Human in the Loop: Artificial Intelligence, Disability, and Hidden Ableism

Abstract

The film Humans in the Loop, directed by Aranya Sahay, presents a compelling narrative about Nehma, an Adivasi woman from Jharkhand who becomes part of the invisible workforce training artificial intelligence systems through data labelling work. This research paper contextualises the film within emerging scholarship on disability studies and artificial intelligence, examining how AI systems perpetuate ableist assumptions, reinforce normalcy, and marginalise disabled communities. Through analysis of the film’s narrative, combined with critical examination of disability-inclusive AI frameworks, this paper argues that the intersection of AI development and disability justice reveals fundamental contradictions in how technology is created, designed, and deployed. The paper explores three critical dimensions: first, how human-in-the-loop systems rely on marginalised labour while obscuring the biases embedded in the training data; second, how AI systems encode ableist assumptions that treat disability as a deviation from normalcy requiring correction rather than as a valued mode of human difference; and third, how accessibility to whom and legibility for what become central questions when examining whose needs AI is designed to serve. By situating Nehma’s labour within broader conversations about disability, technology, and justice, this paper demonstrates that building truly inclusive AI requires centering disability perspectives from the earliest stages of design, recognising disabled people as knowledge producers rather than mere subjects of intervention, and fundamentally reimagining what we mean by access, participation, and equity in technological systems.

Introduction: The Invisible Work Behind Intelligent Machines

When we encounter artificial intelligence in our daily lives—in recommendation algorithms, facial recognition systems, or content moderation tools—we rarely think about the human beings whose labour made those systems possible. The work is invisible by design. Somewhere, in offices and centres across the Global South, thousands of workers sit before computer screens, clicking, labelling, and categorising images, text, and data to train machine learning models. This work, often called “ghost work” in the literature, is the foundation upon which the glamorous world of AI innovation rests. Yet this foundational labour remains hidden, underpaid, and often performed by those most marginalised within global systems of power: women, people from indigenous communities, people living in rural areas, and individuals from economically disadvantaged backgrounds.

Aranya Sahay’s 2024 film Humans in the Loop brings this invisible world into sharp focus. The film follows Nehma, played by Sonal Madhushankar, an Oraon tribal woman who returns to her village in Jharkhand after her marriage dissolves. Facing a custody battle and economic precarity, Nehma takes a job at a data labelling centre operated by an international technology firm. Her task is simple on the surface: look at images and mark what she sees according to a predetermined set of categories. Label caterpillars as pests. Mark faces according to ethnicity. Categorise objects by type. These decisions, made by Nehma and workers like her, directly feed into the training datasets that shape how AI systems “understand” and classify the world.

The brilliance of Sahay’s film lies in how it forces viewers to confront several uncomfortable truths simultaneously. The first is economic: Nehma’s labour is essential to the functioning of global AI systems, yet she receives minimal compensation and no recognition for her contribution to systems worth billions of dollars. The second is cultural and epistemic: her knowledge—knowing that caterpillars are not pests but actually protect plants—is actively suppressed in favour of standardised, corporate-determined categories. The third is deeply personal: the demands of this labour directly harm her relationship with her daughter Dhaanu, creating a tragic irony in which she must suppress her own voice and human intuition to train systems designed ostensibly to “help” humanity.

It is within this layered complexity that disability studies offers a crucial analytical lens. While Humans in the Loop does not explicitly address disability, the mechanisms it depicts—the devaluing of human knowledge in favour of standardised metrics, the erasure of individual contexts and experiences, the treatment of deviation from normality as error to be corrected—are precisely the mechanisms through which AI systems enact harm against disabled people. When we examine the film through a disability justice framework, we see that the ableist logic embedded in data labelling work mirrors the ableist logic embedded in AI systems themselves. Both assume a “normal” user, a “normal” body, a “normal” way of processing and understanding the world. Both treat anything that deviates from this imagined normality as a problem to be solved, corrected, or excluded altogether.

This paper brings together three critical conversations: the analysis of invisible labour in AI systems as depicted in Humans in the Loop, the emerging scholarship on disability, exclusion, and artificial intelligence, and the practical frameworks for thinking about accessibility and legibility in technological systems. By situating these conversations in dialogue with one another, we argue that a genuinely inclusive AI ecosystem requires centering disability perspectives not as an afterthought or accessibility feature, but as a fundamental principle of design and development from the very beginning. More broadly, this paper contends that the film’s depiction of human-in-the-loop systems reveals something critical about AI development in general: that the same mechanisms through which marginalised workers are exploited are the mechanisms through which AI systems systematically exclude and harm marginalised communities, including disabled people.

