategy, M&A plans, unannounced products.
  6. Proprietary source code — the famous case: in 2023 Samsung engineers reportedly pasted internal code into ChatGPT, and the company restricted generative AI use in response.
  7. Passwords, API keys and credentials — never, in any tool.
  8. Documents under NDA — a confidentiality agreement doesn't distinguish between a person and a chatbot.
  9. Employee data — salaries, performance reviews, HR records.
  10. Classified material — anything above 'unclassified' must never touch an unapproved system, full stop.

How often does this actually happen? The numbers

  • A Cyberhaven study (2023) found that about 11% of what employees paste into AI tools is sensitive company data.
  • IBM's Cost of a Data Breach 2024 report puts the average cost of a breach at $4.88 million.
  • Regulators are catching up: in Israel, Amendment 13 to the Privacy Protection Law (in force since August 2025) enables significant fines for unlawful data processing.

How do you still use AI on sensitive documents?

The answer isn't blocking AI — employees route around blocks. The answer is a safe path: strip or replace the identifying details before the document ever meets the tool.

  1. Publish a short, clear AI policy: which tools are allowed, for which data.
  2. Prefer business plans (Enterprise/Team) with a no-training commitment.
  3. Adopt anonymization: replace names, IDs, phones, customers and amounts with consistent tokens — the AI sees the structure without the identities.
  4. Train employees on the list above — most leaks are honest mistakes.

Frequently asked questions

Does ChatGPT store what I type?

Yes — conversations are saved to your account, and on consumer plans they may be used for model training unless you opt out in Data Controls. On business plans, data isn't used for training by default.

Can I paste a client's document into ChatGPT?

If it contains identifying details or confidential client information — not without consent or without removing those details first. It's a transfer of data to a third party under privacy law, and usually an NDA breach as well.

What's the difference between redaction and consistent anonymization?

Redaction (blacking out) destroys context — you can't tell it's the same person across the document. Consistent anonymization replaces each detail with a stable token (PERSON_001), so the AI understands the relationships without receiving the identities.

Doesn't an enterprise AI plan solve this?

It reduces part of the risk (no training on your data), but not all of it: the data still leaves your organization, sits in logs, and is exposed to permission mistakes. Anonymizing before the paste adds a protection layer that doesn't depend on any vendor.