Safe AI prompts: prepare, configure, and verify the result
An AI prompt can include more than typed words. Uploaded files, screenshots, images, voice, pasted logs, browsing connectors, and workspace integrations can carry hidden identifiers, metadata, comments, or complete records.
Redaction is useful only if the remaining details cannot readily identify the person, account, business, or event. The safest prompt is often a synthetic example that preserves the structure of the problem without preserving the sensitive facts.
Classify before opening the tool
Decide whether the material is public, internal, confidential, regulated, privileged, or secret before convenience shapes the decision.
Minimize the task, not just the names
A rare job title, precise date, location, transaction, and diagnosis can re-identify someone even after their name is removed.
Settings are not a universal permission
A training opt-out may not change retention, administrator access, abuse review, connectors, legal preservation, or the rules of an employer.
Outputs need their own review
Generated text can reproduce sensitive input, infer private facts, or create unsupported claims that should not enter a customer or personnel record.
Make one deliberate change at a time
Define the smallest useful question
Ask for a generic structure, checklist, formula, or critique before deciding whether any real example is necessary.
Strip direct and indirect identifiers
Remove names, addresses, account numbers, faces, document properties, unique events, exact values, and combinations that point back to a subject.
Replace rather than blank
Use consistent fictional labels and perturbed values so the model can follow relationships without receiving the original data.
Check product and workplace rules
Confirm account tier, retention, training, region, connected services, administrator access, and the policy governing the task.
Review and clean up
Inspect the output for sensitive echoes and errors, then delete chats or files where appropriate and record consequential decisions outside the model.
Check behavior, not just the toggle
| What you see | What to check | Safer adjustment |
|---|---|---|
| Prompt contains a password or API key | The secret must be treated as exposed | Revoke or rotate it; deleting the chat is not a substitute |
| Document contains hidden comments | Upload may reveal more than visible pages | Create a clean, flattened, metadata-reviewed copy or use a synthetic document |
| Free consumer tool is used for client work | Terms, retention, or organizational permission may not fit | Stop and use an approved environment or a non-AI workflow |
| Names removed but case is distinctive | Re-identification may still be easy | Generalize dates, locations, roles, amounts, and rare events or do not submit it |
Recheck product settings before sensitive work changes
Retention, training, connector, administrator, and workspace controls can change with the product or account tier. Review them before a new data class or task is approved, and delete old files or conversations where policy and operational needs allow.
Sources for Safe AI prompts
Confirm current instructions and limits for this specific task before acting.
Preserve the shape of the problem without preserving the sensitive facts
Redaction must consider combinations, not only obvious names. A rare role, exact date, location, transaction value, diagnosis, customer history, or quoted message may identify a person or event when combined. Replace those details with consistent fictional labels and altered values while keeping only the relationships needed to answer the question.
Inspect files as data containers before uploading them. Comments, tracked changes, document properties, hidden rows, image metadata, and embedded content can disclose more than the visible page. When a real file is unnecessary, create a clean synthetic sample. When the task is governed by workplace, contractual, legal, or professional duties, use the approved process rather than personal judgment alone.
Treat exposed credentials as an incident, not a redaction mistake that deletion can undo. Revoke or rotate passwords, tokens, keys, and secret links that entered a prompt. Review the output for sensitive echoes and unsupported claims before it is shared or copied into a consequential record.
Questions to answer before relying on it
Is a training opt-out enough for confidential data?+
No. It addresses one possible use. Retention, access, logging, connectors, contractual terms, and organizational policy still matter.
Can I paste medical or legal documents after removing the name?+
Not automatically. The document may remain identifiable and may be regulated or privileged. Use an approved professional workflow and obtain the necessary authority.
What should I do after accidentally pasting a secret?+
Rotate credentials immediately, notify the appropriate owner, preserve the relevant audit facts, and follow the provider and workplace incident process.
Connect this setup to the controls around it
No single detector settles a consequential decision, so connect the surrounding verification controls. Use the AI privacy & protection hub for the complete topic pathway, then continue with How AI models use your data, AI data permissions & sharing controls, and Protecting customer data.
For team-wide rules beyond one prompt, build a practical workplace AI policy.