Part One: The Film as Documentation of Systemic Exclusion

The Context of Data Labelling in Global AI

To understand the significance of Humans in the Loop, we must first understand the material reality of data labelling work in the global AI economy. Sahay’s film is based on an article by journalist Karishma Mehrotra titled “Human Touch,” published in FiftyTwo magazine, which documents the experiences of women workers in Jharkhand engaged in exactly this kind of work. The article, and the film that follows, are rooted in meticulously researched reality.

Data labelling is among the most labour-intensive aspects of training modern AI systems. Large language models, computer vision systems, and other contemporary AI applications require massive datasets of human-annotated examples. Researchers and engineers at major technology companies—working for firms like Google, Amazon, Facebook, and numerous other multinational corporations—cannot possibly generate these datasets themselves. Instead, they outsource the work to global networks of contract workers, often based in economically disadvantaged regions where labour is cheap. These workers, the vast majority of whom are women, spend their days doing repetitive, mentally demanding labour: looking at images and deciding what category they belong to, reading text snippets and labelling their sentiment, listening to audio and transcribing what they hear.

What makes this work particularly insidious is how it is presented, particularly in the marketing of technology companies and in mainstream media coverage of AI. The public narrative around AI emphasises innovation, genius, and forward-thinking design. The engineers and researchers who conceptualise these systems are celebrated as visionaries. Yet the actual construction of AI systems depends fundamentally on the exploited labour of those whom society deems least visible and least valuable. This is what Sahay captures in his film: the contradiction between the grand promises of AI technology and the ground-level reality of how these systems are actually built.

In Humans in the Loop, Nehma’s work is portrayed with careful attention to its texture and rhythm. We see her sitting at a computer terminal, hour after hour, making split-second decisions about how to categorise images and data. The film does not romanticise this work; nor does it present Nehma as particularly clever or special in her ability to do it. Rather, it presents her as what she is: a working woman doing her best to survive, to keep her family together, and to maintain her dignity in a system that offers her little compensation and less recognition.

Nehma’s Labour and the Suppression of Knowledge

One of the most powerful scenes in the film occurs when Nehma is asked to label caterpillars as pests. The task seems straightforward: look at an image of a caterpillar, mark it as a pest, submit. But Nehma hesitates. She knows, from her lived experience as a farmer and as someone embedded in a knowledge system shaped by generations of working the land, that caterpillars are not simply pests. They serve an ecological function; they actually protect plants by eating away the rotting parts. To label them as pests would be to embed a false understanding of ecological relationships into the AI system’s training data.

Nehma’s resistance to this categorisation is significant. It represents the moment where her embodied, contextual knowledge—knowledge grounded in place, in indigenous practices, in generations of agricultural wisdom—comes into direct conflict with the decontextualised, standardised categories required by the AI system. The tech firm’s response is instructive: they do not engage with Nehma’s epistemic authority or the validity of her knowledge. Instead, they dismiss her concerns and eventually reprimand her for providing “faulty” labels. In the logic of AI training data, there is no room for contextual nuance, no space for indigenous knowledge systems, no possibility that the standardised category might be wrong.

This moment in the film encapsulates something crucial about how AI systems perpetuate not only economic exploitation but also epistemic injustice. Epistemic injustice, a concept developed by philosopher Miranda Fricker, refers to the wrongful exclusion of certain people or forms of knowledge from knowledge-producing practices. When Nehma’s knowledge about caterpillars and ecological relationships is dismissed in favour of a standardised categorisation, she is being subjected to epistemic injustice. Her knowledge, grounded in place and practice, is deemed less valid than the abstract category determined by engineers thousands of kilometres away.

This becomes even more troubling when we consider that the AI system trained on Nehma’s (coerced) labels will now carry forward this misunderstanding of ecological relationships. The system, once deployed, might be used in agricultural contexts where it could cause real harm. A farmer relying on the AI system to identify pests might destroy beneficial insects, leading to crop damage. In this way, the suppression of Nehma’s knowledge does not simply constitute an injustice against her as a worker; it potentially cascades outward to cause harm to others.

