Deidentified vs anonymised data: what to redact for AI

Anonymised, pseudonymised and de-identified get used as if they were three shades of one colour. They aren’t. Each comes from a different rulebook, and each makes a different promise about whether someone could still work out who the data describes.

That difference decides what you can truthfully say about a spreadsheet or a case note. It also decides how much you still have to strip from a document before it leaves you. Below I take the words one at a time, then end on what to take out before a file reaches an AI tool.

Deidentified vs anonymised: which law uses which word

The GDPR (EU law, mirrored in the UK GDPR) knows only two of these words. Recital 26 treats information as anonymous when the person can no longer be identified by the means reasonably likely to be used, weighing cost, time and the technology available. Data that passes that test sits outside the Regulation altogether.

Article 4(5) defines pseudonymisation: records that point to a person again only once you add information stored apart, under its own safeguards. Still personal data. De-identified appears in neither provision; it is defined in US health law and Australia’s Privacy Act. UK firms use the same vocabulary when picking AI tools they can defend.

When a contract or a vendor page promises “de-identified” data, I want to know which rulebook it has in mind, because only some of them attach a test to the word. Firms working across borders already ask the same of the EU AI Act’s reach beyond Europe.

What “digitally anonymised” means in a documentary or the news

In a film credit or a news caption, “digitally anonymised” means the editors altered the footage so viewers can’t tell who is speaking. Usually that’s a blurred or pixelated face and a pitch-shifted voice, sometimes an actor reading the words. It names an editing job, not a legal status.

Whether it works depends on everything else in the frame. The GDPR’s question still fits: could someone, using means reasonably likely to be used, put a name to this person? For a stranger the blur is enough. For a neighbour who knows the kitchen, the dog and the accent, it may change nothing.

If I were the one being filmed, I’d ask about the voice and the room before the face. The same thinking applies when a transcript of that interview ends up typed into a chatbot: the name is gone, but the story may still point straight at you.

The “anonymized language” quiz question about AI, answered

A quiz on school or university AI guidelines asks what “anonymized language” means and offers four choices: fake names that belong to other real students, removing names and IDs and writing “the student”, a secret number code known only to staff, or writing in a different language.

The right answer is the second one. Removing identifiers and using general terms is what anonymising means. The staff-only code is pseudonymisation, because the school keeps the key. Borrowed names of real classmates attach the story to someone else, and translation changes the words, not the person.

For a teacher writing those guidelines, I’d add the line the quiz leaves out: “the student” only works if the year group, the incident and the date don’t point to one child. Notes about health or support needs deserve the most care, as the rules on special category data under UK GDPR show.

Is pseudonymised data anonymous? The EDPB draft and the SRB ruling

Not for whoever made it. The EDPB’s Guidelines 01/2025, a consultation draft adopted on 16 January 2025, treat such records as personal even when one party has them and another has the key, if the two could reasonably be combined. Erasing the key settles nothing unless what remains meets the conditions for anonymity.

The EU Court of Justice took a more relative line in EDPS v SRB (C-413/23 P) on 4 September 2025, under the regulation for EU bodies, whose definitions match the GDPR’s. A recipient who can’t lift the pseudonyms or re-identify by cross-checking may hold non-personal data. The body keeping the key still holds personal data.

I’d question any compliance plan built on the recipient’s half of that judgment. If you swapped the names, you hold the key, so the file is personal data for you, and sending it is still a disclosure you may need to account for. The decision below sorts a file by who holds what.

HIPAA de-identification: Safe Harbor and Expert Determination

In the US, de-identified has a precise meaning for health information. Under 45 CFR 164.514 there are two routes. A person with statistical expertise documents that the risk of identification by the expected recipient is “very small”, or the Safe Harbor method removes 18 kinds of identifier.

Safe Harbor carries a second condition: the covered entity must have no actual knowledge that what remains could identify someone. The list also reaches past the patient: a spouse’s name or the employer’s phone number goes too. Before a clinic hands any of this to a vendor, the HIPAA questions for AI tools come first.

Outside healthcare, I’d still borrow the list as a checklist, because it names things people forget, like fax numbers, device serials and full-face photos. A US employer’s own staff files sit outside HIPAA, as how PHI differs from PII in an HR file explains, yet the same fields identify people.

De-anonymised data: how re-identification happens

De-anonymised, or re-identified, means someone has put a name back on data that was meant to be anonymous. It rarely needs a key. It needs a few ordinary facts that together describe one person, the quasi-identifiers that NIST SP 800-188 says can “triangulate a specific individual”.

Hashing doesn’t prevent it. A July 2024 post from the FTC’s Office of Technology calls a hash a unique signature that can follow a person or device over time, and warns firms not to call hashed data anonymised. Australia’s OAIC adds that data can be personal in one setting and de-identified in another.

The calculator below does the sum a curious colleague does in their head. Below about ten matches, I don’t see how anyone can call the note anonymous, whatever was taken out. Councils releasing files meet this daily, as AI and public records requests in US local government shows.

What to redact before a document goes into an AI chat

Whichever word fits in the end, the work before an upload starts the same way. Out go the direct identifiers: names, including those of relatives and colleagues, every reference number, email addresses, phone numbers, street addresses and photos of faces. The Safe Harbor list above is a fair inventory of that first pass.

Then come details that identify only in combination: an exact date, a small town, a rare job, an employer with a dozen staff. These rarely need deleting. A month, a region or a sector in their place usually keeps what the model needs. Our field-by-field guide to what personal data should go covers each kind.

What should survive, to my mind, is the matter itself: the injury, the dispute, the figures the AI has to work with. Strip those and the answer is useless, and the next person simply pastes the raw file. The table walks one invented note through all three treatments.

Reversible labels or data nobody can trace: what to give an AI tool

For most work with ChatGPT, Claude or Copilot you want the answer back with the real names in it. That needs reversible labels, which makes the text pseudonymised and personal data for you. Call it that. I don’t trust any product page that describes a reversible swap as anonymisation.

That swap is what the Nonimo app does on your computer: names and numbers turn into placeholders like [PERSON_1] before the prompt is sent, and the real values return in the answer. In the desktop app, that text never passes through our servers.

An insurer’s questions on staff AI use come down to the same point: which of the two did you do, and can you show it? If a file never needs its names back, anonymise it properly instead and drop the key.

Sources

• Regulation (EU) 2016/679, the GDPR (EUR-Lex): Recital 26 on anonymous information and the means reasonably likely to be used; Article 4(5) on pseudonymisation.
• Judgment in EDPS v SRB, C-413/23 P, 4 September 2025 (EUR-Lex): paragraphs 76, 77, 85 and 86 on who holds personal data after pseudonymisation.
• EDPB, Guidelines 01/2025 on Pseudonymisation, public consultation: adopted for consultation on 16 January 2025; pseudonymised data remains personal.
• 45 CFR 164.514, 2024 edition (govinfo.gov): Expert Determination, the 18 Safe Harbor identifiers, actual knowledge, re-identification codes.
• FTC Office of Technology, “No, hashing still doesn’t make your data anonymous”: 24 July 2024; hashes still track people.
• OAIC, De-identification and the Privacy Act: de-identification depends on context and a reasonable likelihood of re-identification.
• NIST SP 800-188, De-Identifying Government Datasets: quasi-identifiers, linkage and residual risk; final version, September 2023.

Accessed 10 November 2026.