De-identification and anonymization both mean removing the link between data and a person, but the legal tests differ. HIPAA offers two de-identification methods. The GDPR treats only truly anonymous data as outside its scope, and pseudonymized data stays personal. The CCPA adds its own definition. This is general information, not legal advice.
This page is general information, not legal advice. Whether your company can license a given dataset depends on your contracts, your privacy notices and the law where you and your customers are. Review it with your own counsel before signing anything.
Why the words matter
People use de-identified, anonymized and pseudonymized as if they meant the same thing. In law they do not. Each term belongs to a particular legal regime and carries its own test. Data that meets one test can fail another.
The difference decides which rules apply to a dataset after it has been processed. If data is anonymous under the GDPR, the GDPR no longer applies to it. If it is only pseudonymized, every GDPR duty remains. If health data is de-identified under HIPAA, it is no longer protected health information. A seller that describes a dataset with the wrong word may be making a legal claim it cannot support.
This guide sets out the main definitions in the United States and the European Union in general terms. It is not legal advice. Which standard applies to a given dataset depends on where the data subjects are, what kind of data it is and who is handling it. Confirm the position with counsel.
HIPAA: two routes to de-identification
HIPAA is the US federal law that protects health information held by covered entities, such as providers and health plans, and their business associates. Its Privacy Rule says that health information is de-identified, and no longer protected, when one of two methods is satisfied.
The first is Safe Harbor. The entity removes 18 listed categories of identifiers relating to the individual and to relatives, employers and household members. The entity must also have no actual knowledge that the remaining information could be used to identify the person. Safe Harbor is a checklist: it is predictable, but it removes detail such as exact dates and small geographic areas that can matter to a buyer.
The second is Expert Determination. A person with appropriate knowledge of statistical and scientific methods analyzes the data and determines that the risk of identifying an individual is very small, taking into account who is expected to receive it. The expert documents the methods and the result. This route can keep more detail, at the cost of an expert engagement and a written analysis.
- Names
- Geographic units smaller than a state, with a limited exception for the first digits of a ZIP code
- All elements of dates, except year, that relate to an individual, and ages over 89
- Telephone numbers, fax numbers and email addresses
- Social Security, medical record, health plan and account numbers
- Certificate, license, vehicle and device identifiers
- Web addresses and IP addresses
- Biometric identifiers and full-face photographs
- Any other unique identifying number, characteristic or code
GDPR: anonymization and pseudonymization
The GDPR applies to personal data, meaning any information relating to an identified or identifiable natural person. Recital 26 explains where that stops. The principles of data protection do not apply to anonymous information: information that does not relate to an identifiable person, or personal data rendered anonymous so that the person is no longer identifiable.
To decide whether someone is identifiable, Recital 26 says account should be taken of all the means reasonably likely to be used to identify the person, by the controller or by anyone else. The cost, the time required and the available technology are all relevant. This is a risk-based test, but regulators have read it strictly. Anonymization in the GDPR sense is meant to be effectively irreversible.
Pseudonymization is a different thing. It means processing personal data so that it can no longer be attributed to a specific person without additional information, where that additional information is kept separately and protected. Replacing a name with a code, or an email address with a hash, is pseudonymization.
Why pseudonymized data is still personal data in the EU
Recital 26 is direct on this point. Personal data that has undergone pseudonymization, and that could be attributed to a person by the use of additional information, should be considered information on an identifiable person. The GDPR treats pseudonymization as a security measure that lowers risk. It does not treat it as an exit.
The reason is the key. With a pseudonymized dataset, someone holds the table or the salt that maps codes back to people. As long as that link exists and could reasonably be used, the people remain identifiable. Even without the key, a pseudonymized record can sometimes be matched to a person through the other fields it contains.
European courts have also considered the position of a recipient who receives pseudonymized data without any access to the key. Whether the data is personal in that recipient's hands can depend on the means of re-identification reasonably available to it. This is a developing area. A seller should assume that pseudonymized data about people in the EU remains personal data unless counsel advises otherwise.
CCPA: the deidentified definition
The California Consumer Privacy Act excludes deidentified information from its definition of personal information. The definition has a factual part and a conduct part.
The factual part is that the information cannot reasonably be used to infer information about, or otherwise be linked to, a particular consumer. The conduct part places three duties on the business that holds it. The business takes reasonable measures to ensure the information cannot be associated with a consumer or household. It publicly commits to maintain and use the information in deidentified form and not to attempt to reidentify it. And it contractually obligates any recipients to comply with the same requirements.
The third duty matters in a data license. A California seller relying on this definition needs a clause in the agreement that binds the buyer not to reidentify. Several other US states have adopted similar definitions in their privacy laws, though the wording varies.
The standards side by side
The table summarizes the tests. It is a simplification. Each regime has detail and guidance behind it that the summary leaves out.
| Term | Regime | Test in brief | Effect |
|---|---|---|---|
| De-identified (Safe Harbor) | HIPAA | 18 identifier categories removed; no actual knowledge of identifiability | No longer protected health information |
| De-identified (Expert Determination) | HIPAA | Qualified expert finds the identification risk very small and documents it | No longer protected health information |
| Anonymous | GDPR | Person not identifiable by any means reasonably likely to be used | Outside the GDPR |
| Pseudonymized | GDPR | Not attributable without separately held additional information | Still personal data; all duties apply |
| Deidentified | CCPA | Cannot reasonably be linked to a consumer; public commitment and contract terms | Not personal information under the CCPA |
Re-identification risk and k-anonymity in plain words
Re-identification is working out who a record is about after the obvious identifiers are gone. It usually happens by linkage. A record with no name may still show a job title, a town and a birth year. If an outside source, such as a public profile, lists the same three facts next to a name, the record is identified.
Fields that enable this are called quasi-identifiers. None identifies a person alone. Together they can. K-anonymity is a simple way to measure the exposure. Group the records by their quasi-identifier values and find the smallest group. That size is k. If every combination is shared by at least k records, anyone trying to link a person can only narrow the search to a group of k.
A record that is unique on its quasi-identifiers has a k of one, which is the highest risk. Raising k means generalizing values, such as replacing a birth date with a year or a town with a region, or removing outliers. K-anonymity is a useful measure but not a legal threshold, and it does not address identifying details inside free text. The checker at /tools/reidentification-risk-checker reports group sizes for the columns you choose.
What this means when licensing data
For a company preparing data for sale, the practical lessons are short. Outrich Data records the de-identification method for each dataset it buys in the data card, so the buyer can see which standard is being claimed and why.
- Use the word that matches the test you have met. Do not call hashed data anonymous.
- Work out which regimes apply before choosing a method: health data, EU data subjects and California consumers each bring their own.
- Remove direct identifiers first. The scrubber at /tools/pii-scrubber handles structured identifiers in text.
- Then measure quasi-identifiers and generalize where groups are small.
- Read samples of free text by hand. Pattern matching misses names and context.
- Put a no-reidentification clause in the license.
- Document the method, the tools and the tests, and state them in the data card.
- De-identification does not override a contract. Data a customer agreement forbids you to share stays forbidden.