Social scoring: Article 5(1)(c) and AI scores that punish people in unrelated contexts

What is social scoring under point (c)?

Social scoring, as described in point (c), involves evaluating or classifying natural persons or groups over a certain period of time, based on social behaviour or on known, inferred or predicted personal or personality characteristics. The evaluation is longitudinal: it builds a picture over time, rather than deciding a single transaction on its own facts. The data can include observed behaviour, inferred traits and predicted characteristics, and the score can be a number, a category, a ranking or a label. A record should describe what is evaluated, over what period, from what data, and what classification results, because each of those facts feeds the later conditions. A one-off eligibility check on its own facts is not the same pattern, although it can still be high-risk or subject to other duties.

What is the unrelated-context condition?

The first alternative condition is that the social score leads to detrimental or unfavourable treatment in social contexts that are unrelated to the contexts in which the data was originally generated or collected. The concern is transfer: behaviour in one sphere of life is used to punish or disadvantage a person in another sphere. A score built from community behaviour, online activity or everyday conduct, and then used to deny housing, employment or services, illustrates the pattern the condition describes. The record should map the origin context of each data source and the context in which the score is used, and ask whether the two are related in a way the person could reasonably have understood when the data were generated.

For the broader regulatory framework, see prohibited practices under Article 5.

What is the disproportionate-treatment condition?

The second alternative condition is that the social score leads to detrimental or unfavourable treatment that is unjustified or disproportionate to the social behaviour or its gravity. Even where the context is related, a punishment that far exceeds the behaviour it responds to can meet this condition. A minor infraction that leads to a severe, long-lasting exclusion, or a single behaviour that leads to exclusion from many unrelated benefits, needs examination on proportionality. The assessment should consider the gravity of the behaviour, the severity and duration of the treatment, the number of contexts affected, and whether the person has a way to understand, challenge or correct the score. Only one of the two conditions needs to be met, so a system should be screened against both.

How does social scoring differ from credit scoring?

Credit scoring and similar eligibility assessments are not automatically social scoring under point (c). A score used to assess creditworthiness, insurance risk or eligibility for a specific service, based on data relevant to that decision and applied proportionately within that context, follows a different pattern from a general social score transferred across contexts. That distinction does not make every credit or insurance score lawful under the Act: Annex III includes specific use cases for credit scoring and for life and health insurance pricing and assessment, which can be high-risk, and data protection law, consumer law and anti-discrimination law can apply. The screening record should therefore distinguish the Article 5 question from the Article 6 question, and answer each separately.

What should a scoring system record show?

A scoring record should name the data sources and their origin contexts, the period over which behaviour is evaluated, the characteristics inferred or predicted, the score or classification produced, the decisions it informs, and the contexts of those decisions. It should state the proportionality reasoning: why the treatment follows from the behaviour, why it is limited to the relevant context, and how long the score affects the person. It should also record the person's ability to know the score exists, to understand the factors behind it, and to seek correction. Where a score is supplied by a third party, the deployer should obtain the same information from the supplier, because the deployer's use of the score is itself within the prohibition.

Continue with exploitation of vulnerability.

What is not decided by this guide?

This guide explains the two conditions in point (c) and the distinction from context-specific scoring. It does not decide whether any particular scoring system is social scoring, whether its use is in an unrelated context, or whether its treatment is disproportionate. Those questions need the data map, the score design, the decisions affected, and the proportionality analysis for the real system. Scoring systems that affect access to work, housing, services or benefits should be confirmed with a competent authority or legal adviser before use, because the consequence of a match is prohibition, not a managed high-risk regime.

Frequently asked questions

Continue through this AI Act cluster

• prohibited practices under Article 5
• exploitation of vulnerability
• predictive policing prohibition
• the risk-based approach
• high-risk AI systems
• Understand AI Act risk categories