Anonymising company data for AI training: our standard for removing personal data

Where a buyer removes personal data after transfer, we tell you in writing first and your own solicitor confirms the transfer route before anything moves. This page explains the law and publishes the standard we work to. It is not legal advice: take your own before licensing any records.

Is anonymised data still personal data under UK GDPR?

No, provided the anonymisation is effective: data protection law does not apply to anonymous information. But the test is practical and strict. People must not be identifiable by any means reasonably likely to be used, by you or by anyone who might obtain the data. Anything short of that is still personal data^2,3.

The ICO's anonymisation guidance, published on 28 March 2025 and now under review after the Data (Use and Access) Act 2025, treats identifiability as a spectrum^1,3. At one end, information identifies people directly; at the other, nobody handling it can reliably link it to a person. In between, the answer depends on who receives the data, what else they hold, what is public and what technology can do. The law does not demand zero risk, but the risk must be "sufficiently remote", and the judgement must be kept under review as technology changes³.

Two consequences follow for licensing. Anonymising is itself processing, so you need a lawful basis, a defined purpose and transparency with the people concerned². And the same dataset can be personal data in your hands, because you keep the originals, yet anonymous in a buyer's hands if the buyer has no reasonable means to identify anyone^3,4.

A note on words. The ICO does not encourage "de-identified" as a synonym for anonymous, because UK law uses the term only in the re-identification offence². We use de-identification for the work of removing personal data, and call a dataset anonymised only when the ICO's test is met. The regulator itself is now the Information Commission, which took over the Information Commissioner's functions on 30 September 2026 and is still known as the ICO, the Information Commission's Office^8,16,17.

Why does pseudonymised data stay personal data?

Because the people can still be identified by anyone holding the key or enough other information. Replacing every name with a consistent code, such as Employee_17, is pseudonymisation. The ICO says pseudonymous data is still personal data, and warns that datasets described as anonymised are often only pseudonymised².

Consistent codes are common in this market because they keep conversations readable: a buyer can follow Employee_17 through a thread. Labelbox Prism, for example, says names, emails and internal IDs are "tokenized before egress"²². That protects people far better than raw data, but it does not end the analysis.

For you, the data stays personal data, because you hold the originals that link every code back to a person⁴. For a buyer that never receives the key, the ICO says pseudonymised data "may be anonymous information in their hands"⁴, and the EU's highest court took a similar, recipient-specific view in EDPS v SRB¹³. Whether that holds depends on what is left in the text.

Free text is the hard part. Emails and chat mention people by role ("the new finance director"), by event ("back from paternity leave") and by details no filter recognises as a name. A lawyer quoted by MIT Technology Review points to the doctor appointments, school pickups and date-night plans that sit in ordinary emails and calendars²³. Every name can be replaced and a colleague could still tell who is who.

Why is linked business data hard to anonymise?

Because the links that make records valuable also identify people. Buyers want a ticket joined to its email thread, its CRM record and its outcome. Each extra link makes it easier to single someone out, combine sources or infer new facts about them, which are the three routes to identification the ICO describes³.

• Singling out: isolating the records that relate to one person, such as the only engineer based in one depot.
• Linkability: combining records across systems or with public sources, the "mosaic effect". The ICO says removing direct identifiers alone is not enough³.
• Inference: deducing something new about an identifiable person, such as who was about to resign.

The proposed Spirit Airlines sale shows the tension. Its dataset was built to keep "referential integrity", the links between records across systems, and the flight attendants' union argued those links could let employees be identified by pattern even with names removed¹³.

Workplace records also come with unusually well-informed intruders. The ICO notes that someone with prior knowledge, such as an employee who knows a colleague is on long-term sick leave, may recognise people a stranger could not³. For company records that means former staff, customers, suppliers and competitors.

If your records include people in the EU, or you have an EU office, ask your solicitor whether the EU GDPR applies too. The European Data Protection Board's draft Guidelines 02/2026, adopted for public consultation on 7 July 2026, test anonymity with three criteria (no record isolation, no linkage, no inference) and say record-level data with many attributes is the most exposed¹⁴.

Which records should be excluded completely?

Some records should stay out however carefully they could be processed. Under our standard, HR and payroll files, health and other special category data, and material you hold in confidence for clients stay out entirely. We also advise leaving out grievances and disciplinary matters, criminal records checks, private direct messages, credentials and worker monitoring data. You confirm the final list of exclusions with your solicitor.

These overlap closely with the categories unions won out of the proposed Spirit sale: medical examinations, payroll, drug testing, crew pairings, biometrics and union grievance filings¹⁵. Some sectors need more: recruitment agencies hold candidate data at scale (see staffing and recruitment), and every sector's limits are on the sectors page.

Where does de-identification happen, and does raw data leave the UK?

