Automated PII Detection Prompt AI Prompt

Design a prompt and validation framework for using LLMs to detect PII in unstructured text at scale.

Data type: {{data_type}} (customer emails, support tickets, free-text form fields, documents)
Volume: {{volume}}
Acceptable false negative rate: {{fnr}} (missed PII — lower is better for compliance)

1. The detection prompt (to be applied to each text sample):

   System instruction:
   'You are a privacy compliance assistant. Identify all personally identifiable information (PII) in the following text. Be conservative — when in doubt, flag it.'

   Task instruction:
   'Scan this text and identify every instance of PII. For each instance found:
   - Quote the exact text
   - Classify the PII type: name / email / phone / address / SSN / date-of-birth / financial / health / government-ID / IP-address / username / other
   - Confidence: High (clearly PII) / Medium (likely PII, context-dependent) / Low (possible PII, may be fictional or generic)

   If no PII is found, return: {"pii_found": false}

   Return ONLY a JSON object matching this schema:
   {
     "pii_found": true,
     "instances": [
       {"text": "...", "type": "...", "confidence": "High|Medium|Low", "start_char": N, "end_char": N}
     ]
   }'

2. Sensitivity settings by use case:
   - For compliance scanning (minimize false negatives): flag all Medium and Low confidence instances
   - For redaction workflows (minimize false positives): flag only High confidence instances
   - For audit sampling: flag High + Medium; review Low manually

3. Validation framework:
   - Create a golden test set of 200 labeled text samples (100 with PII, 100 without)
   - Measure: precision, recall, F1 at each confidence threshold
   - Acceptable recall for compliance: ≥ 95% (missing < 5% of true PII)
   - Measure false positive rate: flag non-PII flagged as PII (acceptable up to 15% for initial triage)

4. Known failure modes to test:
   - Fictional PII (novel character names, example data) — should not be flagged
   - Partial PII (first name only with no other context) — judgment call, document the policy
   - PII in non-English text — test language coverage
   - Obfuscated PII (john[at]email[dot]com) — should be flagged
   - PII in code or SQL queries embedded in text

5. Redaction approach (after detection):
   - Replace detected PII with: [REDACTED-{type}] (e.g. [REDACTED-EMAIL])
   - Log: original text hash, PII types found, redaction timestamp, operator ID
   - Never log the actual PII values in the audit log

Return: the detection prompt, JSON schema, validation framework, golden test set design, and redaction specification.