Before sending data for annotation, remove or replace every piece of personal information the task doesn't need: names, contact details, ID and account numbers, faces, voices and anything else that could identify someone. Replace them with consistent tags such as [NAME] or realistic stand-ins, so the data stays useful for labeling. Then check what's left for combinations of details that could still point to one person.
Done well, this protects the people in your data, reduces your legal exposure under the GDPR and makes working with an outside labeling team much simpler.
This article explains common practice in plain words. It is not legal advice; for a specific dataset, consult a data protection specialist.
Anonymised or Pseudonymised: Why the Difference Matters
The two words sound similar but have very different legal consequences.
| Pseudonymised | Anonymised | |
|---|---|---|
| Definition | Data that can't be linked to a person without extra information, which is kept separately and protected (GDPR Article 4(5)) | Data from which nobody can identify a person, using all the means reasonably likely to be used (Recital 26) |
| Example | "Maria Ingram" replaced by "Customer 4821", with the key stored elsewhere | All names, IDs and rare details removed, with no key anywhere |
| Still personal data? | Yes | No |
| GDPR applies? | Yes, with lower risk | No |
The European Data Protection Board confirmed this in its Guidelines 01/2025 on pseudonymisation, adopted in January 2025: pseudonymised data remains personal data, but pseudonymisation is a valuable safeguard, provided the information that could re-identify people is kept separate and protected.
In practice, most annotation projects end up pseudonymised, not fully anonymised, because some context must stay for the task to make sense. That is fine and still worthwhile: it lowers risk a lot. Just don't call the data "anonymous" in contracts or privacy notices unless it really is.
Techniques by Data Type
| Data type | What to look for | Common techniques |
|---|---|---|
| Text (emails, chats, tickets) | Names, emails, phone numbers, addresses, account and ID numbers, dates of birth, free-text personal details | Replace with tags or realistic stand-ins; remove signatures and footers; generalise exact dates |
| Documents (invoices, contracts, forms) | All of the above, plus signatures, stamps, photos, bank details | Black out or replace fields not needed; keep layout intact so the task still works |
| Images and video | Faces, licence plates, house numbers, screens, name badges, reflections | Blur or mask; crop out irrelevant areas; remove location data in file metadata |
| Audio | Voices, names and numbers spoken aloud | Bleep or cut segments; in transcripts, replace spoken details with tags; voice changing where the task allows |
| Any file | Hidden metadata: author, GPS location, device, creation date | Strip metadata before sharing |
A before-and-after example for a support email:
Before: "Hi, this is Maria Ingram (born 05/12/1987). My card ending 4417 was charged twice on 03.03.2026. Call me on 0722 123 456. Michigan Avenue 10, Chicago."
After: "Hi, this is [NAME] (born [DATE_OF_BIRTH]). My card ending [CARD_DIGITS] was charged twice on [DATE]. Call me on [PHONE]. [ADDRESS]."
The annotator can still label this email "Billing dispute", which is all the task needs.
Keeping the Data Useful
Anonymisation that destroys what the model needs to learn is wasted effort. Three choices keep data useful:
1. Tags or stand-ins?
| Approach | Example | Best when |
|---|---|---|
| Type tags | "[NAME] paid [AMOUNT]" | The task doesn't depend on the details themselves |
| Realistic stand-ins (surrogates) | "Ellen Marshall paid $450" with a made-up name | The model must learn to handle real-looking text, such as for entity recognition |
2. Be consistent within a record. If "Maria" appears five times in a conversation, replace her with the same tag or stand-in each time ([NAME_1]). Otherwise the annotator and the model, lose track of who is who.
3. Keep what the task needs. If the model must find names, you can't remove all names; use stand-ins instead. If it must understand dates relative to each other, shift all dates in a record by the same amount instead of deleting them.
When personal data must stay (because finding it is the task), limit access to a small, trained team, block downloads and make sure your contract and the GDPR basis cover it. See GDPR and AI Training Data.
A Step-By-Step Process
- List what the task needs. Which details must the annotator see to do the job? Everything else is a candidate for removal.
- Find the personal data. Use automatic detection tools for common patterns (emails, phone numbers, IBANs, faces), then check a sample by hand, because tools miss things.
- Choose a technique per field: remove, tag, replace with a stand-in, blur or generalise.
- Apply it consistently across the dataset and within each record.
- Keep the key separate, if you pseudonymise: the table linking stand-ins to real people stays with you, never with the vendor.
- Check for indirect identifiers. A rare job title, a small town and an exact date together can point to one person even without a name.
- Review a sample by hand before sending anything.
- Document what you did, for your records and your GDPR accountability.
Common mistakes:
| Mistake | Why it matters |
|---|---|
| Relying only on automatic tools | They miss unusual names, misspellings and details in free text |
| Forgetting metadata | Photos and documents can carry location, author and device data |
| Removing names but keeping unique combinations | People can be re-identified from context |
| Inconsistent replacement | Annotators and models lose track of who is who |
| Sending the re-identification key with the data | The data is then effectively not pseudonymised at all |
| Calling pseudonymised data "anonymous" | It is still personal data, with GDPR duties |
Your Data, Handled Carefully
AnnoTrue signs an NDA and a data processing agreement before seeing any data, stores it in the EU and our annotators can't download it. We can also advise on what to remove before you send it. Start with a free pilot.
Sources
- GDPR Article 4: Definitions
- GDPR Recital 26: Not applicable to anonymous data
- Hunton Andrews Kurth, EDPB advises on pseudonymisation for GDPR compliance (on EDPB Guidelines 01/2025)