Populations vulnerable to AI systems differ from one another about as much as human groups can. A child and an older adult. A refugee and a person with a visual impairment. A person in deep poverty and a speaker of an under-resourced language. They share little in the way of circumstances or needs.
Yet AI systems fail them in strikingly similar ways. The Foundation's examination of specific populations keeps surfacing the same underlying structure, a four-part failure pattern that recurs regardless of which group is affected. Understanding the pattern matters, because a recurring structural failure has a structural fix, while a thousand unrelated failures do not.
Part one: the presumed typical user
Every AI system is designed around an implicit picture of who will use it, their language, literacy, device, connectivity, abilities, attention, and circumstances. This typical user is almost never written down and almost never examined. They are simply built into the system through a long series of small, unremarked design decisions.
Vulnerable populations, by definition, differ from this presumed typical user in some specific and consequential way. So the very first layer of the system, its basic shape, was built for someone who is not them. Everything after that inherits the mismatch.
Part two: underrepresentation in the data
AI systems learn from data. Vulnerable populations are routinely underrepresented in that data, they are harder and costlier to reach, historical datasets already undercount them, or no one assembling the data considered them. A model trained on data that thinly represents a group performs worse for that group; this is one of the most consistent findings in the field.
The result is a grim inversion. The people who can least afford a system that works poorly are the people for whom the system works most poorly. Performance is lowest exactly where the stakes are highest.
Part three: underrepresentation in building and testing
Vulnerable populations are also underrepresented among the people who build AI systems, and among the people those systems are tested with before release. This matters because a failure mode is most reliably caught by someone who would be affected by it, or by someone who designed the test with that person in mind.
When neither is present, the failure mode is not caught. It survives into the deployed system, and it is then discovered the hard way, in the world, by the people it harms, after launch, when correction is slowest and the damage is already done.
Part four: least access to recourse
When the system fails a vulnerable user, that user is typically the least equipped to do anything about it. Recourse is a chain: knowing the decision was automated, understanding it well enough to challenge it, finding the appeal route, and having the time, knowledge, and standing to use it. Vulnerable users are most likely to have that chain break at one or more links.
So the original failure is not corrected. It stands, and the inability to contest it becomes a second harm layered on the first.
Why the four parts compound
These are not four separate problems. They are one mechanism, and the parts reinforce each other. The presumed typical user shapes a system not built for the group. Thin data makes it perform worst for the group. Absence from building and testing means the resulting failures are not caught. And weak recourse means that when those failures land, they cannot be undone. Each part makes the next one worse.
The compounding is also why the harm is so consistent across such different populations. The specifics, what the presumed typical user looks like, which data is thin, which failure modes go uncaught, vary enormously from group to group. The structure does not vary at all. And a problem with a consistent structure can be addressed structurally: examine and widen the presumed user, address the data gap, bring affected populations into building and testing, and build genuine recourse. The pattern that fails every group the same way can, in principle, be fixed for every group the same way.
