AiAmigo logo mark

GDPR and AI data protection

Anonymisation vs Pseudonymisation: GDPR Guide

Updated

Replacing a name is useful, but it rarely makes a prompt anonymous. Use this practical guide to choose the right protection and describe it accurately.

Anonymisation and pseudonymisation both reduce privacy risk, but they have different legal consequences. This matters when employees paste customer emails, HR cases, contracts or support tickets into generative AI: masking obvious identifiers can make the text safer to use without necessarily taking it outside the GDPR.

The short answer: anonymised information does not relate to an identified or identifiable person and falls outside the GDPR. Pseudonymised data cannot be attributed to a person without additional information that is kept separately and protected, but it is normally still personal data for the controller that can reconnect it. Pseudonymisation is therefore a safeguard, not a way to avoid GDPR duties.

Conceptual comparison — purpose: anonymisation is intended to prevent identification for the relevant parties; pseudonymisation is intended to prevent attribution inside a defined domain unless authorised additional information is used. Reversibility: genuine anonymisation must withstand means reasonably likely to be used; pseudonymisation may deliberately preserve a controlled route back to the person. Legal status: anonymous information is outside the GDPR; pseudonymised personal data remains subject to lawful basis, transparency, minimisation, retention, security and data-subject-right requirements where the person remains identifiable.

A practical prompt example shows the gap. “Summarise María López’s complaint about her rare condition at the only clinic in town” is plainly personal data. Changing the name to “[CUSTOMER_01]” is pseudonymisation if your CRM or a colleague can reconnect the token. Even without a lookup table, the rare condition, clinic and detailed circumstances may single María out. A safer working version removes unnecessary details, generalises the location and condition, and asks only for the analysis needed — but whether the result is anonymous still requires a contextual assessment.

The same applies to documents. An HR spreadsheet where “Anna Jensen” becomes “EMP-4827” remains pseudonymised if HR holds the employee-number mapping. A contract with names masked may still identify a party through an address, case number, signature, metadata or a distinctive sequence of events. A support ticket can remain personal data through an email quoted in an attachment or through enough dates, product details and location clues to identify one customer.

The EDPB Guidelines 02/2026 propose a practical three-part test: No Record Isolation, No Linkage and No Inference. As at 23 August 2026, these guidelines were adopted for public consultation and feedback remained open until 30 October 2026; they were not final guidance. They are nevertheless a useful current framework, but organisations should monitor the final text and relevant case law.

“No record isolation” asks whether a unique combination of attributes isolates one person’s record — the practical singling-out risk. “No linkage” asks whether records can be connected with other data about the same person, including public or internal sources. “No inference” asks whether specific and meaningful information about an identifiable person can be deduced from the data and available context. Passing all three supports an anonymity conclusion; failing one calls for further analysis rather than an automatic conclusion.

Context matters. The 2026 EDPB draft reflects recent EU case law by assessing what means are reasonably likely to be used from each relevant entity’s perspective. A dataset may be anonymous for an independent recipient that cannot lawfully or practically obtain the key, while remaining personal data for the organisation that holds the source records. If a service provider processes the material on the controller’s instructions, the controller’s perspective still matters; outsourcing is not a shortcut around the GDPR.

Anonymisation also has limits. Free text is high-dimensional and can contain rare facts, writing style, chronology and hidden identifiers. Hashing predictable identifiers without a strong secret can be reversed by guessing. Synthetic or aggregated outputs can leak membership or enable inference. Re-identification capabilities and available external data change, so risk assessments, testing and documentation should be reviewed over time. Anonymising personal data is itself processing and needs a lawful, secure and transparent design.

Apply the 2026 EDPB test before calling data anonymous

  • No record isolation / singling out: can a unique job title, exact date, postcode or combination of facts isolate one person, even without a name? Source
  • No linkage: can the prompt or document be matched to CRM records, public profiles, another dataset, a token mapping or information held by a third party? Source
  • No inference: could a reader or AI system deduce a specific and meaningful fact — such as a diagnosis, salary or complaint — about an identifiable person? Source

Checklist before sending personal information to an AI tool

  • Define the minimum purpose: ask for only the transformation, summary or analysis you actually need.
  • Map direct and indirect identifiers across the prompt, document body, tables, filenames, metadata and attachments.
  • Remove or mask names, email addresses, phone numbers, account numbers and other direct identifiers before sending.
  • Generalise exact dates, granular locations, rare roles and unusual events where precision is not necessary.
  • Decide whether consistent tokens such as [CUSTOMER_01] are needed; if they are linkable, classify the material as pseudonymised personal data.
  • Keep any lookup table, original document or secret separately, restrict access and protect it with technical and organisational controls.
  • Test record isolation, linkage and inference against internal knowledge, public sources and plausible future recipients.
  • Check the AI provider, plan, retention, training, subprocessors, transfer arrangements and contractual terms; masking does not replace vendor due diligence.
  • Record the assessment, responsible owner, intended recipients and review date, and reassess when context, data or technology changes.
  • When uncertainty remains, treat the content as personal data and involve your data protection officer or privacy adviser.

When does masked content remain personal data?

Assume it remains personal data when your organisation retains the original or token key, when a person can be identified from context, when the same token enables records to be linked, or when a recipient can reasonably obtain additional information. Removing names alone is de-identification, not proof of anonymity. The conclusion depends on the data, purpose, recipients, available means and current technology — and should be documented rather than inferred from the name of a software feature.

Use AIamigo as a pre-send protection layer

AIamigo detects sensitive information and lets users mask it before content is sent to supported AI tools. This can reduce unnecessary exposure and support data minimisation in everyday prompts. It does not by itself prove that every masked prompt or AI output is anonymous: remaining context, consistent replacement tokens, source documents and information held elsewhere may still make a person identifiable. Treat the result according to your documented assessment and wider GDPR controls.

Related resources

Frequently asked questions

What is the difference between anonymisation and pseudonymisation?

Anonymisation makes people no longer identified or identifiable for the relevant parties using means reasonably likely to be used. Pseudonymisation prevents attribution without separately protected additional information, but normally leaves the information as personal data for a controller that can reconnect it.

Is pseudonymised data still personal data under GDPR?

Usually, yes — particularly for the controller holding the original data, key or other means of attribution. A recipient-specific assessment can be more nuanced, but replacing a name with a code does not automatically remove GDPR obligations.

Does removing names make an AI prompt anonymous?

Not necessarily. Dates, locations, rare events, roles, quoted text and other details may single someone out, link the prompt to another source or reveal meaningful information about an identifiable person.

What are the EDPB anonymisation criteria in 2026?

The consultation version of Guidelines 02/2026 uses No Record Isolation, No Linkage and No Inference. Passing all three supports anonymity; if a criterion fails, the EDPB calls for further analysis. As at 23 August 2026, the guidelines were not final.

Can AIamigo guarantee GDPR anonymisation?

No tool can make that legal conclusion from identifier masking alone. AIamigo detects and masks sensitive content before it is sent, reducing exposure, but context or information held elsewhere may still make the content personal data.

Official guidance and further reading