Social scoring is the evaluation or classification of people over a certain period of time based on their social behaviour or personal characteristics, where the resulting score leads to detrimental or unfavourable treatment. The EU AI Act has prohibited this for AI systems since 2 February 2025, for public authorities and companies alike, in two situations: where the detrimental treatment happens in a social context unrelated to the context in which the data was originally generated or collected, or where that treatment is unjustified or disproportionate to the behaviour or its gravity (Article 5(1)(c) AI Act). Not every score is prohibited. A credit score based on relevant financial data, or a fraud check based on relevant data, in principle falls outside the ban, but is often a high-risk AI system.
Suppose a municipality uses a model to give people on welfare benefits a risk score, based among other things on the neighbourhood they live in, the language they speak and the composition of their household. Those with a high score are the first to have their file reviewed. Is that prohibited social scoring, or ordinary fraud prevention?
In short
The ban applies to anyone who places on the market, puts into service or uses an AI system for social scoring, so to providers and deployers. According to the Commission's guidelines, three conditions must be met at the same time: an AI system, an evaluation or classification of people over a certain period of time based on behaviour or characteristics, and a score that leads or can lead to detrimental or unfavourable treatment in one of the two situations of the ban (points (149) and (150)). Being singled out for additional inspection, without further harm, already counts as unfavourable treatment (point (165)). Lawful scoring for a specific purpose, such as credit assessment, fraud prevention and insurance, is not prohibited as such, as long as the data is relevant and the consequences are justified and proportionate (points (175) to (177)). The Digital Omnibus did not change this prohibition. The maximum fine is 35 million euros or, for an undertaking, 7 percent of worldwide annual turnover if that is higher (Article 99(3)). On top of the AI Act, the GDPR rules on profiling, automated decisions and the DPIA apply.
What Article 5(1)(c) prohibits
In short, the provision prohibits the placing on the market, the putting into service or the use of AI systems for the evaluation or classification of natural persons or groups of persons over a certain period of time based on their social behaviour or known, inferred or predicted personal or personality characteristics, with the social score leading to either or both of the following:
- detrimental or unfavourable treatment in social contexts that are unrelated to the contexts in which the data was originally generated or collected;
- detrimental or unfavourable treatment that is unjustified or disproportionate to their social behaviour or its gravity.
Recital 31 AI Act gives the reason: such systems, used by public or private actors, may lead to discriminatory outcomes and the exclusion of certain groups, and may violate the right to dignity and non-discrimination. The same recital also draws the line: the prohibition "should not affect lawful evaluation practices of natural persons that are carried out for a specific purpose in accordance with Union and national law".
The prohibition has applied since 2 February 2025 (Article 113, third paragraph, point (a) AI Act). The Digital Omnibus, Regulation (EU) 2026/1744, added two new prohibitions to Article 5(1) (points (ba) and (bb), on non-consensual intimate material and child sexual abuse material, from 2 December 2026). Point (c) was left untouched; the consolidated text of 27 July 2026 still shows its original wording.
What the Commission means by a social score
The Commission set out its reading of the ban in its guidelines on prohibited AI practices, published on 4 February 2025; the version now on the Commission's page is C(2025) 5052 final of 29 July 2025. Section 4 covers social scoring. The guidelines are not binding; the authoritative interpretation lies with the Court of Justice. The main points:
- Three conditions at once. An AI system that is placed on the market, put into service or used; an evaluation or classification of persons or groups over a certain period of time based on social behaviour or personal characteristics; and a score that leads or can lead to detrimental or unfavourable treatment in one of the two situations of the ban (points (149) and (150)). The ban applies to providers and deployers, each within their own responsibilities (point (150)).
- Any form of score counts. A number between 0 and 1, a ranking or a label (point (151)). Classification is broader than evaluation and can also be a categorisation without a judgement (point (153)). Profiling under the GDPR is a form of evaluation and may be covered by the ban when it is done with an AI system (point (154)); the Commission refers to the SCHUFA judgment on credit scores.
- Over a certain period of time. A one-time rating on data from one specific situation is in principle not covered, but the Commission wants that condition assessed case by case to prevent circumvention (point (155)).
- Behaviour and characteristics, broadly read. Social behaviour also covers the payment of debts and relations with government, police and the law (point (157)). Personal characteristics include ethnicity, family situation, address, income, household members, profession and level of debt (point (158)).
- The score must cause the treatment. The AI output need not be the only cause, but must play a sufficiently important role, also when a human takes the decision (points (160) and (161)). A score produced by another organisation also counts (point (162)).
- Unfavourable is enough. Unfavourable treatment requires no harm: a person singled out by a score for additional inspection on suspicion of fraud is already treated unfavourably. Detrimental means the person suffers harm (point (165)). A benefit for some, such as priority for housing, can also mean less favourable treatment of others (point (169)).