The Erasure of Identity and Presence

Another crucial scene involves Nehma’s supervisor showing her AI-generated images of tribal women. These images were generated by an AI system trained, at least in part, on data that Nehma herself had labelled. Yet the images do not reflect Nehma’s own appearance or any recognisable representation of tribal women as Nehma knows them. Instead, the AI has generated images of women with light skin, light-coloured hair, and facial features that conform to dominant aesthetic standards. In other words, the AI has taken Nehma’s labour and produced images that erase the very people it purports to represent.

This scene is particularly poignant because it demonstrates how AI systems do not neutrally encode information. Rather, they encode the biases, assumptions, and aesthetic preferences of those who built them and of the datasets they were trained on. By the time Nehma’s labour contributes to these systems, biases have already been embedded at multiple levels: in how the data was collected, in how categories were chosen, in whose images were included and whose were excluded. Her labour, rather than correcting these biases, simply helps to perpetuate and amplify them.

The erasure of Nehma’s presence in the AI-generated images also reflects a broader pattern in how technology companies approach diversity and representation. Many technology firms, facing criticism about the lack of diversity in their products and services, have moved toward collecting more diverse training data and implementing more inclusive design practices. Yet as Sahay’s film demonstrates, this move toward diversity can itself become a mechanism of erasure. Diverse data is collected from marginalised communities, processed through systems designed by predominantly non-disabled, non-marginalised engineers, and then used to generate products that conform to dominant standards. The labour of those communities is appropriated while their actual agency, knowledge, and preferences remain invisible.

The Human Cost: Family, Relationship, and Emotional Labour

While the film engages with the macro-level questions of labour exploitation and epistemic injustice, it also explores the deeply human cost of this work. Nehma is not only a worker; she is also a mother. She is trying to maintain a relationship with her daughter Dhaanu while navigating the demands of her job and a difficult custody situation. The film portrays how the demands of her labour directly harm her ability to be present for her daughter. When Nehma comes home exhausted from work, she has little emotional energy to invest in her relationship with Dhaanu. When she must suppress her own voice and knowledge to comply with the corporation’s demands, she is modelling for her daughter a particular relationship to authority and self-silencing.

This dimension of the film—the intersection of reproductive labour, emotional labour, and waged labour—is particularly relevant to our analysis because it demonstrates how the harms of exploitative labour systems are not confined to the workplace. They ripple outward into intimate family relationships and into the psychic and emotional lives of workers. For Nehma, the work at the data labelling centre is not simply an economic transaction; it is a demand that she reshape her entire being—her knowledge, her voice, her intuitions—to fit the requirements of an abstract system.

This brings us to a critical insight: the labour demanded by AI training is not simply manual labour. It is also cognitive labour and emotional labour. It requires workers to make split-second judgments, to suppress their own contextual knowledge in favour of standardised categories, to maintain focus and accuracy over long hours of repetitive work. This form of labour is particularly demanding, and its costs are borne most heavily by those with the fewest resources and the most precarious positions within systems of power.

Part Two: Ableism, Normalcy, and the Architecture of AI Systems

Defining Ableism in Technology Development

To understand how AI systems harm disabled people, we must first have a clear understanding of what ableism is and how it functions in technological contexts. Ableism refers to a system of beliefs and practices that privileges able-bodiedness and able-mindedness as the norm, treats disability as a deviation from this norm, and structures society to advantage non-disabled people while disadvantaging and excluding disabled people. Ableism is not simply individual prejudice; it is a systemic and structural phenomenon embedded in institutions, policies, practices, and cultural assumptions.

In the context of technology development, ableism manifests in multiple ways. First, it shapes which people are hired to design and build AI systems. The vast majority of AI engineers and researchers are non-disabled people, predominantly from privileged socioeconomic backgrounds. This homogeneity means that the design decisions embedded in AI systems reflect the experiences and assumptions of non-disabled people. Systems are built around the assumption of a “normal” user with “normal” abilities, and any deviation from this imagined norm is treated as an edge case or an exception to be handled separately, if at all.

Second, ableism shapes the data used to train AI systems. Because disabled people represent a statistical minority (though this varies depending on how disability is defined and counted), and because data collection efforts often exclude or underrepresent disabled people, AI training datasets frequently contain little or no information about disabled people’s experiences, needs, or ways of being in the world. An AI system trained on such data will inevitably develop a distorted understanding of humanity that erases or misrepresents disabled people.

