AI Safety and Privacy: What Not to Put in a Prompt

Use a red, yellow, or green check before you prompt

Red data should not go into a general AI prompt: passwords, OTPs, API keys, card or bank details, Aadhaar or PAN numbers, patient records, private customer data, confidential contracts, unreleased source code, and another person’s private work. Yellow data includes internal notes or documents you may use only with permission, an approved account, and the minimum necessary details. Green data is public, non-sensitive information or material you own and are comfortable sharing under the tool’s terms.

• Ask: “Whose data is this, and do I have permission to share it?”
• Ask: “Can I complete the task with placeholders or a smaller excerpt?”
• Ask: “Would disclosure harm a person, customer, employer, or account?”

Redact by changing the example, not hiding the format

Replace names with labels such as “Customer A”, round or invent non-sensitive example numbers, remove exact dates and addresses, and share only the paragraph or error needed for the task. Base64 encoding is reversible and is not encryption or anonymisation. A person can sometimes still be identified after obvious fields are removed, so do not upload a full document merely because you deleted the name.

Check account controls—but understand their limits

Consumer AI products provide different controls for activity, model improvement, temporary conversations, deletion, and connected services. These controls and retention periods can change. Check the official settings page for the account you are using; work and school administrators may control options differently. Turning off model improvement does not grant permission to upload employer or customer data, and deleting a chat may not remove information already sent to a connected service.

Use stricter rules for work, school, health, and money

CERT-In advises people to avoid sharing personal, sensitive, confidential client, and intellectual-property information with public generative AI services. For professional tasks, use only tools approved by the organisation and follow its acceptable-use, access, retention, and incident-reporting process. Keep diagnosis, employment, lending, legal, tax, disciplinary, and other high-impact decisions with qualified humans.

If you shared something sensitive by mistake

Stop the conversation and do not repeat the data. Delete the chat or activity where the product allows it, disconnect any unnecessary integration, and rotate exposed passwords, keys, tokens, or payment credentials immediately. For workplace, customer, health, or regulated data, notify the appropriate IT, security, privacy, or compliance contact promptly; deletion alone may not complete the required response.

1. 1Define the exact outcome you want from What Not to Put in a Prompt.
2. 2Give the AI the necessary context, constraints, examples, and preferred format.
3. 3Review the result for accuracy, tone, privacy, and completeness before using it.
4. 4Save or reuse the prompt only after it produces a reliable result for your use case.

• Pasting an unredacted log, bank statement, medical record, contract, source file, or customer conversation into a public AI service.
• Assuming deletion, temporary-chat mode, Base64 encoding, or turning off model training is the same as confidentiality.

Recommended tools for this workflow

Frequently asked questions