A pillar of functioning social systems is to establish the needs of individuals or groups of people who may be vulnerable to specific forms of harm and to provide adequate protection.

People can be or become vulnerable – that is ‘unable to take care or protect themselves or others from harm or exploitation’ for a variety of reasons. The more commonly known ones range from structural injustices, such as when social safety nets collapse and expose people to state dependency, to sudden events like meteorological disasters, to changes affecting people’s surroundings or social life, like the introduction of a new system to access public services. Usually, it is relatively easy to see, if not anticipate, who is or will be vulnerable, and why.

The use of AI systems, as well as simpler data-driven technologies, complicates this picture.

These tools compute aspects of our lives according to new, complex and unpredictable protocols. This may render people vulnerable in novel ways, ways that are incomparable to those society has had to respond to before; that compound or hard-code existing forms of structural injustice; and that expose people who are not usually identified as vulnerable (e.g. people who are not identified through a protected characteristic) to increased risk. Indeed, what we see more and more often is that anyone, at different times, could become vulnerable to AI-driven harm in unanticipated ways.

The notion of vulnerability has a long history of varied applications and meanings across domains such as law, disaster relief and cybersecurity. It is often associated with concepts of inequality, discrimination and bias, but differs from them as someone’s state of vulnerability can be difficult to evidence and use as the basis for making recourse to justice.

In this blog post, we grapple with the new kind of vulnerability precipitated by the use of AI technologies and ask whether the notion of vulnerability itself could help reframe the discourse around – and prevent – AI-driven harm. We introduce and interrogate the concept of ‘vector of vulnerability’ which we take to be the combination of why, how, in which context of use and when the application of an AI tool can lead to the exposure of specific groups of people to excess risk of harm.

Exploring this kind of vulnerability is especially important in the current climate where increased investments in technological development are not matched with protections for people and society. And there is little to no formal recourse should people experience harm due to the use of AI technologies.

Our over-arching questions are:

• Can applying the lens of vulnerability to AI design choices, applications of use and policies help anticipate both the types of harm and the groups of people who are more likely to experience it?
• Can this approach help us bake necessary safeguards into the development process of both AI technologies and the policies governing them?
• If we can name and describe the pathways leading to AI-driven harm, can we establish where existing protections are working and find ways to improve them when they don’t?
• As AI applications may expose everybody to excess risk at different times and in different contexts, can conceptualising protection around the notion of vulnerability improve safeguards for all?

Vulnerability as captured in existing research

In this blog post, we index vulnerability to a time – before harm happens – and a metaphorical space that people enter upon interacting with a data-intensive technology. Those who find themselves in such a ‘space’ may already be identified as vulnerable or minoritised in society or they may be people who usually aren’t identified as vulnerable.

To explore how this space emerges, we undertook a thematic analysis of Ada’s research publications between 2021 and 2025. We observed the complex interaction of sociotechnical factors and policy decisions determining who faces AI-driven harm and noted that the pathway from cause to negative effect is not always clear.

Vulnerability can arise from several factors – the kind of data used when programming a technology, how tools are designed and the resources they use – and, in each case, we can try to draw a vector describing how these factors intersect each other.

Vectors of vulnerability may cut across the context of application of a tool. For example, using an advanced AI assistant for everyday tasks might not carry substantial risk and make someone vulnerable, but using it to manage mental health issues could introduce a vulnerability when someone relies and becomes dependent on the assistant, replacing human connections.

Vulnerability may be indexed to whether a person can choose and opt-out of using an AI technology. Users of immersive technologies, for instance, could become vulnerable in different ways, by virtue of using the same technology under different conditions and being subjected to different power dynamics. Workers employed in factories that deploy immersive technologies, for example, are exposed to intrusive monitoring and novel forms of workplace surveillance. Those using immersive technologies for entertainment may be vulnerable to under-regulated and understudied data collection and privacy breaches or be exposed to traumatising virtual assaults.

In another example, a recent report focusing on the financial sector highlighted how daily users of generative AI applications for financial decisions are more exposed to risk. In this case, early adoption of the technology may lead to people becoming vulnerable and facing accrued financial and privacy risks.

In other cases, the adoption of an AI tool in a field of production may change the field entirely and how employment is organised within it. In the creative industries, including fashion, TV and cinema, the introduction of generative AI models has brought about specific forms of vulnerability that are relatively new to these sectors, including heightened threats to contractual rights and people’s livelihoods.

Vectors of vulnerability become visible when we compare different life experiences and ways of interacting with data-driven technologies. And the diversity of pathways to vulnerability suggests the need for a comprehensive analytical framework to explore how AI affects the lives of different groups of people.

Examples of vulnerability and possible approaches to vulnerability analysis

The details of how someone may be made vulnerable by using AI technologies become clearer when we consider empirical examples.

If we use the notion of vulnerability to analyse and prevent pathways to harm, we need to develop consistent approaches to our examples. We can look at AI-driven vulnerability from various perspectives, each approach bringing a different understanding of the vectors leading to an amplification of risk.

If we focus on sectoral AI adoptions, where safeguards are in place to protect structurally minoritised groups, we can study how AI and data-driven technologies are making unexpected groups of people vulnerable. If instead we start with a specific tool and track its different use cases, we can analyse which specific applications make people vulnerable and who these people are likely to be.

In the table below, we offer two different potential entry points for interrogating vulnerability and begin to show the extent of the complexity and entanglement of different factors that the notion of AI-driven vulnerability might bring.

The use of AI is leading to forms of vulnerability which are novel both in nature and in scale. While there is still little evidence on the exact pathways leading to vulnerability, we know they are numerous, complex and hard to predict. Exploring these pathways can steer policymaking towards more effective protections that are proactive rather than reactive. These protections should be for any individual who may otherwise fail to reasonably protect themselves as well as for those that we, as a society, already acknowledge need protection.

We need a new agenda that defines and names the causes of vulnerability, and can push for the institution of the right safeguards. Currently, without the ability to name and describe the nature and scale of the problem, many of us who are vulnerable will be harmed and will likely be unable to seek redress.