Third, ableism shapes the very conceptualisation of what AI is supposed to do and how it is supposed to function. AI systems are often framed as tools to increase efficiency, standardise processes, and eliminate variability. Yet human variation—including disability-related variation—is treated as something to be eliminated rather than accommodated or valued. This logic treats disability as something inherently problematic that needs to be engineered away or corrected.

Disability Models and AI Design

To understand how ableism becomes encoded in AI systems, it is useful to consider different models of disability and how each shapes technological design. The medical model of disability treats disability as an individual deficit or deficiency located within the body or mind of a disabled person. From this perspective, disability is a problem that should ideally be cured or corrected. The social model of disability, by contrast, locates disability not in individual bodies but in the interaction between individuals and social environments. From this perspective, disability results from the mismatch between people’s capacities and the demands and design of social structures. A person might have a mobility impairment, but that person is not “disabled” until they encounter stairs, narrow doorways, or other environmental barriers that their bodies cannot navigate.

These two models have profoundly different implications for AI design. If developers adopt a medical model, they will seek to design AI systems that identify disability and correct it. They might develop AI systems designed to detect and intervene in autism, to identify and classify mental illness, or to predict who will become disabled. Such systems often embed eugenic logic—the idea that disability should be eliminated or prevented. They also treat disabled people as passive subjects of intervention rather than as active agents in their own lives.

If developers instead adopt a social model of disability, they will ask different questions about AI. Rather than asking how to use AI to correct or eliminate disability, they will ask how to design AI systems that accommodate diverse ways of being and doing. They will consider not whether a person can see or hear in “normal” ways, but whether the AI system can adapt to multiple forms of sensory input. They will recognise that many people experience their disabilities not as personal tragedies but as integral aspects of their identity and experience.

Yet in practice, most AI systems have been developed with neither explicit consideration of the medical model nor the social model of disability. Instead, they have been built with what we might call an implicit ableist model: the assumption that there is a single “normal” way to see, hear, move, think, and interact with the world, and that anything deviating from this norm is an outlier or an error. This implicit model is particularly insidious because it often goes unrecognised and unexamined. Developers do not think of themselves as embedding ableist assumptions; they simply think of themselves as building systems that work for “users” or “people,” not recognising that their conceptualisation of what a user is and what people are like is deeply shaped by their own non-disabled perspectives and experiences.

Algorithmic Marginalization and Disability

When we combine the homogeneity of the AI development workforce with the underrepresentation of disabled people in AI training data and the implicit ableist assumptions embedded in system design, we see the emergence of what scholars call “algorithmic marginalization.” This term refers to how AI systems disproportionately harm those already marginalised by existing systems of power. For disabled people, algorithmic marginalization manifests in multiple concrete ways.

In hiring, for example, some companies have begun using AI systems to screen job applicants. These systems are trained on historical hiring data and make predictions about which candidates are most likely to succeed. Because disabled people have historically faced discrimination in hiring and have been underrepresented in many fields, the AI systems trained on this historical data learn to predict that disabled people are less likely to succeed. When the AI system is then used to screen new applicants, it systematically excludes or deprioritises disabled candidates, reproducing and amplifying historical discrimination. A disabled person with atypical eye movements might be filtered out before their actual qualifications are even considered. A person with a speech disability might be rejected by an automated phone screening system that cannot interpret non-standard speech patterns. The discrimination is enacted not by a human hiring manager making a conscious choice but by an algorithm whose biases have been hidden inside mathematical operations.

Similarly, in healthcare, AI systems used to predict which patients are at risk of certain conditions or which patients would benefit most from certain interventions have been found to systematically misclassify or underestimate the needs of disabled people. A system trained primarily on data from non-disabled people might fail to recognise how a particular symptom manifests in someone with a different sensory or cognitive system. The result is that disabled people receive worse healthcare decisions, not because individual doctors are biased, but because the AI system has encoded an ableist understanding of what health and illness look like.

In education, AI tutoring systems and automated grading systems often fail to accommodate disabled students. A student with dyslexia might struggle with a tutoring system that presents information in a particular format without the ability to customise that presentation. A student who is Deaf might find that an automated grading system that relies on audio input cannot evaluate their work at all. The AI system, designed around assumptions about how learning happens and how knowledge should be demonstrated, becomes a barrier rather than a tool.

