---
title: "De-identification"
description: "De-identification removes the 18 HIPAA Safe Harbor identifiers from health data so it falls outside HIPAA scope. How it differs from anonymization and pseudonymization, and what it means for AI workflows."
url: "https://sonomos.ai/glossary/de-identification/"
source: "https://sonomos.ai/glossary/de-identification/"
---

# De-identification

The process of removing or obscuring identifiers from health data so it no longer identifies an individual, removing it from HIPAA's scope.

## In depth

Under HIPAA, de-identified health information is no longer Protected Health Information and can be used or disclosed without restriction. HIPAA recognises two methods: Safe Harbor requires removing 18 specific categories of identifier (names, dates other than year, geographic subdivisions smaller than a state, phone numbers, email addresses, SSNs, medical record numbers, health plan beneficiary numbers, account numbers, certificate and license numbers, URLs, IP addresses, device identifiers, biometric identifiers, full-face photographs, and any other unique identifier) and having no actual knowledge that the remaining information could re-identify an individual. Expert Determination requires a qualified statistician to certify that the risk of re-identification is very small. De-identification is distinct from anonymization (a broader concept without a fixed legal standard) and from pseudonymization (where identifiers are replaced by a code that could be reversed with a key). Properly de-identified data loses HIPAA protection, which is a feature — but the de-identification process itself must be rigorous, since quasi-identifiers (ZIP code + birth date + sex) can re-identify more than 85% of the U.S. population when combined.

## Examples

- Removing the 18 HIPAA Safe Harbor identifiers from a patient dataset before sharing it with an AI vendor for clinical research.
- A statistician applying Expert Determination to a health claims dataset to confirm re-identification risk is under 0.05%.
- Replacing patient dates of birth with birth year only and removing all geographic detail below state level before AI analysis.

## How Locke handles de-identification

Locke's detection layer is aligned with HIPAA's Safe Harbor identifier list — the 18 categories of identifiers that must be removed for de-identification. By flagging these categories locally before a prompt is sent, Locke helps covered entities apply a de-identification-consistent approach to AI prompts without requiring a formal de-identification workflow for every interaction.

[See the product](https://sonomos.ai/locke/)

## Related terms

- Anonymization

  Irreversibly transforming personal data so that no person can be identified, even with additional information.
- Pseudonymization

  Replacing direct identifiers with stable pseudonyms so the data can no longer identify a person without additional information held separately.
- Protected Health Information (PHI)

  Individually identifiable health information held or transmitted by a covered entity, regulated under HIPAA in the United States.

## Reading on this topic

- [HIPAA-Compliant AI: A Complete Checklist for Healthcare Organizations in 2026](https://sonomos.ai/blog/hipaa-compliant-ai-checklist-2026/)
- [PHI vs PII vs Personal Data: A Plain-English Compliance Glossary for 2026](https://sonomos.ai/blog/phi-vs-pii-vs-personal-data-glossary/)