Protecting sensitive data in AI workflows
Reduce exposure across annotation, inference, review, and operational logs.
List the places where data is copied
Follow a record through upload, labeling, training, evaluation, and inference. Include reviewer exports, debugging traces, and temporary files. Each copy needs an owner, a purpose, access rules, and a retention period.
Minimize what each stage can see
Replace identifiers with stable pseudonyms where reviewers only need to link records. Use synthetic fixtures for testing logs and exports. Review whether sensitive fields appear in traces or downloads, and check that deletion workflows handle the copies your process actually creates.
Trace one record through the full workflow
Consider a document uploaded for annotation. Copies may appear in object storage, queued tasks, reviewer previews, model traces, evaluation exports, and support logs. Record which components handle the original and which can use a redacted derivative. Include backup and deletion behavior so retention decisions cover the places a record actually lives.
Minimize the reviewer’s view
For a routing task, reviewers may need the message content but only a pseudonymous account reference. For a visual task, crop irrelevant regions or mask identifiers where that preserves the annotation goal. Test these transformations on realistic fixtures. Excessive redaction can remove context needed for a label, so define the minimal fields for each task and check their effect on quality.
Limit what diagnostics and exports retain
Prefer operation names, identifiers, timestamps, and status fields for routine diagnostics. Provide access to sensitive evidence only through a scoped workflow where it is needed. Check journal entries, error messages, and download previews with synthetic records that include clearly recognizable marker strings. Examine the resulting artifacts to confirm that only the fields intended for each audience appear.
Test retention and deletion with a synthetic record
Place a unique test identifier in a record and follow it through the normal workflow. Remove the record using the documented process and verify what disappears from the application, queued work, and derived datasets. Where backups expire on a different schedule, document that separately. Give each data location an accountable owner so gaps can be corrected.