Updated for 2026
The Audit Question AI Cannot Answer
A HIPAA auditor asks: "Why was this clinical note de-identified?"
"The algorithm processed it" is not an answer.
HIPAA's Expert Determination method sets a clear bar. A qualified person must apply statistical and scientific principles. That person must show that re-identification risk is very small. The standard requires clear, on-record method — not black-box output.
Legal discovery sets the same bar. A special master asks: "Why was this paragraph redacted?" The response must name the privilege ground. It must describe the withheld material under FRCP Rule 26(b)(5). "The tool flagged it" does not satisfy that rule.
IAPP research from 2025 found that 34% of DPOs report insufficient tools for automated anonymization compliance documentation. The gap is not in detection. It is in documenting what was found and why.
What HIPAA Requires
HIPAA gives two paths under 45 CFR 164.514.
Safe Harbor: Remove all 18 specified PHI identifiers. Auditors check which entity types the tool found and how each was handled.
Expert Determination: A qualified person applies statistical principles. They document the method, the risk analysis, and their own qualifications.
Both paths share one key demand. Auditors must understand what was done. They cannot just be told it happened. A system that gives de-identified output with no method records fails both paths.
What GDPR Adds
GDPR enforcement is rising. EDPB issued 900+ enforcement decisions in 2024. GDPR fines hit €1.2 billion that year — a record.
GDPR Article 5(2) sets the accountability rule. Controllers must be able to demonstrate compliance — not just achieve it. The duty is active proof, not passive compliance.
For teams using automated anonymization tools, this rule covers the tools. A DPO must document technical measures. They need to name what the tool finds. They need to name how it finds it. They need to state what confidence is required and what action is taken. A tool that gives none of this blocks the audit duty.
Four Fields That Build the Audit Trail
An explainable redaction system must record four items per redaction.
Entity type: "PERSON" or "SSN" or "DATE_OF_BIRTH" — the class of data found. Each class maps to a HIPAA PHI type or a GDPR personal data type.
Detection method: Was this a regex match on a fixed pattern? Or an NLP model match based on context? Regex matches are fully reproducible. NLP matches carry confidence levels. That difference matters for audit records.
Confidence score: For NLP matches, this is the probability that the span is the claimed entity type. A score of 0.94 for a person name is documentable. A binary "flagged/not flagged" is not.
Operator applied: Was the entity replaced with a token, hashed, redacted, or suppressed? Naming the operator supports audit review.
These four fields are the audit trail. HIPAA Expert Determination needs it. Legal discovery privilege logs need it. GDPR accountability records need it. Without it, automated redaction cannot be defended to auditors, courts, or supervisory authorities.
See how anonym.legal captures this at the compliance overview and security practices pages. For a walkthrough of HIPAA Safe Harbor processing, see the batch HIPAA clinical notes guide.