PII Removal Before Sharing Data with AI

A practical checklist and workflow for identifying and removing personally identifiable information from M&E data before using AI tools for analysis.

AI for M&E•Ben Playfair••6 min read
aiprivacyPIIdata protectionGDPRanonymization

Why This Matters

M&E practitioners increasingly want to use AI tools - ChatGPT, Claude, Gemini - for qualitative analysis, report drafting, and data interpretation. But M&E data regularly contains personally identifiable information (PII): beneficiary names, locations, health status, contact details, and sensitive demographic combinations.

Cloud AI transfers data to another processing environment. Check the purpose, permission or lawful basis, applicable law, organizational rules, funding agreement, and actual service contract before sharing. Review storage, access, onward transfers, retention, deletion, and model-training use separately. A vendor name does not establish these controls.

Removing names is only one safeguard. Codes, roles, locations, and narrative details can still identify people. Use an approved environment or withhold the data when the remaining risks cannot be managed.

What Counts as PII in M&E Data

Direct Identifiers (Must Always Remove)

These identify a specific individual on their own:

Names & Personal Details:

  • Full names, first names, surnames, nicknames
  • Mother's or father's names
  • Community or clan names

ID Numbers:

  • National ID, passport, social security numbers
  • Employee, beneficiary, or participant IDs
  • Health insurance or medical record numbers

Contact Information:

  • Phone numbers (mobile or landline)
  • Email addresses
  • Physical addresses, GPS coordinates
  • Postal codes (if precise enough to identify location)

Dates:

  • Birth dates or exact ages
  • Enrollment or registration dates
  • Dates of specific incidents (replace with year only or age ranges)

Digital Identifiers:

  • IP addresses, device IDs
  • Social media handles
  • Biometric data, photos showing faces

Indirect Identifiers (Review Carefully)

These can identify individuals in combination:

  • Small geographic areas (specific village, school, clinic)
  • Age + gender + location combinations
  • Unique job titles or organizational roles
  • Rare medical conditions or disabilities
  • Open-ended text with personal narratives
  • Unique characteristics ("the only refugee from X country")
  • Small groups or distinctive combinations, with risk dependent on context rather than a universal safe headcount
  • Sensitive combinations (HIV status + village + age group)

The combination test: If combining two or more data points could identify a specific person, treat the combination as PII even if individual fields seem safe.

The Removal Workflow

Step 1: Audit Your Data

Before processing, catalog what PII exists:

For structured data (Excel/CSV):

  • Review all column headers for PII fields
  • Check merged cells or hidden columns
  • Remove file metadata (author, last modified by)
  • Look for PII in unexpected columns (comments, notes fields)

For qualitative data (transcripts, interviews):

  • Search for proper nouns (names, places, organizations)
  • Flag specific locations or landmarks mentioned
  • Identify unique personal stories that could identify someone
  • Check for accidentally included contact information

For reports and documents:

  • Remove participant names from narratives
  • Delete acknowledgments with beneficiary names
  • Check headers/footers for file paths or author names
  • Review annexes for raw data tables

Step 2: Choose Your Approach

| Method | Best For | Effort | Reversible? | |---|---|---|---| | Manual redaction | Small datasets (fewer than 50 records), documents | High | No | | Find-and-replace | Structured data with known PII columns | Medium | No | | Approved detection plus human review | Larger datasets or mixed content | Depends on tool and data | Depends on the transformation and retained keys | | Pseudonymization with a separate key | Approved analysis requiring linkage | Medium | Yes; the data remains potentially identifiable |

Step 3: Replace, Don't Just Delete

Deleting PII can make data unusable. Instead, replace with consistent placeholders:

| Original | Bad (deleted) | Good (replaced) | |---|---|---| | "Maria Gonzalez" | [BLANK] | "PARTICIPANT_047" | | "Kibera, Nairobi" | [BLANK] | "URBAN_SETTLEMENT_A" | | "Age: 34" | [BLANK] | "Age: 30-39" | | Sensitive health status | [BLANK] | Omit unless necessary and approved; coding alone does not remove sensitivity |

Replacement rules:

  • Names → sequential codes (PARTICIPANT_001, PARTICIPANT_002)
  • Locations → generalized categories (RURAL_AREA_A, URBAN_CENTER_B)
  • Exact ages → age ranges (18-24, 25-34, 35-44)
  • Dates → year only or quarter
  • Sensitive conditions → retain only if necessary and approved; a generic code can still reveal sensitive information

These are risk-reduction techniques, not automatic anonymization. Reversible tokenization is pseudonymization. Keep the linking key separately with restricted access and a retention plan; assess whether the remaining information can identify people without it.

Step 4: Verify Before Sharing

Before pasting any data into an AI tool, run a final check:

  • [ ] No names remain anywhere in the data
  • [ ] No contact information (phone, email, address)
  • [ ] No ID numbers or unique identifiers
  • [ ] Locations generalized to appropriate level
  • [ ] Ages converted to ranges
  • [ ] Sensitive health/legal status coded
  • [ ] Open-ended text reviewed for embedded PII
  • [ ] Residual identification and community-harm risks assessed for the intended recipient and use
  • [ ] Processing purpose and authorization confirmed
  • [ ] Access, storage, transfers, retention, deletion and training use checked against the actual contract
  • [ ] File metadata cleaned

Choosing a Processing Environment

Use tools approved for the data and purpose. Automated detection can miss indirect identifiers or redact analytically important material; verify the result against the protected original. Local processing can reduce external transfers, but device access, logs, backups, telemetry, and deletion still need controls. Do not interpret local hosting or encrypted keys as proof of anonymity or compliance.

Key Principles

  1. Authorize and Minimize Before Processing: use only data and environments approved for the task
  2. Replace, Don't Delete - maintain data utility while removing identifiers
  3. Check Combinations - individual fields may be safe, combinations may not be
  4. Document Your Process - keep a record of what was anonymized and how
  5. Use the Right Tool - manual review for small datasets, automated detection for large ones
  6. Stop When Safeguards Are Unclear: use synthetic examples or a non-AI workflow until the responsible reviewer resolves the risk

Privacy source: ICO guidance on security and data minimization in AI. This is UK guidance; confirm the law and policy applicable to your processing.

Back to AI for M&E