The cycle’s argument is that capable machines press the human question to the surface—that understanding AI requires understanding what we owe one another in a world where the machinery of social sorting has been automated. The excoded are the people for whom the sorting has gone wrong, and Buolamwini insists that their experience is not peripheral to the cycle’s central questions but central to them: the question of who gets to be fully human in the world we are building is answered, in practice, by which faces the systems recognize, which voices they transcribe faithfully, which people’s existence they represent accurately in the synthetic culture they produce.
The excoded also constitute the cycle’s most direct answer to abstraction. It is easy to debate the AI transition as a question of capability and scale, of productivity and displacement, of governance and alignment. The excoded insists on the person at the other end of every deployed system—the specific individual who is sorted, scored, seen, or overlooked—and refuses the analytical operation that turns her into an acceptable error rate.
Buolamwini introduced the term in Unmasking AI (2023) alongside the coded gaze, treating the two as companion concepts: the coded gaze names the structural mechanism, the excoded names its human consequence. The pairing is methodologically consistent with her career—every empirical finding is given both a structural name and a human name, because both are necessary to produce accountability: the structural name for institutions to address, the human name for courts to protect.
The Gender Shades study that established the disparity in facial-analysis systems was the founding empirical case for the excoded, but Buolamwini was always explicit that the concept generalizes beyond faces to every algorithmic system that makes consequential decisions about people. Hiring algorithms, credit-scoring models, recidivism-prediction tools, healthcare triage systems—each produces its own population of excoded, and the populations overlap: the people most likely to be wrongly assessed by one system are typically the people most likely to be wrongly assessed by the others.
Collectivizing the harm. The most important function of the term is to resist the industry’s atomization of algorithmic harm. Individual errors become isolated incidents; the excoded become a class. And a class can assert collective rights, demand systemic remedies, and exercise political pressure in ways that isolated individuals cannot. The naming is a political act as much as a descriptive one.
Double exposure. The excoded are not merely the people harmed by systems that fail them. They are also, disproportionately, the people targeted by systems that work. The community that is most likely to be misidentified by facial recognition is also the community most likely to be scanned by it. The worker whose résumé is most likely to be discarded by a hiring algorithm is also the worker who most needs algorithmic screening to work correctly. The harms compound at the intersection.
Invisibility of recourse. The excoded are vulnerable in a specific way: the harm is often invisible to them. They may never learn that a model rejected them, let alone why. Even where they do learn, the avenues of redress are thin—no appeal, no explanation, no human who will take responsibility. Buolamwini has been pointed about this gap: even significant AI policy falls short precisely on the questions of redress and consequence for companies whose systems harm real people.
Generative expansion. The generative turn expands the category beyond sorting and denial to representation and fabrication. A person who is consistently misrendered, stereotyped, or erased by the dominant generative systems is excoded in a new sense—not denied a service but denied accurate existence in the synthetic infrastructure through which culture is increasingly produced. The harm is more diffuse but no less real, and it falls along the same familiar lines.