The term “vulnerability” should be understood as a state of being susceptible to harm, manipulation, and exploitation, based on age, disability, or socio-economic status.⁷ The AI Act protects children from manipulative AI toys, or games that encourage risky behaviour that can lead to harm.

In case of the elderly, the focus is on preventing deceptive personalised offers or coercive tactics by AI systems that exploit cognitive capabilities. Such practices can result in financial or psychological harm.⁸ For example, AI systems designed to help financial decision-making must ensure not to target economically vulnerable individuals.

While harmful exploitation is prohibited, the AI act does not prohibit practices resulting in a beneficial effect. Hence, AI systems providing learning tools for children, assistive robots for the elderly or impaired people can be allowed.

Prohibited social scoring practices include unfair evaluation or classification of individuals or groups, based on their social behaviour or personal characteristics over a certain time. The assessment must lead to either or both (i) a detrimental or unfavourable treatment in social contexts that are unrelated to the context in which the data was originally collected, (ii) detrimental or unfavourable treatment of people that is unjustified or disproportional considering their social behaviour.⁹ The prohibition applies both to public and private sectors and targets systems that assign scores, rankings, etc. The main objective of the prohibition is to prevent discriminatory and unfair outcomes, to protect human dignity, privacy, non-discrimination, and others.

This practice involves assessing a risk of the possibility to commit a criminal offence by an individual. This prediction must be based solely on profiling or personal attributes.¹⁰ The objective of this prohibition is to ensure that an individual is not assessed on AI-predicted behaviour. It targets crime prediction systems that use historical data to predict the possibility of committing a crime in future, potentially reinforcing bias and compromising public trust. The prohibition contains a few exclusions: (i) It does not cover petty offences nor administrative offences, (ii) it does not cover predictions based on location, (iii) the risk assessment does not apply to legal entities, (iv) the prohibition does not cover AI systems supporting human assessment based on objective facts about criminal activity.¹¹

Untargeted scraping of facial images means gathering them without any restrictions (such as consent). In 2017, a US company, Clearview AI, developed a software which allowed to identify an individual using a photograph that was extracted from publicly available images on the Internet. As of 2 February 2025, this type of practice is banned with a few exceptions. For example, if the scraping of facial images is focused on individuals belonging to a specific group, this type of activity would be “targeted scraping” and would not be covered by the AI Act. Similarly facial images collected with explicit consent, not commercialised, and used only for research purposes are also exempted.

AI systems that are capable of inferring emotions are prohibited in the workplace, for education, or in training institutions, unless used for medical or safety reasons. The sole detection of human’s expressions, gestures, movements is not considered emotion recognition, unless used to infer or identify emotions, based on biometric data. For example, inferring physical states such as pain, fatigue is not emotion recognition. However, AI systems inferring emotions, such as happiness, from specific movements, (in)voluntary motions,  falls within the scope of prohibition.¹²

AI systems capable of inferring, based on biometric data, a characteristic belonging to a special category of personal data, as per the GDPR¹³,  shall be prohibited. AI systems capable of deducing a person’s religious belief based on their tattoos, or inferring an individual’s race from their tone of voice would be prohibited. However, categorising images of patients according to their skin colour to make a proper cancer diagnosis would be permitted.¹⁴

Deployers are the only group within the scope, since only the “use” of the system is prohibited. The rationale behind the prohibition is to protect individual rights and freedoms from the intrusive nature of such systems. In 2020, a Spanish supermarket chain used a real-time RBI in their CCTV cameras to collect facial images of individuals, to compare their faces against a police database.¹⁵ This case would be an example of a prohibited practice under article 5(1)h. The AI Act introduces three exceptions where the real-time RBI systems may be used, provided that it is strictly necessary. These are (i) search for missing persons, (ii) preventing an imminent threat, (iii) finding and identifying a criminal suspect for the purpose of their prosecution. Member States must regulate such use to ensure appropriate safeguards.¹⁶