These examples all share a common pattern: AI systems trained on data from predominantly non-disabled populations, designed by predominantly non-disabled teams, and evaluated according to metrics that do not account for disabled people’s experiences, systematically disadvantage disabled people. The harm is not intentional, but it is nonetheless real and consequential.

The Problem of “Fixing” Disability Through Design

One particularly troubling approach to disability and technology is what some scholars call “technoableism”—the impulse to use technology to “fix” or “overcome” disability. This approach has a superficial appeal: if we can design technology that allows disabled people to see, hear, walk, or think in ways closer to non-disabled norms, surely that is a good thing? Yet this framing contains several hidden problems.

First, it treats disability as an individual problem located in the disabled person’s body, rather than recognising disability as produced through the interaction between people and their environments. It obscures the role of social barriers, discrimination, and inaccessible design in creating disability. By locating the “problem” in the disabled person, technoableism reinforces the medical model of disability and the idea that disabled people need to be corrected or fixed.

Second, it places the burden of adaptation entirely on disabled people. Rather than asking how systems and environments can be redesigned to accommodate diverse human variation, technoableism asks: how can we technology to make disabled people more like non-disabled people? This places an enormous burden on disabled people to constantly adapt, to use technologies that are often uncomfortable or cumbersome, and to hide or suppress their disabilities to fit into systems not designed for them.

Third, technoableism often obscures the economic interests driving technological development. When a company develops a new assistive technology and markets it to disabled people as a solution to their problems, what often goes unexamined is whether the technology actually solves the problems disabled people have identified, and whether it is accessible and affordable to the people who might benefit from it. Many assistive technologies are prohibitively expensive, designed by non-disabled people without input from disabled communities, and inaccessible even to disabled people due to other accessibility barriers.

These problems are directly relevant to understanding AI systems and disability. Many proposals for using AI to address disability problems are rooted in technoableist logic. Companies propose AI systems to detect autism in young children, with the unstated but clear implication that autism should be detected and possibly prevented or “corrected.” Researchers develop AI systems to “predict” who will develop mental illness, often with the goal of early intervention or prevention. These approaches may have some benefit, but they also embed ableist assumptions: that disability is a problem located in individuals that should ideally be prevented or corrected, rather than that society should be designed to accommodate disabled people as we are.

Part Three: Accessibility for Whom? Legibility for What?

Rethinking Access in Technological Systems

The concept of “accessibility” in technology often refers to features or accommodations added to existing systems to make them usable by people with disabilities. A website might have alt text for images, allowing people using screen readers to access the content. Software might have keyboard navigation, allowing people who cannot use a mouse to navigate the interface. These accommodations are valuable, but they often embody a particular logic: the system is designed for a “normal” user, and accommodations are added afterward for people with disabilities.

This logic has several implications worth examining. First, it assumes that the “normal” design is correct, and that people with disabilities must adapt to fit into that design. Second, it treats disability access as an addition to the core product rather than as integrated into its fundamental architecture. Third, it often results in accommodations that are clunky, difficult to use, or that mark disabled people as different or exceptional. A screen reader interface, for example, can be more difficult to use than a visual interface, even though the information should be equally accessible. A keyboard navigation system might require many more keystrokes than mouse navigation to accomplish the same task.

More fundamentally, the question “accessibility for whom?” pushes us to interrogate what accessibility actually means and who gets to define it. In many cases, accessibility features are designed by non-disabled engineers based on their assumptions about what disabled people need, rather than through collaboration with disabled people themselves. The result is often that accessibility features address some needs while creating new barriers for others. A video captioning system, for example, might be helpful for Deaf people or people in noisy environments, but if the captions are poorly formatted or contain errors, they might be less useful than no captions at all.

Similarly, the question “legible to what?” pushes us to ask whose interpretations and understandings matter. When an AI system is trained to recognise faces, the question is not simply whether it can recognise faces, but whose faces it can recognise and according to what definitions of “face” or “person.” If a Deaf person communicates through sign language and facial expressions, is that a recognisable “face” to the AI system? If a person wears a niqab or other religious covering, how does the system interpret that? If a person has facial differences due to a condition like Treacher Collins syndrome, does the system still recognise them as a person?

