The Safe Input Decision
Decide what to type, upload, record, or connect before information leaves your hands.
In this lesson, you will learn to:
- Identify information that should not enter an unsuitable service.
- Transform a task so it uses less or synthetic data without losing its purpose.
The Safe Input Decision
Teaches direct and indirect identifiers, sensitivity, minimization, masking limits, synthetic examples, and safe prompt transformation.
Names are only the beginning
Sensitive input can include names, faces, voices, addresses, health details, finances, credentials, private messages, unreleased work, source code, legal material, client records, location, and security findings. Removing a name is not enough when a rare job, exact date, small town, case description, or unique combination can point back to a person. Pseudonyms reduce direct visibility but may still be reversible.
Ask whether the service needs the real record at all. A writing task may need structure and tone, not identities. A classification test may work with invented examples. A spreadsheet question may need column names and a few synthetic rows, not the production file. If the task cannot work without sensitive data, the correct next step may be a service specifically suitable for that data—or completing the task without AI.
Rewrite the task before rewriting the data
Start by stating the purpose in abstract terms. Replace names and identifiers with neutral roles, remove fields irrelevant to the task, generalize dates or locations, and substitute representative synthetic values. Preserve only the pattern the model must process. Then inspect the transformed input as if you were an outsider: could it still reveal a person, secret, customer, investigation, or system?
For example, do not paste a complete complaint to improve its tone. Extract the communication pattern: “Rewrite this fictional response so it acknowledges frustration, explains a two-day delay, and avoids admitting unsupported facts.” Do not assume instructions such as “keep this confidential” change how the service handles data. The service receives the data before it can follow the instruction.
Screenshots, recordings, and connectors are inputs too
People often remember not to paste text but forget that screenshots can expose names, tabs, URLs, messages, and hidden notifications. Meeting assistants may capture voices, attendance, side conversations, and screen shares. Browser extensions and connected assistants may read far more than the current prompt. A pasted link can trigger retrieval of a page that contains personal or confidential information.
Before uploading or connecting, inspect the complete data path. Crop images, close unrelated windows, obtain appropriate consent for recordings, and choose the narrowest connector scope. Treat an AI feature embedded inside another application as its own service with its own terms and data behavior. Convenience does not remove disclosure; it changes how disclosure happens.
Resources
- CISA: Stay Safe Online When Using AI — Use CISA’s public awareness sheet as a concise reminder not to share personal or confidential details and to verify AI content.
- EDPB Opinion 28/2024 — Read the European Data Protection Board’s analysis of personal data, model anonymity, legitimate interest, and deployment considerations.