Article 5(1)(g) is a prohibition on inference, not on biometrics. It bars deducing a specific closed list of traits from biometric data, and it contains two express carve-outs that keep large categories of ordinary biometric processing outside it.
The closed list of traits
Article 5(1)(g) of Regulation (EU) 2024/1689 prohibits biometric categorisation systems that categorise individually natural persons based on their biometric data to deduce or infer their race, political opinions, trade union membership, religious or philosophical beliefs, sex life or sexual orientation.
The list is exhaustive and it maps closely, though not exactly, onto the special categories in GDPR Article 9(1). Two differences are worth holding on to. Health data and genetic data appear in GDPR Article 9 but not in this list, so inferring a health condition from a face is not prohibited by point (g) — it is very likely prohibited under GDPR Article 9, and it may be a medical device, but not here. And “ethnic origin” in the GDPR appears as “race” here, which is narrower on its face.
“Categorise individually” is also doing work: the system must assign the individual person to a category. Aggregate statistical reporting that never attaches a label to an identified individual is a weaker fit, though the boundary is untested and a system that computes per-person labels internally has categorised individually whatever it reports.
Not legal advice. This provision sits in the top penalty tier under Article 99(3) — up to EUR 35,000,000 or 7% of worldwide annual turnover — and every system it touches is also processing biometric data under the GDPR, which has its own prohibition and its own regulator. Take advice on the specific system.
What is a biometric categorisation system
Article 3(40) defines it as an AI system for the purpose of assigning natural persons to specific categories on the basis of their biometric data, unless it is ancillary to another commercial service and strictly necessary for objective technical reasons.
That trailing clause is a real exclusion built into the definition rather than into the prohibition, which means a system falling inside it is not a biometric categorisation system at all for the Regulation’s purposes. The paradigm case is image processing incidental to a service the user asked for: a photo application detecting skin tone to apply an exposure correction is categorising on the basis of biometric data, ancillary to the service, and technically necessary to do the thing requested.
The clause is conjunctive. Ancillary and strictly necessary for objective technical reasons. A feature that could be delivered without the categorisation fails the second half, and “it improves our targeting” is not an objective technical reason. How much slack the clause has is not settled; it is a definitional escape hatch that has not been tested and should not be leaned on hard.
Verification is a different thing entirely
The most common confusion on this provision is between categorisation and verification, and the Act keeps them separate at the definitional level.
- Verification — Article 3(36) defines biometric verification as automated, one-to-one verification, including authentication, of the identity of natural persons by comparing their biometric data to previously provided biometric data. Unlocking a device, confirming an enrolled user at a gate, matching a selfie to a document photo. No trait is inferred; a claim of identity is confirmed or rejected.
- Identification — Article 3(35) covers automated recognition of physical, physiological or behavioural characteristics for the purpose of establishing identity by comparing against a database, one-to-many. Regulated under Annex III point 1(a) and, for real-time law enforcement use in public spaces, restricted by Article 5(1)(h).
- Categorisation — assigning to a category. Only this one engages Article 5(1)(g), and only when the category is on the closed list.
Annex III point 1(a) also expressly excludes AI systems intended to be used for biometric verification whose sole purpose is to confirm that a specific natural person is the person they claim to be. So one-to-one verification is neither prohibited nor high-risk under the Act, which is a genuinely permissive position and one worth knowing before redesigning an authentication flow. The GDPR still applies to it in full.
The lawfully acquired dataset carve-out
Article 5(1)(g) states that the prohibition does not cover any labelling or filtering of lawfully acquired biometric datasets, such as images, based on biometric data, or categorising of biometric data in the area of law enforcement.
Two distinct things are excluded and they should not be run together.
Dataset labelling and filtering
The condition is that the dataset was lawfully acquired. Filtering a lawful image corpus by biometric attribute — to balance it, to audit it for representativeness, to build the evidence Article 10(2)(f) demands about possible biases — is expressly outside the prohibition. This is the carve-out that makes bias testing possible, and it interlocks with Article 10(5), which permits exceptional processing of special categories of personal data where strictly necessary for bias detection and correction, subject to a list of safeguards.
Note what “lawfully acquired” excludes: a dataset built by the untargeted scraping prohibited under Article 5(1)(e) is not lawfully acquired, so the carve-out does not reach it.
Law enforcement categorisation
A blunter exclusion, and one of the harder-edged compromises in Article 5. It is not obvious from the text how far it extends — whether it covers any categorisation by any authority in any law enforcement context, or something narrower shaped by Directive (EU) 2016/680 and by the necessity and proportionality requirements of the Charter. It is not yet clear, and it is the kind of question that will reach the Court of Justice eventually.
What remains regulated
Outside the prohibition, biometric categorisation according to sensitive or protected attributes is high-risk under Annex III point 1(b), which is a category that only makes sense once you notice how much point (g) leaves out — inference of age, gender, health or disposition are not on the prohibited list and land here instead. The Chapter III obligations then apply in full.
Article 50(3) separately requires deployers of a biometric categorisation system to inform the natural persons exposed to it of its operation. And GDPR Article 9 operates independently: biometric data processed for the purpose of uniquely identifying a natural person is a special category, and inferred data about beliefs or orientation is special category data in its own right regardless of how it was derived — see GDPR Article 9 applied to biometric AI.
The practical summary is that the Act prohibits a specific inference, regulates a wider class of categorisation as high-risk, and leaves one-to-one verification alone. Which of those three a system is doing is the first question to answer, and it is answered by what the system outputs rather than by what data it consumes. The neighbouring prohibition on inferring emotional states is at Article 5(1)(f).
Top comments (0)