AI Bias in Disability Rehabilitation: An Urgent Equity Crisis

There are moments in institutional life when innovation, rather than liberating, quietly reorganizes inequality into more efficient forms. Vocational rehabilitation, historically positioned as a pathway toward dignity through work, now stands within such a moment. The integration of artificial intelligence into its decision-making processes signals not simply technological advancement, but a transformation in how human capacity itself is interpreted, measured, and ultimately sanctioned. What emerges is not a neutral system of allocation, but a patterned logic—one that risks deepening the very inequities it was designed to resolve.

At first glance, the appeal of artificial intelligence within vocational rehabilitation is compelling. Systems that promise predictive accuracy and streamlined service delivery appear to offer clarity in a field often marked by complexity. Yet these systems do not arise in isolation. They are constructed from data that bear the imprint of prior exclusions, historical disparities, and institutional blind spots. As Virginia Eubanks writes, automated systems often “profile, police, and punish the poor” [2, p. 11]. In this sense, AI does not stand outside the social order; it crystallizes it.

This concern becomes especially salient within federally governed vocational rehabilitation systems structured under the Rehabilitation Act of 1973, the Americans with Disabilities Act (ADA), and the Workforce Innovation and Opportunity Act (WIOA). These frameworks articulate a commitment to equity grounded in individualized assessment and nondiscrimination [12][13]. Yet algorithmic systems introduce a competing logic—one rooted in aggregation, prediction, and probabilistic classification.

From a clinical psychological perspective, this tension is profound. Vocational rehabilitation is not merely a technical intervention; it is a process of identity reconstruction. Individuals must rebuild self-efficacy, navigate stigma, and reassert agency. AI systems, however, privilege measurable indicators over lived experience. As Strauser (2014) notes, psychosocial variables central to rehabilitation outcomes often resist standardization [3]. When these are excluded, the system ceases to understand the individual and instead approximates them through statistical inference.

Medical sociology reframes this dynamic by situating disability within systems of power. The social model of disability emphasizes that disadvantage emerges from structural barriers rather than individual deficits. Yet AI systems often revert to a medicalized logic, interpreting deviation from normative datasets as pathology. This shift is not merely technical—it is epistemological.

Here, the insights of Michel Foucault become particularly instructive. Foucault’s concept of biopower describes the ways in which modern institutions regulate populations through classification, surveillance, and normalization. In contemporary algorithmic systems, this regulatory logic is intensified. As Rouvroy and Berns (2013) argue, algorithmic governance represents a form of “governmentality without subject,” in which individuals are managed through data patterns rather than engaged as autonomous agents [17, p. 173]. Within vocational rehabilitation, this manifests as a shift from individualized care to population-level optimization.

Disability, within this framework, becomes a site of algorithmic regulation. Individuals are sorted, scored, and categorized according to predictive models that reflect historical norms of productivity and independence. This process echoes what Foucault described as the “normalizing gaze,” now extended through digital infrastructures. The risk is not merely exclusion, but the subtle redefinition of what counts as a viable life.

This dynamic converges with the logic of surveillance capitalism, as articulated by Shoshana Zuboff (2019). Zuboff argues that contemporary data systems extract behavioral information to predict and shape human action, transforming experience into a form of economic resource [18]. In vocational contexts, this translates into systems that monitor, evaluate, and preemptively categorize workers. As she writes, “surveillance capitalism claims human experience as free raw material for translation into behavioral data” [18, p. 8].

Empirical research supports this convergence. Kellogg, Valentine, and Christin (2020) demonstrate how algorithmic management systems exert control over workers through continuous monitoring and evaluation, reshaping autonomy within labor processes [7]. Similarly, Ajunwa (2020) finds that automated hiring systems systematically disadvantage individuals whose work histories deviate from normative patterns—disproportionately affecting disabled populations [6].

Within vocational rehabilitation, these dynamics intersect with legal mandates for individualized assessment. The Supreme Court’s decision in Olmstead v. L.C. established that unjustified segregation constitutes discrimination under the ADA, requiring services to be delivered in the most integrated setting appropriate [14]. Central to this ruling is the rejection of categorical assumptions about disability.

Algorithmic systems, however, risk reintroducing such assumptions through statistical generalization. Decisions are no longer explicitly categorical, yet they operate through patterns that replicate historical exclusion. Segregation becomes computational—embedded within risk scores and eligibility thresholds. The individual is no longer seen directly, but inferred through data.

Critical disability theory sharpens this critique. As Ruha Benjamin (2019) argues, technological systems can produce a “New Jim Code,” embedding racialized and ableist inequities within seemingly neutral processes [10]. Tom Shakespeare (2014) similarly emphasizes that disability emerges from the interaction between bodies and social environments, not from deviation alone [16]. Algorithmic systems, by privileging normative baselines, obscure this relational understanding.

