Article 5(1)(g) prohibits the placing on the market, the putting into service or the use of AI systems to 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 prohibition does not cover labelling or filtering of lawfully acquired biometric data sets, such as images, based on biometric data, or categorising biometric data in the area of law enforcement, in the circumstances stated in the article. The carve-outs are narrow and must not be read as a general permission for sensitive-attribute inference.
What does biometric categorisation mean here?
Biometric categorisation means assigning a person to a category based on biometric data, in order to deduce or infer an attribute about that person. Biometric data include data from physical, physiological or behavioural characteristics that allow or confirm unique identification, such as facial images, fingerprints, voice patterns and gait. Categorisation differs from identification: identification asks who this person is, while categorisation asks what attribute this person has. A system that sorts people by an inferred attribute, even without naming them, can be categorisation. The record should describe the biometric input, the categories produced, and the inference from data to attribute, because the prohibition is about that inference step.
Which attributes are listed in point (g)?
Point (g) lists race, political opinions, trade union membership, religious or philosophical beliefs, sex life, and sexual orientation. The list is closed in the sense that the prohibition is tied to those attributes, but the list covers the most sensitive inferences a biometric system can make. A system that infers age or sex for a lawful, non-listed purpose follows a different analysis, although data protection law and Annex III can still apply. A system that infers a listed attribute as a by-product, or as an intermediate step toward another output, still needs the point (g) analysis, because the prohibition concerns categorising to deduce or infer the attribute, not only systems whose headline output is the attribute.
For the broader regulatory framework, see prohibited practices under Article 5.
How does the labelling and filtering carve-out work?
The article states that the prohibition does not apply to labelling or filtering of lawfully acquired biometric data sets, such as images, based on biometric data, or to categorising biometric data in the area of law enforcement. Labelling or filtering means organising a data set, for example sorting images by a technical feature, without categorising individual natural persons to infer a listed attribute about them. The data set must be lawfully acquired, and the activity must remain labelling or filtering rather than becoming individual categorisation. A data set operation that produces a profile of a person, or that is used to infer a listed attribute about an identifiable person, moves beyond the carve-out. The record should show the data set source, the labelling operation, and the absence of individual inference.
What about law enforcement categorisation?
Categorising biometric data in the area of law enforcement is treated separately in the carve-out, in the circumstances the article and the applicable law enforcement framework state. That treatment is not a blanket permission for law enforcement to infer sensitive attributes from biometric data in every case. The specific legal basis, purpose, necessity and safeguards for the law enforcement activity must be identified, and data protection law for law enforcement, including Directive (EU) 2016/680, must be checked. A commercial supplier should not assume that a law enforcement customer brings every use within the carve-out. The supplier should record the customer's stated legal basis and purpose, and should seek qualified advice where the basis is unclear.
How does point (g) relate to Annex III and data protection?
Biometric uses that fall outside point (g) can still be high-risk under Annex III, which lists remote biometric identification, biometric categorisation according to sensitive or protected attributes or characteristics, and emotion recognition. Data protection law treats biometric data as a special category with strict conditions, and the GDPR requires an appropriate lawful basis and safeguards. A system cleared under point (g) therefore still needs the Annex III and GDPR analyses. Conversely, a system that is high-risk under Annex III should not be assumed to be outside point (g): the two tests use different wording, and a system can meet the Annex III description while also meeting the point (g) prohibition, in which case prohibition prevails.
Continue with emotion inference at work and in education.
What is not settled by this guide?
This guide explains the prohibition and its carve-outs in outline. It does not decide whether any particular system categorises individuals, whether it infers a listed attribute, or whether a labelling or law enforcement carve-out applies. Those questions need the biometric input, the categories, the inference logic, and the purpose and legal basis evidence, read against the consolidated text. Biometric categorisation touches sensitive attributes and can affect people without their knowledge, so borderline systems should be confirmed with a competent authority or legal adviser before they are placed on the market, put into service or used.
Frequently asked questions
Is sorting photos by face prohibited?
Not necessarily. Labelling or filtering of lawfully acquired biometric data sets, such as images, based on biometric data is carved out, provided the activity remains labelling or filtering and does not categorise individual persons to infer a listed attribute about them.
Does point (g) cover inferring age from a face?
Age is not one of the attributes listed in point (g). Age inference can still engage data protection law, Annex III, and other provisions, and it must be checked separately, especially where the inference affects children or access to age-restricted services.
Can a system infer a listed attribute if the user consents?
Point (g) does not contain a general consent exception. Consent may be relevant under data protection law, but it does not by itself take a system outside the Article 5 prohibition. The wording of point (g) and its carve-outs must be tested first.
Continue through this AI Act cluster
This article is an informational overview, not a legal determination. Check the current legal text and official implementation guidance for your system and role.
Consolidated Regulation (EU) 2024/1689 ↗AI Act Service Desk timeline ↗European Commission regulatory framework for AI ↗