De-identified vs pseudonymised under Australia's Privacy Act

People use de-identified and pseudonymised as if one were a stronger dose of the other. Australian privacy law doesn’t. The Privacy Act defines de-identified as an outcome, and never mentions pseudonymised at all.

That gap decides what you can safely call a file, and which privacy principles still follow it when it leaves. Whether a given AI tool is allowed in the first place is covered in which AI tools the Privacy Act lets you use.

Is de-identified the same as pseudonymised?

No, and the statute shows why. Compilation No. 104 of the Act, dated 4 June 2026, defines personal information and de-identified, one the mirror of the other. Open it on the Federal Register of Legislation and search for “pseudonymis”: nothing comes up.

The nearest words, pseudonym and pseudonymity, belong to APP 2. It gives a customer the option of dealing with you under a made-up name. Where the law needs identity, as with a rental application, the applicant can still limit reuse by sending a Victorian licence copy watermarked for that agent.

The OAIC’s APP 2 guidelines say using a pseudonym “does not necessarily mean that an individual cannot be identified”. Written about customers, it fits a file of [PERSON_1] labels just as well. Overseas law splits the words differently, as the UK, EU and US vocabulary guide shows.

Another word for de-identified, and words that only sound like it

Anonymised is the word most people reach for, and in Europe it does the job de-identified does here. Health research has its own vocabulary. The ALRC noted in 2008 that the National Statement on research ethics avoided “de-identified data” because its meaning was unclear.

Contracts and vendor pages add a third layer: masked, tokenised, hashed, coded. Those name what was done to a field, not where the file ended up. In Victoria, OVIC quotes a description of pseudonymisation as masking with a pseudonym, and says plainly that it doesn’t make information de-identified.

When a supplier’s contract says “anonymised” with no method beside it, ask for the method in writing. Pick a word below to see what it promises.

Pseudonymised data and the APPs: which ones still apply

Whoever holds the key holds personal information, and every principle that applies to them keeps applying. The OAIC’s de-identification guide makes the point with a custodian who keeps the original. The same dataset can be personal information on that desk and de-identified with another entity.

So ask twice: what the file is in your hands, and what it becomes in the recipient’s. A labelled extract sent to an overseas contractor who could identify anyone still raises APP 8. Councils answer to state law instead, set out in which privacy law covers an Australian council.

What surprises people is APP 11.2. Keeping a pseudonymised copy with its key is not destroying or de-identifying anything.

In Victoria, agencies and councils looking at AI tools run the same comparison under the PDP Act, whose s 3 definition of de-identified is close to word for word the federal one.

Can de-identified data be re-identified? The reasonable likelihood test

Yes, and that is the whole test. The OAIC treats information as de-identified when there is no reasonable likelihood of re-identification, which it also puts as a very low risk. Possible isn’t enough; likely is what counts.

As a floor, the OAIC uses a motivated intruder: a competent person with no specialist skills, holding the data, the internet and public documents. It also looks at the data access environment, in four parts: other data, people, infrastructure and governance. A locked research lab and a shared inbox score differently.

OVIC goes further and tells Victorian agencies to assume unit-record data can be re-identified once it meets other data. A client spreadsheet deserves the same suspicion: one row per person is exactly what that warning is about.

De-identified data examples from two OAIC cases

Two published OAIC matters show both outcomes, and they’re more useful than any definition. In 2016 the Department of Health put MBS and PBS claims for a random 10% sample of people on data.gov.au. Service dates were shifted by up to 14 days, and provider numbers were encrypted.

The encryption was reversible. University of Melbourne researchers told the ABS on 8 September 2016, and the Commissioner later found patients weren’t reasonably identifiable but providers were. That split, inside one release, is the clearest picture of pseudonymised against de-identified I know of.

I-MED went the other way: the Commissioner was satisfied its patient data had been de-identified sufficiently. That finding sits in a report on preliminary inquiries (31 July 2025), not a determination, and the OAIC says it isn’t an endorsement. Clinics weighing their own data should also read the Ahpra AI guidelines on patient data.

Pseudonymised vs de-identified: one client record three ways

Here is a made-up note a bookkeeper might want help with, treated three ways. The first version names everyone. The second swaps identifiers for labels, which is the step a desktop tool can do. The third is the rewrite only a person who knows the job can do.

The labels took out the name, birth date and address, which is worth having. They didn’t touch the trade or the town, and the town’s only mobile dog groomer is easy to find. The accountants’ page lists the identifiers that turn up most in that kind of file.

The third version keeps the tax question whole. The same discipline applies to a TFN, which comes out of a prompt before anything else.

In a Word file, check the reviewer comments as well, because tracked changes and comments can still carry the client’s name.

Pseudonymised client data in ChatGPT and other AI tools

The OAIC’s guidance on commercially available AI products, published 21 October 2024, has a worked example: insurer staff paste a claim into a public chatbot. The OAIC calls that a disclosure to the chatbot’s owners, which puts APP 6 in play.

Its best practice line is to keep personal information, and particularly sensitive information, out of publicly available generative AI tools. It also asks for the minimum input a task needs. Labels help with the minimum. They don’t settle whether anyone is still reasonably identifiable at the provider’s end.

Tick what’s still in your labelled text. Towns and first names are the easiest to overlook.

Court material sits under an extra layer in NSW, where the Supreme Court’s AI practice note looks at the platform rather than the labels.

De-identified or pseudonymised: wording for a privacy or AI policy

The two words make different promises in writing. “De-identified” commits you to a legal result someone can test. “Pseudonymised” commits you to a method you can show you followed. When the second is the true one, I’d write the second: it’s the promise you can keep.

The same care belongs in the data paragraph of an engagement letter, where a loose word becomes a promise to a client.

In our AI policy template they belong in clause 5, on what may go in and on what condition.

The first line describes a technical step. Practices with their own software often build it on the open-source Presidio library, though the proxy around it still has to be written.

Calling a file de-identified: the check before it leaves

Before anything goes out, ask about the reader: holding only this text, could they work out who it’s about? If the answer is “maybe”, I wouldn’t let the word de-identified anywhere near it.

A tool that puts the real values back after the model replies keeps a key by design, as Limina’s re-identify route shows. Whoever runs it is holding personal information, whatever the output looks like.

Nonimo is built that way too, and doesn’t claim otherwise. On a practice computer in Australia it swaps names and numbers for labels, shows each swap, and restores the originals in the reply you copy.

So the file stays personal information on your side, and the judgement about the story remains yours. Your text never reaches us, as the security page explains.

The plans are on the pricing page.

Sources

• Privacy Act 1988 (Cth), Compilation No. 104. s 6(1) definitions of personal information and de-identified; APP 2.1 and 2.2; APP 11.2 and 11.3; no use of pseudonymised or pseudonymisation.
• OAIC, De-identification and the Privacy Act. No reasonable likelihood test, motivated intruder, custodian example, four parts of the environment, hashing obscures rather than removes.
• OAIC, Report into preliminary inquiries of I-MED. The steps taken, the scale, the outcome and the non-endorsement.
• CSIRO, A framework for data de-identification. The De-identification Decision-Making Framework with the OAIC: audit, risk analysis and control, impact management.
• OVIC, Key concepts. PDP Act s 3 definition; not a permanent state; pseudonymised information isn’t de-identified (paras 1.43 to 1.48).
• ALRC, For Your Information (Report 108), What is not personal information?. National Statement terms; master lists and coded data; identifiability as contextual.

Accessed 7 November 2026.