De-identified vs Anonymized Data: What AI Buyers Require

De-identified data has had identifiers removed or transformed to meet a specific legal test, such as HIPAA's Safe Harbor or California's CPRA definition, with some residual risk managed by controls. Anonymized data, in the GDPR sense, can't be linked to a person by any means reasonably likely to be used. Pseudonymized data, with names swapped for codes, is neither by default.

Why do the labels matter to a company selling data?

They decide which law still applies after the data leaves you. If the delivered copy is genuinely anonymous under the GDPR, the GDPR stops applying to it. If it's "deidentified" under the CPRA, it's outside California's definition of personal information. If it's only pseudonymized, every obligation stays attached.

The words also show up in marketing, and that's where companies get hurt. The FTC has said that hashing identifiers doesn't make data anonymous, and that misleading anonymity claims can be deceptive. If your notice to employees or customers says "anonymized," make sure a lawyer would agree under the standard you're actually relying on.

How do the standards compare?

What's the difference between HIPAA Safe Harbor and Expert Determination?

HIPAA gives two routes, and only two, for turning protected health information into de-identified data (HHS).

Safe Harbor is a checklist. You remove 18 categories of identifiers about the individual and their relatives, employers and household members. They include:

• Names
• Geographic units smaller than a state (the first three ZIP digits may stay if that area has more than 20,000 people)
• All date elements except the year, and ages over 89
• Phone and fax numbers, email addresses and Social Security numbers
• Medical record, health plan, account, certificate and license numbers
• Vehicle and device identifiers, URLs and IP addresses
• Biometric identifiers and full-face photos
• A catch-all: "any other unique identifying number, characteristic, or code"

You must also have no actual knowledge that what remains could identify someone. The checklist applies to free text, not only to database fields.

Expert Determination is a judgment. A person with "appropriate knowledge of and experience with generally accepted statistical and scientific principles" determines that the risk is "very small" that the information could be used, alone or with other reasonably available information, "by an anticipated recipient" to identify someone. The expert documents the methods and results. The rule sets no numeric threshold, and HHS's guidance notes that some practitioners issue time-limited determinations because the available data and technology change.

For workplace data, the trade-off is stark. Safe Harbor's catch-all and its strict date rules strip much of what makes a conversation useful. Expert Determination can keep more, but only for a defined recipient and environment. Medical practices considering a license should treat both as off-limits until counsel confirms how the buyer meets one of them; see healthcare practices.

What is the GDPR anonymization standard?

Recital 26 sets the bar. To judge whether someone is identifiable, "account should be taken of all the means reasonably likely to be used, such as singling out, either by the controller or by another person." That includes the cost, the time, and the technology available now and coming. Data protection rules "should therefore not apply to anonymous information." The same recital says pseudonymised data that could be attributed to a person "by the use of additional information" is still information about an identifiable person.

European regulators' Opinion 05/2014 turned this into three tests a technique should defeat:

1. Singling out an individual
2. Linking records about the same individual
3. Inferring information about an individual

Consistent pseudonyms in a Slack export fail the linkability test by design: the point is that "Person 07" is the same person in every message.

In September 2025, the EU Court of Justice added nuance in EDPS v SRB (C-413/23 P). Pseudonymised data may not be personal data in the hands of a recipient that has no means reasonably likely to re-identify anyone. But the original controller must still tell people at the time of collection who will receive their data (Jones Day summary). For a seller, the practical point is that the buyer's position may be lighter than yours. Your notice obligations to your own staff don't go away. Mode's referral page lists the UK among its markets, and the UK GDPR uses the same concepts.

What does the CPRA mean by "deidentified"?

California's definition in §1798.140(m) covers information that "cannot reasonably be used to infer information about, or otherwise be linked to, a particular consumer," but only if the business that holds it does three things:

1. "Takes reasonable measures to ensure that the information cannot be associated with a consumer or household."
2. "Publicly commits to maintain and use the information in deidentified form and not to attempt to reidentify the information." Testing the process is the one exception.
3. "Contractually obligates any recipients of the information to comply with all provisions of this subdivision."

Two consequences follow for workplace data. Employees are consumers: the employee-data exemption expired on January 1, 2023. And the third prong is about contracts. The seller has to bind the buyer, and a buyer that wants to treat the data as deidentified has to bind its own recipients in turn. Ask to see both.

Is aggregated data the same thing?

No, and it's rarely what AI buyers want. The CCPA separately excludes "aggregate consumer information," defined in §1798.140(b) as information about a group or category of consumers, with individual identities removed, that isn't linked or reasonably linkable to any consumer or household. "Average ticket resolution time by department" is aggregate. A thread showing how one refund dispute was escalated and fixed is record-level. Workflow buyers pay for the record-level version, so the aggregate exclusion seldom helps a seller.

Can de-identified data be re-identified later?

Sometimes on purpose, and that's allowed within limits. HIPAA lets a covered entity attach a re-identification code to de-identified data, provided the code isn't derived from information about the person and the covered entity doesn't disclose it or the mechanism for re-identification (45 CFR §164.514(c)). The CPRA lets the business attempt re-identification only to test whether its own process works.

Re-identification by anyone else is the risk you're trying to contract away. The standards assume someone keeps the key under control, and the CPRA expressly requires recipients to promise not to try. If a buyer can't say who holds the pseudonym lookup table and when it's destroyed, it can't claim the strongest version of any of these labels. The same is true if its downstream contracts don't ban re-identification.

What do AI data buyers actually require?

Mode doesn't publish a legal standard that it certifies to, and we haven't found a workplace-data buyer that does. What buyers describe is a working method:

• Mode says it "removes sensitive material and transforms personal identifiers in the agreed copy before reselling it for AI research and training" (program page). Its demo shows "Company 01 / Person 01" labels. Dan Novaes described the same idea on air: "Dan's going to become … Derek." Mode names no third-party certification.
• Google's Spirit Airlines bid, a $10 million winning bid that still needs court approval, covered "internal data and custom software." Google said it was "not buying their customer or credit card information" (Business Insider). Passenger databases were excluded and Tonic.ai was hired to de-identify the rest. Unions and 121 members of Congress still objected that conventional de-identification may not hold up against AI re-identification (Travel Weekly). The hearing is set for November 6, 2026.

The pattern is consistent. Buyers want coherent, pseudonymized conversations with direct identifiers removed. Legally, that's pseudonymization. It can meet the CPRA's "deidentified" definition if the measures, the public commitment and the recipient contracts are all in place. It is unlikely to be GDPR-anonymous from the seller's side, because the seller, or the buyer's de-identification environment, could re-link it.

We're a paid referral partner of Mode and earn a fee if your company licenses data through it (disclosure). That's one more reason to verify these points against your own contract rather than our summary.

Which standard should your company aim for?

Match the standard to the people in the data, not to the buyer's marketing:

For the mechanics of scrubbing Slack, email and Jira, see data anonymization for AI training. For contracts, employee notice and the rest of the pre-signing review, see our safety checklist. None of this is legal advice. Talk to your counsel before you sign.