Short answer
De-identification removes or alters the information that identifies people or organizations, so the remaining records can be used with less privacy risk. For AI training data it means stripping names, contact details, account numbers and similar identifiers from business records. It reduces risk but does not eliminate it, especially in free-form text.
Every serious AI data licensing deal includes de-identification. It is the step that lets a company license records of how it works without exposing the people and customers who appear in them. It is also widely misunderstood. This guide explains what de-identification is, which legal standards define it, how it is done and where it falls short.
Key takeaways
- De-identification removes or changes identifying details so records can no longer reasonably be linked to a person.
- Different laws set different standards. HIPAA, the CCPA and GDPR each define it in their own way.
- Structured fields are far easier to de-identify than free-form text such as chat and email.
- De-identification protects identity. It does not make confidential content non-confidential.
- For high-risk categories, exclusion is safer than de-identification.
Three terms that are often confused
| Term | What it means | Can it be reversed? |
|---|---|---|
| De-identification | Removing or altering identifiers so data is not reasonably linkable to a person | Should not be, under the applicable standard |
| Anonymization | Irreversibly removing any link to an individual; the term GDPR uses, with a high bar | No |
| Pseudonymization | Replacing identifiers with codes while a key exists somewhere | Yes, by whoever holds the key |
The difference matters legally. Under GDPR, pseudonymized data is still personal data. Under US laws, data that meets the statutory de-identification standard generally falls outside the definition of personal information.
The legal standards
HIPAA: two formal methods
For health information in the United States, the HIPAA Privacy Rule recognizes two routes.
- Safe Harbor. Eighteen categories of identifiers are removed, including names, small geographic units, dates tied to an individual, phone numbers, email addresses and account numbers, and the organization has no actual knowledge that the remaining information could identify someone.
- Expert Determination. A qualified expert applies statistical or scientific principles and concludes that the risk of re-identification is very small.
Most companies outside healthcare are not bound by HIPAA, but its Safe Harbor list is a useful checklist of what counts as an identifier.
CCPA: three conditions
Under the California Consumer Privacy Act, information is deidentified only if the business:
- takes reasonable measures to ensure it cannot be associated with a consumer or household,
- publicly commits to keep it in deidentified form and not to attempt re-identification, and
- contractually obligates any recipient to follow the same rules.
This matters for AI data deals. A paid transfer of personal information can be a "sale" under the CCPA. Properly deidentified data is not personal information, so meeting all three conditions is what changes the legal picture. Note the third condition: the obligation has to be written into the license.
GDPR: a higher bar
For data about people in the EU and UK, anonymization must make identification not reasonably likely by any means. That is a demanding test, and many techniques that satisfy US standards do not meet it.
What gets removed from business records
In a typical deal for operational records, three groups of details are de-identified:
- Personal information. Names, email addresses, phone numbers, home addresses, ID and account numbers, and personal details mentioned in passing.
- Vendor names. The suppliers and partners that appear throughout a company's records.
- Identifiable customer details. Customer names, contacts and anything that would point to a specific customer.
How it is done
| Technique | What it does | Example |
|---|---|---|
| Redaction | Deletes the identifier | "Call [REDACTED] at [REDACTED]" |
| Replacement | Swaps in a consistent placeholder | "Maria Lopez" becomes "Person_14" everywhere |
| Generalization | Makes a value less precise | A birth date becomes a year; a city becomes a region |
| Suppression | Drops a field or record entirely | Remove the notes field; remove a whole channel |
Replacement is common for AI training data because it keeps conversations readable. A thread still makes sense when the same person is "Person_14" throughout.
Structured data is the simpler case. In a CRM, the name is in the name field and the phone number is in the phone field. Those columns can be dropped or replaced reliably.
Unstructured data is harder. In chat and email, identifiers appear anywhere, in any form: a nickname, a signature block, a street mentioned in passing. Automated detection finds most of them, not all, which is why verification matters.
The limits you should know
A law firm that advises on these deals highlights three limits.
Free text is difficult at scale. Removing identifiers from millions of messages is much harder than removing them from structured fields.
Context identifies people. "The plant manager at our Ohio site" may point to one person with no name attached. Buyers also want records linked across systems, and links can narrow a description down to an individual or a small group.
Sensitive stays sensitive. A complaint about pay, a health disclosure or a customer's confidential pricing does not become harmless when the name is removed. De-identification does not satisfy a confidentiality clause.
The practical conclusion: use de-identification for ordinary operational records, and exclude high-risk categories entirely. See what data can't be licensed for AI training.
Who does it, and what to put in the agreement
In many deals the buyer, or a provider the buyer uses, performs the de-identification. That is workable, but the company licensing the data should stay in control of the standard. Legal advisers recommend that the company:
- chooses, or jointly chooses, the de-identification provider,
- approves the written protocol before any data moves,
- requires certification that the protocol was followed,
- limits access to raw data and uses secure transfer,
- binds the buyer by contract not to attempt re-identification,
- agrees who bears the risk if something is missed.
See key terms in an AI data licensing agreement.
Where DataRelai stands
We only work with buyers who anonymize and de-identify data. In the deal you approve, personal information, vendor names and identifiable customer details are de-identified. The buyer works with you to qualify, transfer and de-identify the dataset. We never process or receive your data ourselves. When you describe your company on our intake form, leave out customer names and personal data.
Frequently asked questions
Is de-identified data still personal data?
It depends on the law. Data that meets the CCPA's deidentification conditions is not personal information under that statute. Under GDPR, only truly anonymized data falls outside the rules, and pseudonymized data does not.
Can de-identified data be re-identified?
Sometimes, especially when records are detailed or linked to other sources. That is why agreements prohibit re-identification attempts and why high-risk data is excluded.
Does de-identification make any data licensable?
No. It addresses privacy risk. It does not override confidentiality clauses, privilege, platform terms or sector rules.
Who is responsible if identifying details are missed?
That depends on the agreement. Allocation of privacy risk is a term to negotiate, not an assumption to make.
Sources
- U.S. Department of Health and Human Services, guidance on de-identification under the HIPAA Privacy Rule
- Frankfurt Kurnit Klein & Selz, "Selling Company Data for AI Training: What to Know Before You Say Yes" (October 1, 2026)
This guide is general information, not legal, tax or financial advice. Whether a particular dataset can be licensed depends on what it contains and on obligations your company already has.