obscurGDPR and AI
GDPR and artificial intelligence: the measures a controller can prove
Published on
In short: the GDPR does not forbid using artificial intelligence on documents that contain personal data. It asks the data controller to take appropriate measures and to be able to demonstrate it (Art. 5(2) and Art. 24). It does not ask for perfection: it asks for evidence. This page lists those measures article by article, and reports how well automatic recognition of personal data actually works, as measured.
Why “demonstrate” matters more than “protect”
In an inspection, or when a client complains, the question is not whether the firm did its best. It is what it can show. A default setting that protects without anyone having to remember, an activity log, a sheet ready for the impact assessment: these are what hold up. A vendor's promise, on its own, does not.
The measures, article by article
| Article | What it asks | The measure, as obscurIA applies it |
|---|---|---|
| Art. 4(5) | Pseudonymise: the information that links back to the person kept separate and protected | Placeholders instead of data; mapping encrypted on the device only; restoring requires a secret token |
| Art. 5(1)(c) · 25(1) | Data minimisation and protection by design | Only protected text reaches the AI; the document never leaves the computer; the final check runs again exactly where text leaves |
| Art. 25(2) | Protection by default | Every text is reviewed and confirmed before it goes to the AI or is exported, without anyone having to remember |
| Art. 5(1)(e) | Storage limitation | Expiry for saved documents (7, 30, 90, 365 days or never), with deletion and database compaction |
| Art. 32 | Security appropriate to the risk | Encrypted local storage, keys in the system keychain and wiped from memory, credentials recognised and covered |
| Arts. 9 and 10 | Special categories and criminal data | Dedicated recognition of health, beliefs, politics, trade union, sex life, genetic and biometric data, convictions; a warning if the text still talks about them |
| Art. 5(2) · 24 | Accountability: being able to prove it | Tamper-evident activity log with no content; a per-document attestation with counts only |
| Art. 35 | Impact assessment | A DPO sheet with data flows, data, measures and measured residual risk |
How reliable is automatic recognition? We measured it
We ran the recognition engine on four collections of documents written by third parties who did not know how it works: minutes, medical records, contracts, bank statements, chats, configuration files, forms in English. In total 523 documents and 4,887 personal data items to find.
| Collection | Documents | Data items | Recognised by the engine |
|---|---|---|---|
| 1 | 104 | 1,000 | 82% |
| 2 | 122 | 1,380 | 86% |
| 3 | 146 | 1,176 | 89% |
| 4 | 151 | 1,331 | 88% |
Codes with a check digit, such as the Italian tax code, and API keys never slipped through. What does slip through is mostly first or last names mentioned on their own, dates, foreign documents and plates, health or religious phrases written in an unusual way. The conclusion is simple: no automatic tool can honestly promise “100% anonymous” data, and being wary of anyone who does is already a prudent measure.
Why a human review is needed
If the machine recognises about nine data items out of ten, the measure that covers the tenth is a person re-reading the protected text before it leaves. That is what protection by default means (Art. 25(2)): in obscurIA every text is reviewed and confirmed before it goes to the AI or is exported, and the confirmation is recorded in the activity log. The controller can limit the review to the most sensitive texts: that is their choice, and it is recorded too.
What to prepare for the impact assessment (DPIA)
- The flows: what stays on the computer, what reaches the AI provider, what reaches the software vendor.
- The categories of data processed, including the special categories of Articles 9 and 10.
- The AI provider chosen, the legal basis and the processing agreement with them.
- The measures in the table above, with the measured residual risk and the review that covers it.
- Who can restore the real data, and how long documents are kept.
For these points obscurIA provides the DPO sheet and, for each document, a signed attestation that reports only counts per category, never values. Technical details are on the Security and privacy page.
What pseudonymisation does not do
Pseudonymised text is still personal data under the GDPR (Recital 26), because the mapping exists; and context can make a person recognisable even without a name. The difference between the two operations is explained in Anonymous or pseudonymised data?. Finally, a tool protects what goes through it: what people type directly on an AI service's website must be governed by an internal policy.
Informational article, not legal advice. The measures described here help the controller support their own compliance; the assessment of the specific case remains theirs and their DPO's.