De-identification vs anonymization depends on which framework uses the words. HIPAA calls health information de-identified when there is no reasonable basis to believe it can identify a person [1]. The GDPR never defines "anonymisation" and leaves anonymous information outside its scope [3]. NIST uses de-identification as the general term and recommends avoiding the word anonymization [4]. Pseudonymization, the replacement of identifiers with codes that a separately held key or table links back to the person, is one type of de-identification in NIST's terms, and the GDPR still treats its output as personal data [3, 4, 11].
The difference matters whenever records about people are shared, licensed or used to train models, because each label makes a different promise, and the EU's test for anonymity moved in 2025 and 2026 [10, 11, 12]. This article is research on what laws, regulators and standards say, never legal advice.
De-identification vs anonymization
| Framework | Term | What it means there |
|---|---|---|
| HIPAA | De-identified | No reasonable basis to believe the information can identify a person, reached by one of two methods. No longer protected health information [1, 2] |
| GDPR | Pseudonymized | Cannot be attributed to a person without additional information kept separately. Still personal data [3] |
| GDPR | Anonymous | Does not relate to an identifiable person, judged by the means reasonably likely to be used. Outside the Regulation [3] |
| NIST | De-identification | Any process that removes the association between identifying data and the person [4] |
| NIST | Anonymization | De-identification that cannot be reversed, a word NIST recommends avoiding [4] |
| CCPA | Deidentified | Cannot reasonably be linked to a consumer, backed by measures, a public commitment and contracts [13] |
NIST's reason for avoiding anonymization is practical. An auxiliary dataset could reverse a de-identification, and "it is not possible to know whether such an auxiliary dataset exists" [4]. The GDPR's English text uses "pseudonymisation" 14 times and never uses the words "anonymisation" or "de-identification" (counted for this article on 8 November 2026) [3]. The UK Information Commissioner's Office does not encourage "de-identified" as a synonym for anonymous or pseudonymous data, because UK data protection law does not define it [17].
Pseudonymization keeps a key
The GDPR defines "pseudonymisation" as processing after which data "can no longer be attributed to a specific data subject without the use of additional information", which is kept separately and protected [3]. Recital 26 adds that pseudonymized data that additional information could attribute to a person are information on an identifiable person [3]. The Article 29 Working Party put it plainly in 2014: "pseudonymisation is not a method of anonymisation" [16].
The European Data Protection Board (EDPB) named the additional information in January 2025. Cryptographic keys and tables matching pseudonyms to the data they replace are "pseudonymisation secrets", and "Lookup tables are personal data since they allow the identification of data subjects" [11]. HIPAA has its own version. A covered entity may keep a code that re-identifies de-identified records, if the code is not derived from information about the person and the re-identification mechanism is not disclosed [1].
Hashing gets its own warnings. NIST counts encryption, keyed hashing and realistic surrogate values as pseudonymization techniques, and warns that hashing without a key "generally does not confer security", because an attacker can hash every possible value [4]. The FTC's technology staff wrote in July 2024 that a hash "still creates a unique signature that can track a person or device over time", and that companies should not claim hashing makes data anonymous [14].
HIPAA's two routes: Safe Harbor and Expert Determination
Safe Harbor removes 18 kinds of identifiers of the person and of relatives, employers and household members, among them names, geographic units smaller than a state, date elements other than the year for dates tied to the person, phone numbers, record and account numbers, device identifiers, IP addresses and full-face photographs [1]. The first three ZIP digits may stay only for areas of more than 20,000 people [1]. The covered entity must also have no actual knowledge that what remains could identify someone [1].
Expert Determination asks a person with statistical and scientific expertise to find, and document, that the risk is "very small" that an anticipated recipient could identify someone [1]. HHS sets no numeric level that meets that test, and some practitioners issue time-limited certifications because technology and available information change [2]. Under either route, HHS writes, the risk "is not zero" [2].
A limited data set shows where the line sits. It removes 16 direct identifiers but may keep town or city, state and ZIP code, remains protected health information and needs a data use agreement [1].