The two situations of the ban
| Situation | Test | Example from the guidelines |
|---|---|---|
| 1. An unrelated context (point (i)) | The data comes from a context with no connection to the purpose of the score, or leads to generalised surveillance | A social welfare agency estimating the probability of benefit fraud using a spouse's nationality or ethnic origin, having an internet connection or behaviour on social platforms (point (166)) |
| 2. Unjustified or disproportionate (point (ii)) | The consequences are out of proportion to the gravity of the behaviour, or lack a legitimate aim | A municipality scoring the trustworthiness of residents on library books returned late, rubbish left out on the wrong day and late payment of local taxes, and using the score to withdraw benefits (point (167)) |
| Both at once | Unrelated data and disproportionate consequences | A tax authority detecting childcare benefit fraud using criteria such as low income and dual nationality, after which benefits are stopped and reclaimed (point (168)) |
The two situations are alternatives and may also apply in combination; a case-by-case analysis must show that at least one of them is met (point (164)). For the childcare benefit example, the Commission refers in a footnote to the Dutch childcare benefits scandal.
Public and private
The ban applies regardless of whether the AI system or the score is provided or used by a public or a private party (point (170)). The Commission gives two private examples: an insurer using spending and other financial information from a bank to set the premium for life insurance or to refuse a contract, although that information is unrelated to eligibility for that insurance; and a private credit agency using AI to decide on a loan for housing based on unrelated personal characteristics.
In a check by the market surveillance authority, the Commission expects the provider and the deployer to demonstrate that their practice is legitimate and justified: by explaining how the system works, which types of data and sources it uses, whether only data from the context of the score is processed, and whether the consequences are proportionate (point (171)). Anyone using a score must be able to substantiate it.
What is not social scoring
- Scores of organisations. The ban covers natural persons. Scoring legal entities is in principle outside it, unless the score aggregates the behaviour of people and directly affects them, such as all employees of a company or the students of a school (point (173)).
- Ordinary user ratings, such as stars for a driver or a host, as long as no AI analysis combines them with other data (point (174)).
- Lawful scoring for a specific purpose. Credit scoring, risk scoring and underwriting, more efficient claims handling, specific employee evaluations, fraud prevention and detection, law enforcement and scoring users' behaviour on online platforms are not prohibited per se, if lawful and in line with the AI Act and other law (point (175)). That holds as long as the data comes from the context of the score and the consequences are justified and proportionate (point (176)).
- Concrete examples outside the ban (point (177)): a credit score based on income, expenses and other financial circumstances that complies with consumer protection law; a fraud check based on transactional data and other objectively relevant sources; a motor insurer increasing the premium proportionately for risky driving recorded by a telematics device; an online shop giving customers with few returns a faster returns process while others keep the standard process.
- Not an AI system. The ban only covers AI systems within the meaning of Article 3(1) AI Act. Systems based on rules defined solely by natural persons are not covered by that definition (Recital 12 AI Act). The GDPR still applies in full.
Example 1: a municipality and welfare benefits
Back to the municipality from the introduction. The model is trained on past files and gives each benefit recipient a score; that is an AI system. It evaluates people over a period of time on personal characteristics, and a high score leads to a review of their file. That is already unfavourable treatment (point (165)), and if the review ends in a recovery claim, detrimental treatment as well.
Then the data. According to the Commission, a public authority may use data that is relevant for the allocation of the benefit and lawfully collected to determine the risk of fraud (point (166)). Household composition can, on our reading, be such relevant data, because it helps determine the benefit a person is entitled to. The neighbourhood someone lives in and the language someone speaks say nothing in themselves about entitlement to welfare. They also work as proxies for origin or nationality. A model that selects people for review on that basis runs, on our reading, a high risk of falling under situation 1, and because of the disproportionate consequences possibly under situation 2 as well. Even if the municipality only uses relevant data, the consequences must remain justified and proportionate (point (176)). A detailed indicator-by-indicator test is in AI for welfare fraud risk scoring.
What Dutch courts and the Dutch DPA have already said. These rulings and decisions do not concern the AI Act, which did not exist yet. They do show where the line lies under the law that already applied.
- In the SyRI case (ECLI:NL:RBDHA:2020:865, 5 February 2020; English translation ECLI:NL:RBDHA:2020:1878), the District Court of The Hague reviewed the legislation on the System Risk Indication (SyRI) against Article 8(2) ECHR, the right to respect for private life. The court saw a risk that SyRI would unintentionally make connections based on bias, such as a lower socio-economic status or an immigration background, and noted that SyRI had so far only been deployed in so-called problem neighbourhoods (legal ground 6.93). Whether that risk was sufficiently addressed could not be assessed, for lack of verifiable insight into the risk indicators and the risk model (6.94 and 6.95). The court declared the SyRI provisions of the SUWI Act and the SUWI Decree (section 65 and chapter 5a) non-binding because they conflict with Article 8(2) ECHR (7.2). The court expressly left open whether a risk report is an automated decision under the GDPR (6.60).
