PII Detection Experiments

5 guided experiments to understand how PII detection handles different types of personal information.

What you'll learn: How different PII types are detected, the role of confidence thresholds, and when human review is needed.

Experiments completed: 0 of 5

Available experiments

Obvious PII Detection

Safe to try This experiment won't break anything. You can always reset.

5 min

What you'll learn:

AI reliably detects standard PII: names, addresses, phone numbers

Try it yourself

Copy this text and paste it into the chatbot:

Test a document with clear personal information

Expected response

All names, addresses, phone numbers and emails are redacted

Screenshot showing the expected chatbot response for Obvious PII Detection

Subtle PII Detection

Safe to try This experiment won't break anything. You can always reset.

5 min

What you'll learn:

AI can detect less obvious PII: staff IDs, case references

Try it yourself

Copy this text and paste it into the chatbot:

Test a document with staff IDs and case numbers

Expected response

Staff IDs and case references identified with lower confidence

Screenshot showing the expected chatbot response for Subtle PII Detection

False Positive Management

Safe to try This experiment won't break anything. You can always reset.

10 min

What you'll learn:

Some place names may match person names - review is essential

Try it yourself

Copy this text and paste it into the chatbot:

Test a document mentioning places like Reading, Hamilton, or Chester

Expected response

Some false positives flagged (e.g., 'Reading' detected as a name)

Screenshot showing the expected chatbot response for False Positive Management

Confidence Threshold Adjustment

Safe to try This experiment won't break anything. You can always reset.

5 min

What you'll learn:

Lower threshold catches more PII but increases false positives

Try it yourself

Copy this text and paste it into the chatbot:

Compare results at 80% vs 60% confidence thresholds

Expected response

Lower threshold redacts more items but includes false matches

Screenshot showing the expected chatbot response for Confidence Threshold Adjustment

Before/After Comparison

Safe to try This experiment won't break anything. You can always reset.

5 min

What you'll learn:

Visual comparison helps understand what was redacted

Try it yourself

Copy this text and paste it into the chatbot:

View original and redacted documents side by side

Expected response

Clear visual comparison showing redacted sections

Screenshot showing the expected chatbot response for Before/After Comparison

What's Next?

Architecture Tour

Understand how PII detection and redaction work together.

Take the tour

Test the Limits

Push the boundaries to understand edge cases.

Try challenges