These questions matter because legibility is not neutral. What is legible to a system determines what that system can do and, crucially, what people can do with that system. If a system cannot recognise certain faces, certain people become invisible to that system. If a system cannot interpret certain communication styles, certain people cannot communicate with that system. Legibility, then, is about power: whose ways of being, doing, and communicating are recognised as valid and intelligible by technological systems?

Accessibility and Disability Justice Praxis

Moving beyond accessibility as an add-on feature, disability justice frameworks offer a different way of thinking about design. Disability justice, developed by disability justice organisers and scholars, is an approach that centres the leadership and knowledge of disabled people, particularly disabled people of colour, in imagining and building more just and equitable worlds. Rather than starting with a “normal” product and adding accommodations, disability justice approaches start by asking: what do disabled people actually need? What are our priorities? What would systems look like if they were designed around our needs and preferences rather than adapted to fit existing designs?

This approach has implications for how we think about AI and technology more broadly. If we were to design AI systems from a disability justice perspective, we would not start by asking how to make disabled people fit into systems designed for non-disabled people. Instead, we would ask what disabled people need from AI systems. We would centre disabled people in the design process from the very beginning, not as subjects to be tested on but as co-designers and knowledge producers. We would recognise that disabled people have expertise about disability, about accessibility, and about navigating worlds not designed for our needs. This expertise should be valued and compensated.

Furthermore, a disability justice approach would interrogate the very purposes for which AI is being developed. Is AI being developed to surveil and control disabled people, or to support disabled people’s autonomy and self-determination? Is it being developed to “fix” or “cure” disability, or to make the world more accessible and equitable for disabled people as we are? These are not neutral technical questions; they are political questions about what kind of world we want to build.

Disability-Smart Prompts and Challenging Ableism in AI Use

One practical framework for thinking about how to use AI more equitably is the concept of disability-smart prompts. A disability-smart prompt is a prompt to an AI system (like a large language model or image generation system) that is designed to push back against ableist assumptions and to centre disability perspectives. For example, when asking an AI system to generate an image of a “successful businessperson,” a traditional prompt might result in images of people who conform to able-bodied, often white, able-minded norms. A disability-smart prompt, by contrast, might specify: generate images of successful businesspeople with a range of disabilities, including people using wheelchairs, people who are Deaf, people who are neurodivergent, and people with invisible disabilities. The prompt actively works against the system’s tendency to reproduce ableist stereotypes and erasures.

Disability-smart prompts are valuable because they demonstrate that users have some agency in how they interact with AI systems. Rather than passively accepting whatever output the AI system generates, users can deliberately craft prompts that push back against biases and that actively work to make AI systems more equitable and disability-inclusive. This is particularly important given that most AI systems, as currently designed and deployed, will perpetuate ableist assumptions unless deliberately prompted to do otherwise.

However, disability-smart prompts are not a complete solution. They require users to have knowledge about disability, to be aware of ableist biases, and to have the time and energy to craft careful prompts. They place the burden on individual users to counteract systemic biases rather than addressing the biases in the system itself. More fundamentally, they work only for open-ended generative systems where users have control over prompts. For many AI systems—algorithmic hiring systems, automated medical diagnostic systems, content moderation systems—users do not have this kind of control. The biases are baked in, and individual users cannot prompt their way around them.

This means that while disability-smart prompts are a useful tactical tool, they cannot be a substitute for structural change in how AI systems are designed, developed, and deployed. We still need diverse teams of developers including disabled people. We still need disability-focused data collection and annotation. We still need accountability mechanisms when AI systems cause harm to disabled people. We still need policy changes that make inclusive AI a requirement rather than an optional feature.

Part Four: The Convergence of Labour Exploitation and Ableist Design

Connecting Nehma’s Labour to Broader Systems of Disability Exclusion

We can now return to Humans in the Loop with a deeper understanding of how the film depicts not just labour exploitation but also the mechanisms through which ableist assumptions become embedded in AI systems. When Nehma labels images according to predetermined categories, suppressing her own knowledge about ecological relationships, she is doing more than participating in her own economic exploitation. She is also participating in the creation of AI systems that will be fundamentally ableist in their assumptions and operations.

