Exploiting Vulnerability: Age, Disability and Social or Economic Situation
The second prohibited category targets AI systems that exploit the particular vulnerabilities of specific groups of people. A vulnerable person is someone who, by reason of age, disability, or social or economic situation, is in a weaker position than others to protect their own interests. An AI system exploits that vulnerability if it deliberately manipulates or distorts the person’s behaviour in a way that causes them material or immaterial harm.
The age protection recognises that children and young people often lack the experience and maturity to evaluate certain kinds of messaging, persuasion and social pressure. Minors cannot always recognise when they are being manipulated or exploited. They lack developed decision making strategies that adults use to filter persuasion. They are more susceptible to peer pressure, social proof and appeals to social belonging. An AI system that targets children with personalised manipulation is doubly prohibited: once as manipulation under the first category, and again as exploitation of age-related vulnerability. A social media recommendation algorithm that learns to serve content designed to keep children scrolling is an example of AI that might trigger concerns.
Disability covers both physical and cognitive conditions. This includes mental health conditions, learning disabilities, acquired brain injuries, sensory impairments and mobility impairments. The risk with AI systems is twofold. First, AI systems designed for the general population may fail for people with disabilities, creating barriers. Second, AI systems designed with specific knowledge of a disability might intentionally exploit that disability. An example would be an AI voice assistant that uses patterns of speech known to disorient people with certain cognitive conditions, or a system that learns that people with certain disabilities are more susceptible to particular persuasion techniques.
Social or economic vulnerability includes a broader range of situations. It covers homelessness, poverty, precarious employment, recently arrived migrants with limited knowledge of local systems, elderly people isolated from family support, and other situations where a person has fewer resources or less power to negotiate or protect themselves. An AI system that exploits economic vulnerability might target financially desperate people with predatory lending offers, using an algorithm that identifies signs of financial stress and triggers consumption decisions when the person is most desperate.
The prohibited practice is not merely to target vulnerable groups for marketing or services. It is to deliberately exploit their vulnerability in a way that causes harm. This means the organisation must have knowledge or reasonable foreseeability that the system would cause harm to vulnerable people. Reasonable foreseeability means that the harm should be predictable by someone with appropriate expertise in the domain. It means the harm must be material or immaterial damage to the person’s interests, not trivial inconvenience. It is not enough that a vulnerable person is merely exposed to the system or receives services. There must be causal connection between the deliberate exploitation and the harm caused.
There is no exception to this prohibition for any sector or purpose. Even government services, charities serving vulnerable populations, and healthcare providers cannot use AI systems that exploit vulnerability, except for the narrow law enforcement exception that applies to all Article 5 bans. This is a strong statement of principle: protecting vulnerable people takes priority over other policy objectives.
For an organisation to demonstrate it has not breached this prohibition, it must show that it has identified any vulnerable populations in its user or customer base, conducted an assessment of whether its AI system could exploit their vulnerability, and implemented controls to prevent that harm. This might include design changes to make the system more accessible, monitoring systems to catch signs of exploitation early, safety features that prevent decisions when vulnerability is detected, or accessibility features that protect vulnerable users. The burden is on the organisation to prove it has taken these steps.
Examples of prohibited exploitation of vulnerability include: a credit card AI system designed to identify consumers in financial distress and automatically increase their credit limit to trigger spending when they are most financially vulnerable; an employment assessment AI system that identifies candidates with learning disabilities from employment history patterns and deprioritises them in hiring despite legal protections; an AI system in a benefits administration application that deliberately times complex enrolment procedures for times when elderly users are known to be fatigued; a game or gambling application that uses AI to identify children or young people and personalises content to increase engagement in ways that are age-inappropriate.