The EU AI Act's Ban on Exploiting Vulnerable Groups

Article 5(1)(b) is routinely described as the second half of the manipulation ban. It is a separate prohibition with a separate and shorter test, and reading it as an appendage to Article 5(1)(a) makes it look narrower than it is.

The test, element by element

Article 5(1)(b) of Regulation (EU) 2024/1689 prohibits an AI system that exploits any of the vulnerabilities of a natural person or a specific group of persons due to their age, disability, or a specific social or economic situation, with the objective or the effect of materially distorting the behaviour of that person or of a person belonging to that group, in a manner that causes or is reasonably likely to cause that person or another person significant harm.

Four elements: a vulnerability on one of three listed grounds; exploitation of it; objective or effect of materially distorting behaviour; significant harm, actual or reasonably likely. As with Article 5(1)(a), they are cumulative, and as with (a), effect is sufficient — a provider who did not intend to exploit anything is not outside the provision by virtue of that.

The three things (b) does not require

The side-by-side is the useful part, because a summary that merges the two provisions silently imports elements from (a) that (b) does not have.

• No listed technique. Article 5(1)(a) requires a subliminal, purposefully manipulative or deceptive technique. Article 5(1)(b) requires nothing of the sort. The mechanism can be entirely overt. A system that plainly and truthfully offers something, targeted because the recipient is vulnerable, can engage (b) and cannot engage (a).
• No impairment of informed decision-making. Article 5(1)(a) requires that the distortion operate by appreciably impairing the ability to make an informed decision, and that this cause a decision the person would not otherwise have taken. Article 5(1)(b) goes straight from exploitation to material distortion of behaviour. The intermediate causal step is simply absent.
• No group harm limb, but a wider victim. Where (a) names harm to the person, another person, or a group of persons, (b) names the person or another person. The trade is that (b) expressly catches distortion of the behaviour of any person belonging to the vulnerable group, so it is drafted for targeting a class rather than an individual.

The practical consequence: an assessment that runs the (a) chain, fails at the informed-decision limb, and stops there has not answered the question. If the audience is age-, disability- or situation-defined, (b) has to be run separately.

The three vulnerability grounds

Age

Unqualified in the article text. It covers children, and recital 29 and Article 9(8) both show the Regulation’s general concern for persons under 18, but nothing in (b) limits the ground to minors. Older adults are within it on the plain words.

Disability

Recital 30 ties the term to Directive (EU) 2019/882, the European Accessibility Act, which imports a broad functional understanding of disability — long-term physical, mental, intellectual or sensory impairments which in interaction with various barriers may hinder full participation. That is a wider category than a national disability benefits register would suggest.

A specific social or economic situation

This is the ground that surprises people, and it is the one added during the legislative process rather than present in the 2021 proposal. Recital 30 gives persons living in extreme poverty and ethnic or religious minorities as examples. It is a situational ground, not a status ground: someone in acute financial distress is within it while the distress lasts.

Because it is situational, it is also inferable. A system that does not ask about income but predicts financial distress from behaviour has identified the vulnerability just as effectively, and nothing in the provision requires that the vulnerability be known rather than inferred. Whether that inference itself is lawful is a separate question under the GDPR, and often the more immediately binding one.

What “exploits” adds

The verb is doing real work and it is the least analysed word in the provision. Serving a vulnerable person is not exploiting them. A debt-advice service whose users are by definition in financial distress is not within (b) merely because of who its users are; the system has to take advantage of the vulnerability in a way that turns against the person’s interest.

The natural reading is that exploitation requires the vulnerability to be the reason the distortion works — a causal link between the protected characteristic and the effect on behaviour, running to the person’s detriment. That reading is not authoritative. It is not yet clear how far the term reaches, whether it imports any element of culpability, or how a decision-maker would treat a system whose targeting was accidental. What would settle it is a preliminary reference to the Court of Justice or a national enforcement decision that reasons through the word, and neither exists.

Where this bites in practice

Three shapes are worth checking against (b) rather than only against (a).

• Optimised targeting of credit, gambling or high-cost products. An optimiser that discovers financial distress predicts conversion is doing precisely what the provision describes, with no manipulative technique anywhere in the system.
• Companion and engagement systems used by minors. Age is a listed ground and engagement optimisation is a material distortion of behaviour by design. The live question is the harm limb, which is where the analysis actually happens.
• Accessibility features repurposed for persuasion. A system that knows a user relies on assistive technology and uses that knowledge in a commercial flow rather than in the user’s interest engages the disability ground directly.

Article 5 has bound since 2 February 2025 under Article 113(a), and fines of up to EUR 35,000,000 or 7% of worldwide annual turnover apply under Article 99(3). For the adjacent prohibition on scoring people across contexts, see the social scoring ban and its two limbs; for the data protection angle on children specifically, see GDPR Article 8 and AI products used by children.