- On 7 December 2021 the Dutch Data Protection Authority (Autoriteit Persoonsgegevens, AP) fined the Tax Administration 2.75 million euros, among other things because the nationality of childcare benefit applicants served as an indicator in a system that automatically flagged applications as risky. According to the AP, that processing was unlawful and discriminatory.
- On 12 April 2022 it imposed a fine of 3.7 million euros for the Fraud Signalling Facility (FSV), a blacklist with fraud signals on about 270,000 people. The AP found, among other things, that there was no legal basis and that data was often inaccurate. According to the AP, the Tax Administration's own investigation showed that staff were instructed to base the fraud risk partly on nationality and appearance.
High risk and FRIA. If the system falls outside the ban, it is on our reading a high-risk system under Annex III point 5(a) AI Act (evaluating, granting, reducing, revoking or reclaiming public assistance benefits). The deployer obligations then apply from 2 December 2027 (Article 26 AI Act), and the municipality carries out a fundamental rights impact assessment (FRIA) before first use (Article 27 AI Act). If the system was placed on the market or put into service before that date, those obligations only apply if it is significantly changed in its design from that date; providers and deployers of systems intended to be used by public authorities must comply by 2 August 2030 in any case (Article 111(2) AI Act). That transitional rule is expressly without prejudice to Article 5: an existing system does not escape the ban.
Example 2: the bank and the insurer
A bank assessing the creditworthiness of consumers on income, expenses, debts and other financial circumstances falls outside the ban according to the Commission, if that data is relevant for the purpose and the bank complies with consumer protection law (point (177)). The Consumer Credit Directive, Directive (EU) 2023/2225, requires in Art. 18(3) that the creditworthiness assessment is based on relevant and accurate information on income, expenses and other financial and economic circumstances; according to the Commission, that directive also prohibits special categories of personal data and data from social networks in that assessment (point (182)). If the bank or its credit agency also uses unrelated personal characteristics, such as, on our reading, the neighbourhood, nationality or social media behaviour, the ban comes into view (compare point (170)).
If the credit score falls outside the ban, it is a high-risk system: Annex III point 5(b) covers AI used to evaluate the creditworthiness of natural persons or establish their credit score, with the exception of AI used to detect financial fraud. An insurer using AI for risk assessment and pricing in life and health insurance falls under point 5(c). For both, the high-risk obligations apply from 2 December 2027, and private deployers under point 5(b) and (c) must then carry out a FRIA as well (Article 27(1)). According to the Commission, complying with the high-risk requirements can also help prevent such a system from turning into prohibited social scoring (point (172)). More on the credit decision itself in from scoring algorithm to transparent credit decision.
A fraud check based on transactional data, past history and other objectively relevant sources falls outside the ban according to the Commission, as long as any detrimental treatment is justified and proportionate (point (177)). The GDPR and, where relevant, the high-risk rules still apply. An insurer using bank spending data that says nothing about the risk of a life insurance policy is, by contrast, one of the Commission's prohibited examples (point (170)).
Example 3: employers and education, briefly
An employer evaluating staff on their work performance for a specific purpose, such as an appraisal, is not prohibited as such (point (175)). A score that also takes into account behaviour outside work, such as social media activity, and that determines who is promoted or dismissed, may be covered. In one of its examples the Commission expressly names the detriment that arises when job opportunities depend on social media activity (point (168)). Outside the ban, AI for decisions on promotion, termination, task allocation and evaluating performance and behaviour is high-risk (Annex III point 4(b)). See also AI in recruitment and selection.
The same applies to schools: AI for admission and for evaluating learning outcomes is high-risk (Annex III point 3), but a score on behaviour outside school that leads to a disadvantage at school may be covered by the ban. A score for a whole school that directly affects its students can also count (point (173)). Emotion recognition at work and in education has its own ban; see emotion recognition in the workplace.
Boundaries: criminal offences, high risk and the FRIA
Article 5(1)(d) prohibits AI that predicts the risk of a person committing a criminal offence, based solely on profiling or personality traits; it does not apply to AI supporting a human assessment already based on objective and verifiable facts. That ban is narrower and more specific than the ban on social scoring: it only concerns criminal offences (guidelines, footnote to point (146)). Scoring by police and other law enforcement authorities on relevant data falls outside point (c) if it fits within criminal and police law, but must then still be tested against point (d) (point (177)).
High risk. A system that falls outside the ban is often high-risk, and a high-risk system may still be prohibited in a specific case if all conditions of Article 5 are met (guidelines point (37)). An AI system listed in Annex III that profiles natural persons is always high-risk (Article 6(3) AI Act, last subparagraph).