Anonymous for whom: the EU test after SRB
In January 2025 the EDPB said pseudonymized data stay personal even when the data and the additional information are not in the hands of the same person [11]. On 4 September 2025 the EU Court of Justice held in EDPS v SRB that pseudonymized data must not be regarded as personal data "in all cases and for every person" [12]. Comments shared with an outside valuation firm could be non-personal for that firm, but only if it was not in a position to lift the pseudonymization and the measures also kept it from attributing them by other means, such as cross-checking [12]. For the controller that held the additional information, they stayed personal [12].
On 7 July 2026 the EDPB adopted Guidelines 02/2026 on Anonymisation for public consultation, taking the ruling into account [9, 10]. They assess anonymity "from each relevant entity's perspective" against three criteria: no record isolation, no linkage to another dataset and no inference [10]. The consultation ran until 30 October 2026, and on 8 November 2026 the EDPB's page showed it closed and offered version 1.0 only [9, 10].
The definition itself is the subject of a proposal. The European Commission's Digital Omnibus proposal of 19 November 2025 would write an entity-relative test into Article 4(1) GDPR, and in February 2026 the EDPB and the European Data Protection Supervisor urged the co-legislators not to adopt that change [30, 36].
Why removing names is not enough
Quasi-identifiers do the identifying. ZIP code, gender and date of birth likely made 87% of the U.S. population unique in 1990 census data [18, 19]. Against 2000 census data, the same three made about 63% unique [19]. A 2019 study estimated that 15 demographic attributes would re-identify 99.98% of Americans in any dataset [20]. In a movie-rating dataset of 500,000 subscribers, released for a public prize, an adversary who knows a little about a subscriber can identify that subscriber's record [21]. Four points in time and place were enough to uniquely identify 95% of 1.5 million people in fifteen months of mobility data [22].
NIST describes quasi-identifiers as the material of linkage attacks [4]. A linkage attack joins a record to another record about the same person from a different dataset [4, 10]. It is the matching that entity resolution performs across sources, aimed at people who were meant to stay unnamed. HHS adds that a covered entity's knowledge of such studies does not by itself count as actual knowledge under Safe Harbor [2].
Free text and model training
HHS applies Safe Harbor to identifiers wherever they appear, in structured fields and free text alike [2]. Text is harder to clean [5]. In a 2014 shared task on clinical narratives, three of ten systems scored F1 above 0.90 [24]. A benchmark of 1,268 cases from the European Court of Human Rights marks the text spans to mask beyond predefined categories of identifiers [25]. One open-source toolkit for detecting personally identifiable information (PII) states that its automated detection gives no guarantee of finding all sensitive information [34]. Finding identifiers is a detection task scored against labeled text [24], much like the field tests in how to test document extraction accuracy before you buy.
Models add a second exposure. A study first posted in December 2020 extracted verbatim training text, including public names, phone numbers and email addresses, from a public language model released in 2019 [27]. A 2023 study found that "Scrubbing techniques reduce but do not prevent the risk of PII leakage" [28]. The EDPB's opinion of December 2024 holds that models trained with personal data "cannot, in all cases, be considered anonymous" [29].
What to ask of a de-identified dataset
A claim that data are de-identified or anonymous can be checked with five questions.
- Which standard and method? HIPAA Safe Harbor, Expert Determination, GDPR anonymity for a named recipient, or the CCPA's definition [1, 3, 10, 13].
- Who holds the key? Keys and lookup tables are additional information in the GDPR's sense, and their release compromises identities [4, 11].
- Was re-identification tested? NIST describes a motivated intruder test that uses public, commercial or private datasets [4].
- What binds the recipient? The CCPA's definition requires reasonable measures, a public commitment not to re-identify and contracts that bind recipients [13], the same three protections the FTC set out in 2012 [15].
- When was the risk reassessed? HHS describes time-limited certifications [2], and the EDPB calls periodic reassessment good practice [10].
The other checks on a seller's file are in what to check before you license a dataset. Rillor's training data cleaning removes or replaces personal details to the rules a client sets, and the results of the checks run on every cleaned version ship with the data.
Read the full report
The full report sets out each definition in its own words, follows the EU decisions of 2025 and 2026 in order, covers k-anonymity, differential privacy and the evidence on text, and lists all 36 sources.