Skip to main content

This is a prototype vision of how a future government service could work. It's not a real service yet, but we're exploring what it could look like. Your feedback will help shape the real service.

Step 4: Review Detected PII with Confidence Scores - FOI Redaction Walkthrough

Examine AI-detected PII and understand confidence scores for each detection

Walkthrough progress

Step 4 of 5 • 3 minutes

Step 4 3 minutes

Review Detected PII with Confidence Scores

Examine the PII the AI found and understand confidence scores to know when human review is needed.

Find your FOI Redaction tool URL in CloudFormation Outputs
The FOI Redaction tool ready to process sensitive documents

Expected outcome

  • All PII in the document has been detected
  • Each detection shows a confidence score
  • You understand high/medium/low confidence thresholds
  • You can identify which detections need human review

What you should see

After processing completes, you'll see a results section with detected PII. The exact results depend on which sample document you uploaded:

Understanding confidence scores

Each PII detection includes a confidence score representing how certain the AI is that it found real personal information. This helps you decide which detections to trust and which need human review.

High Confidence 95-100%
Meaning
Very high confidence - PII detected with near certainty. AI is 95%+ sure this is personal information.
Recommended action
Auto-redact - Safe for production redaction without human review
Examples
Email addresses (99%), phone numbers with standard formats (98%), full names in clear context (97%)
Medium Confidence 80-94%
Meaning
Likely PII but some uncertainty. AI is 80-94% confident - usually correct but not certain.
Recommended action
Flag for review - Quick human check before redaction (5-10 seconds per item)
Examples
Partial addresses (no postcode), unusual name formats, ambiguous phone numbers
Low Confidence Below 80%
Meaning
Possible false positive. AI is less than 80% confident - may or may not be actual PII.
Recommended action
Manual review required - Human judgment needed to confirm or reject
Examples
Organization names flagged as person names, street names without numbers, generic email addresses

Production tip: In a real FOI system, you'd set a confidence threshold (e.g., 95%) to automatically redact high-confidence detections and flag lower scores for review. This balances automation with accuracy.

Verify accuracy

Compare the AI detections against the original PDF document:

  1. Open the sample PDF you downloaded earlier

    Find it in your downloads folder and open it alongside the redaction interface

  2. Check each detected PII item

    For each detection in the results, locate it in the PDF. Does the text match exactly? Is it actually PII that should be redacted?

  3. Look for missed PII (false negatives)

    Scan the PDF for any personal information the AI didn't detect. This is rare with our sample documents (97-99% recall rate) but can happen with unusual formats.

  4. Check for false positives

    Are any detections NOT actually PII? For example, organization names flagged as person names. Low confidence scores often indicate these edge cases.

Key observations

Context-aware detection

AI understands context. "Manchester" in an address is flagged as PII, but "Manchester City Council" is correctly identified as an organization, not personal information.

Multi-page coverage

AI processes the entire document, finding PII regardless of where it appears - headers, footers, tables, or body text across all pages.

Format independence

Phone numbers work with any format: 07712 345678, 0121-456-7890, +44 20 7946 0958. No need to standardize formats before processing.

UK-specific patterns

Correctly identifies UK postcodes (M1 4BT, LS1 5AB), National Insurance numbers (AB 12 34 56 C), and UK address formats including flats and building names.

Calculate your time savings

Based on what you just saw, estimate your council's FOI redaction savings:

Troubleshooting

Some PII not detected (false negatives)

If the AI missed some personal information:

  • Check if PII is in an unusual format (all caps, no spaces, abbreviations)
  • Verify text is machine-readable (scanned documents may need better OCR)
  • Look for handwritten text (Comprehend works best on typed text)
  • Check if PII is in tables or complex layouts (may reduce detection accuracy)
  • Unusual UK formats may not be recognized (very old postcode styles, etc.)

Our sample documents should detect 97-99% of PII. In production, always perform final human review before publishing FOI responses to catch edge cases.

Non-PII flagged as personal information (false positives)

If the AI flagged text that isn't actually PII:

  • Check confidence score - false positives often have lower scores (<90%)
  • Common false positives: organization names, building names, generic email addresses (info@council.gov.uk)
  • Review the context - is this actually personal or organizational information?
  • In production, you'd manually un-mark false positives during review step
  • Consider adding common false positives to an allowlist (e.g., council building names)

Setting a higher confidence threshold (95%+) reduces false positives but may increase manual review workload.

Confidence scores seem inaccurate

If confidence scores don't match your expectations:

  • Remember: scores reflect AI's certainty, not ground truth accuracy
  • Low-quality scans or unusual fonts reduce confidence even if detection is correct
  • Uncommon name spellings or address formats may have lower scores despite being accurate
  • Confidence is based on statistical patterns from millions of training documents
  • In production, track actual accuracy vs confidence scores to calibrate your threshold
Note Screenshot placeholder: In production, this page would include screenshots showing the PII detection results table with highlighted confidence scores and a side-by-side comparison with the original PDF.

Need help?

If you're stuck, check the troubleshooting section or contact us.