FRIA. For public authorities, the FRIA is where you record in advance which data the system uses, which groups it affects and which measures are in place. The FRIA may refer to parts already covered by the DPIA (Article 27(4)). A FRIA does not make a prohibited practice lawful. A template is in the FRIA template.
What the GDPR adds
Social scoring almost always involves processing personal data, and according to the Commission it may also breach the GDPR, for example on the legal basis, data minimisation, fairness and transparency (guidelines point (180)). Four points.
A score can be an automated decision. Article 22 GDPR gives the right not to be subject to a decision based solely on automated processing, including profiling, which produces legal effects or similarly significantly affects the person. According to the SCHUFA judgment of the Court of Justice (C-634/21, 7 December 2023), the automated establishment of a probability value by a credit information agency is already such a decision where a third party to which the score is transmitted draws strongly on it to establish, implement or terminate a contractual relationship. If the decision rests on an exception in Article 22(2) GDPR, it may not be based on special categories of personal data unless explicit consent or substantial public interest applies and suitable measures are in place (Article 22(4) GDPR).
Right to an explanation. In Dun & Bradstreet Austria (C-203/22, 27 February 2025), the Court held that in automated decision-making a person may, under Article 15(1)(h) GDPR, require the controller to explain in an understandable way the procedure and principles actually applied to obtain a specific result, such as a credit profile. If the organisation considers that trade secrets or data of third parties are involved, it must provide that information to the supervisory authority or the court, which balances the interests.
A DPIA is almost always required. The GDPR requires a prior DPIA where processing is likely to result in a high risk (Article 35 GDPR). The Dutch DPA's list (Staatscourant 2019, 64418) names fraud prevention, for example by social services or insurers (category 3), credit scores (category 4) and profiling, such as the assessment of economic situation, reliability or behaviour (category 15). How to test your own application is in is your AI application on the Dutch DPA's mandatory DPIA list.
Sensitive characteristics. Data on racial or ethnic origin, health, religion or sexual orientation are special categories of personal data that you may in principle not process (Article 9 GDPR). Nationality is not on that list, but the Dutch DPA found its use as a risk indicator discriminatory and therefore unfair. A score that takes such characteristics into account, directly or through a proxy such as neighbourhood or language, also engages non-discrimination law (guidelines point (181)).
Fines and supervision
A prohibited practice carries the highest fine level: up to 35 million euros or, for an undertaking, up to 7 percent of worldwide annual turnover if that is higher (Article 99(3) AI Act). For SMEs and start-ups, the lower of the two applies (paragraph 6). For small mid-cap enterprises, that reduction only applies to the fines in paragraphs 4 and 5, so not to a prohibited practice (paragraph 6a). Member States lay down the rules on penalties (paragraph 1) and decide whether public authorities can be fined (paragraph 8); depending on the legal system, a court may impose the fine (paragraph 9). The penalty provisions have applied since 2 August 2025 (Article 113, third paragraph, point (b)).
In the Netherlands, no market surveillance authority for the AI Act has been designated yet. According to the planning letter of 31 August 2026, the State Secretary intends to submit the AI Act implementation bill (Uitvoeringswet AI-verordening) to the House of Representatives in the second quarter of 2027 (Parliamentary paper 26643, no. 1554). The ban meanwhile applies directly, and the Dutch DPA can already act under the GDPR against unlawful processing of personal data, as its fines on the Tax Administration show. More in AI Act supervision in the Netherlands and fines under the AI Act and the GDPR for municipalities.
What to do now
Municipality or other public authority
- List every model that gives residents a score, ranking or label, also where a staff member takes the decision.
- Record for each indicator how it relates to the purpose of the score and where the data comes from. Remove indicators without a demonstrable link, and look especially critically at neighbourhood, language, nationality and other proxies for origin.
- Carry out or update the DPIA, and prepare the FRIA for high-risk systems.
Bank or insurer
- Check that the credit score only uses financial and economic data permitted by consumer credit rules, and contains no social network data or special categories of personal data.
- Record how you explain the logic of a score to customers (Article 15(1)(h) GDPR) and how human intervention works (Article 22 GDPR).
- Prepare for the high-risk obligations and the FRIA that apply from 2 December 2027.
Employer or educational institution
- Use only data from the work or learning context that fits the purpose of the evaluation.
- Exclude behaviour outside work or school, such as social media, as input for decisions on promotion, dismissal or admission.
Unsure about your own application? With the case check you describe the application in a few questions and see which AI Act and GDPR obligations come into view.
Frequently asked questions about social scoring
Sources
Newsletter
Every Tuesday, the AI Act week ahead in 5 minutes
A practical briefing on deadlines, new guidance and enforcement, so you know what matters this week. No spam and you can unsubscribe in one click.
Practical and short · No spam · One-click unsubscribe