We prefer de-identification before records leave your systems: inside your own environment, or by a UK processor under contract to you. micro1 and Mercor remove personal data after transfer^18,20. Where a buyer does that, we tell you in writing first and your solicitor confirms the transfer route before anything moves. Personal data sent abroad is a restricted transfer⁷.

Buyers take different routes, so we set out each one in writing before you choose. What buyers say on their own pages (checked 12 November 2026):

When personal data goes to a separate organisation outside the UK, the transfer must be covered by UK adequacy regulations, appropriate safeguards such as the international data transfer agreement or addendum with a transfer risk assessment, or a narrow exception⁷. For the US, the UK's adequacy regulations came into force on 12 October 2023 and cover only organisations listed on the Data Privacy Framework List as participating in the UK Extension¹⁰. The Data (Use and Access) Act 2025 carried them over: regulations of that kind in force before its transfer rules began are treated as made under the new Article 45A of the UK GDPR²⁴. Ask any US buyer that would receive personal data whether it is on that list. De-identifying in the UK first keeps personal data at home and narrows the legal questions.

Who carries out the work, where, and under what contract is set out in writing for each buyer's route before you agree to it. See where this sits in the full process.

How is the result tested and documented?

Under our standard, each package is tested before release with a motivated intruder test: could a determined person, using public information and some inside knowledge, identify anyone? The method, results and remaining risks go into a data protection impact assessment (DPIA) and a release record that you and your solicitor approve.

1. Scope. List the systems, date ranges and whose data they hold: staff, customers, suppliers, the public.
2. Exclude. Remove the categories above before any other processing.
3. Remove direct identifiers from fields, free text and attachments: names, email addresses, phone numbers, addresses, account and reference numbers, signatures, photos and authorship metadata.
4. Reduce indirect identifiers. Generalise dates, places and job titles where they single people out; drop rare events and very small groups.
5. Review. Combine automated detection with human review of samples, because free text defeats pattern matching.
6. Test. Run a motivated intruder test that includes a well-informed insider's view. The ICO's examples of intruders include investigative journalists, estranged partners and industrial spies³.
7. Document. The ICO says you must document key decisions and their reasons, and recommends a DPIA to structure them; it is less likely to take enforcement action where an organisation shows a serious effort to comply and a genuine reason to believe the data was not personal data⁵.
8. Sign off. A named senior person at your company approves each package, with your data protection officer if you have one⁵. Code-to-name mapping tables never go to the buyer.
9. Re-test when the data, the buyer or the technology changes, because identification risk rises over time^3,14.

What must the buyer promise in the licence?

The licence must forbid re-identification and linking with other data, limit use to agreed purposes, bar onward transfer to anyone not bound by the same terms, require security, breach notice and deletion of the delivered data, give you audit rights, and oblige the buyer to tell you at once if it finds identifiable data.

These promises help, but they do not make data anonymous on their own. The EDPB's draft guidelines say contractual terms should only complement technical measures¹⁴, a point William Fry draws out in its analysis of the Spirit sale¹³. UK law adds teeth: it is an offence to knowingly or recklessly re-identify de-identified personal data without the consent of the controller responsible for de-identifying it¹². The ICO also says an organisation that can re-identify data it received as anonymous must treat it as personal data and tell the supplier³. Under the revised terms of the proposed Spirit sale, which still awaits court approval, Google must publicly commit not to knowingly re-identify the data¹⁵.

Be clear about deletion. Labelbox Prism's own FAQ says models already trained on licensed data "cannot be unlearned"²², so deletion usually covers delivered data and future use, not models already built. See the licence terms to check.

What does the law still require of you?

Your company remains the controller of its records, so the legal decisions stay with you: a lawful basis for the de-identification work, a check that licensing fits the purposes you collected the data for, telling staff and customers, and a DPIA. Good de-identification reduces risk; it does not remove these duties.

• Lawful basis and purpose. Anonymising is processing that needs a lawful basis and a clear purpose². Since 5 February 2026, Article 8A of the UK GDPR lists what decides whether a new purpose is compatible with the original: the link between them, the context and relationship with the people, the nature of the data, the possible consequences and safeguards such as pseudonymisation^8,9.
• Transparency. The ICO's data sharing code says to consider when and how to inform people when their data moves to a new controller⁶. For staff, see licensing Slack, Teams and email archives.
• Insolvency. An administrator or liquidator may act as controller "for the time being"⁶; see the office-holder guide.
• Moving guidance. The ICO's guidance is under review¹ and the EDPB's is a consultation draft¹⁴. We will update this standard as they change.

Read our guide to UK law on licensing company records, and the glossary for terms. If you first want to know whether your records are worth this work, see what your company data could be worth; how we are paid explains why sellers pay us nothing.

Talk to us about your records before anything is exported. On a 30-minute call we go through this standard against your systems, and before you agree to any buyer we set out in writing where personal data would be removed and by whom.