Empirical evidence underscores the stakes. Obermeyer et al. (2019) found that healthcare algorithms underestimated the needs of Black patients by more than 50% due to biased proxy variables [9]. Research in algorithmic fairness further demonstrates that predictive systems exhibit higher error rates for marginalized populations, reinforcing cumulative disadvantage [1][5].

The consequences extend beyond policy into lived experience. To be evaluated and excluded by an opaque system is to encounter a form of institutional alienation. Clinical research indicates that perceived injustice and lack of agency are associated with increased psychological distress and disengagement from rehabilitation [11]. In Foucauldian terms, the subject is not only governed, but constituted through these systems—shaped by the categories that define them.

Labor market data reinforces the urgency of this concern. The U.S. Bureau of Labor Statistics (2024) reports that individuals with disabilities have significantly lower employment rates than their non-disabled counterparts [15]. Vocational rehabilitation systems were designed to address this disparity. Yet the integration of AI, if unregulated, risks entrenching it—particularly as algorithmic hiring and management systems become more pervasive.

And yet, within this critique lies the possibility of transformation. If algorithmic systems function as instruments of biopower, they can also be sites of resistance and redesign. A human-centered, sociotechnical approach—integrating clinical expertise, sociological insight, and lived experience—offers a pathway forward. As Selbst et al. (2019) argue, fairness must be situated within its social context, not reduced to technical metrics [5].

Policy reform is essential. Transparency, accountability, and inclusive data practices must be institutionalized. Legal frameworks such as the ADA and WIOA already articulate a commitment to equity; the challenge lies in ensuring that emerging technologies do not undermine these principles. The OECD (2021) emphasizes that AI must align with human rights and inclusive growth [8].

Ultimately, the integration of artificial intelligence into vocational rehabilitation compels a deeper reflection on the nature of governance itself. These systems do not merely allocate resources—they shape the conditions under which individuals are recognized, valued, and included.

If AI is to serve the goals of rehabilitation, it must move beyond efficiency toward recognition. Otherwise, we risk constructing systems that are not only unequal, but quietly transformative in their capacity to redefine what it means to belong.

References 

[1] Barocas, S., Hardt, M., & Narayanan, A. (2019). Fairness and machine learning: Limitations and opportunities. MIT Press.
https://fairmlbook.org

[2] Eubanks, V. (2018). Automating inequality. St. Martin’s Press.
https://us.macmillan.com/books/9781250074317/automatinginequality

[3] Strauser, D. R. (2014). Career development, employment, and disability in rehabilitation (2nd ed.). Springer.
https://link.springer.com/book/10.1007/978-1-4614-7775-5

[4] O’Neil, C. (2016). Weapons of math destruction. Crown.
https://weaponsofmathdestructionbook.com

[5] Selbst, A. D., et al. (2019). Fairness and abstraction in sociotechnical systems.
https://doi.org/10.1145/3287560.3287598

[6] Ajunwa, I. (2020). The paradox of automation as anti-bias intervention. Cardozo Law Review.
https://cardozolawreview.com/the-paradox-of-automation-as-anti-bias-intervention

[7] Kellogg, K. C., Valentine, M. A., & Christin, A. (2020). Algorithms at work.
https://doi.org/10.5465/annals.2018.0174

[8] OECD. (2021). Artificial intelligence, automation and work.
https://doi.org/10.1787/ae747c0d-en

[9] Obermeyer, Z., et al. (2019). Dissecting racial bias in algorithms. Science.
https://doi.org/10.1126/science.aax2342

[10] Benjamin, R. (2019). Race after technology. Polity.
https://www.ruhabenjamin.com/race-after-technology

[11] Corrigan, P. W., & Rao, D. (2012). Self-stigma.
https://doi.org/10.1177/070674371205700804

[12] U.S. Department of Labor.
https://www.dol.gov/agencies/odep

[13] EEOC. ADA statute.
https://www.eeoc.gov/statutes/americans-disabilities-act-1990

[14] Olmstead v. L.C. (1999).
https://supreme.justia.com/cases/federal/us/527/581/

[15] U.S. Bureau of Labor Statistics. (2024).
https://www.bls.gov/news.release/disabl.nr0.htm

[16] Shakespeare, T. (2014). Disability rights and wrongs revisited.
https://www.routledge.com/Disability-Rights-and-Wrongs-Revisited/Shakespeare/p/book/9780415527620

[17] Rouvroy, A., & Berns, T. (2013). Algorithmic governmentality. Réseaux, 177(1), 163–196.
https://doi.org/10.3917/res.177.0163

[18] Zuboff, S. (2019). The age of surveillance capitalism. PublicAffairs.
https://www.publicaffairsbooks.com/titles/shoshana-zuboff/the-age-of-surveillance-capitalism/9781610395694/