Consider the scene where Nehma is told to label images of tribal women, and the resulting AI-generated images erase the actual characteristics of tribal people in favour of conforming to dominant aesthetic standards. This is not simply a problem of lacking diverse training data. It is a problem of how the categories themselves—the predetermined boxes that Nehma is asked to sort information into—embed particular assumptions about what is “normal” or “default” and what is deviant or exceptional. The category “woman” might assume a particular body type, a particular gender presentation, a particular way of appearing. The category “person” might assume a particular way of seeing or hearing or moving. These categories, created by non-disabled designers and implemented through the labour of workers like Nehma, then become baked into the AI systems that will shape how the world is interpreted for millions of people.

Moreover, when Nehma is told that her labels are “faulty” because she included contextual knowledge that contradicted the predetermined categories, this is not simply a case of an individual worker being disciplined by an employer. It is a moment where a particular form of knowledge—contextual, embodied, grounded in place and community—is being actively suppressed in favour of decontextualised, standardised categories. This is precisely the same process through which disabled knowledge, disabled ways of knowing and being, are suppressed in the design and development of AI systems.

The connection becomes even clearer when we consider that disabled people, like indigenous people and like women in the Global South, are often expected to participate in the creation of systems that will then exclude or harm them. Disabled people may be hired to test accessibility features or to label data about disability, their labour appropriated to make AI systems that will then be used in ways that harm disabled communities. Nehma labels data that becomes an AI system that is then deployed in agricultural contexts, where it might cause harm. Similarly, disabled people might be asked to provide data about their disabilities, which is then used to train systems designed to identify and “correct” those disabilities.

This represents a particular form of what we might call epistemic extraction: the appropriation of knowledge, data, and labour from marginalised communities to create systems that serve the interests of dominant groups and that often encode the subordination of those marginalised communities. In the case of disability and AI, epistemic extraction means extracting data about disabled people’s experiences and embodied knowledge, using that extraction to train systems designed by non-disabled people, and then deploying those systems in ways that further marginalise and exclude disabled people.

The Standardisation Problem

One of the most important insights from both the film and from disability studies scholarship is what we might call the standardisation problem. AI systems work by reducing complex phenomena to categories and classifications. They require standardised inputs and produce standardised outputs. This is fundamentally at odds with the reality of human diversity, particularly the diversity that disability represents.

When Nehma is asked to label caterpillars as pests, she is being asked to fit a complex ecological relationship into a binary category. Either something is a pest or it is not. There is no space for nuance, no recognition that the same creature might be a pest in some contexts and beneficial in others, depending on the specific conditions and the goals of the farmer. The standardisation demanded by the AI training process eliminates this contextual variability.

Similarly, when we consider disability and AI, we can see how the demand for standardisation creates problems. A person’s disability is not a fixed category. It is contextual and variable. A person who is Deaf might be fully included in conversations when interpreters are present, but excluded when they are not. A person with chronic pain might be able to work on some days but not others, depending on their pain levels. A person who is neurodivergent might function well in some environments but struggle in others. A person with a mobility disability might be fully accessible to buildings with ramps but excluded from buildings with stairs.

Yet AI systems are built on the assumption of fixed categories. A person is either disabled or not. A person either has a particular diagnosis or does not. There is no space for the contextual, relational, variable nature of disability as experienced by disabled people themselves. The result is that AI systems often fail to serve disabled people well because they are built on categorical assumptions that do not match the reality of how disability actually works.

This is related to what some scholars call the “universality problem” in accessibility and inclusive design. The assumption behind much accessibility work is that if we design for people with disabilities, we will end up with systems that work for everyone. This is true to some extent—many accessibility features, like captions or keyboard navigation, are useful for non-disabled people too. But it is not universally true. Sometimes, designing for disability requires trade-offs or choices that might not be ideal for all users. A system designed to be navigable via keyboard might be slower or more cumbersome than a system designed for mouse navigation. A system designed to accommodate people with severe anxiety might have different interface design than a system optimised for efficiency and speed.

This means that truly inclusive AI systems cannot simply aim for a single, universal design that works equally well for everyone. Instead, they need to be designed for flexibility and variability, with the ability to adapt to different users’ needs and preferences. They need to recognise that disability is not a problem to be solved but a form of human variation to be accommodated and, in many cases, valued.

From Exploitation to Justice: Reimagining Human-in-the-Loop Systems

If we accept the analysis that current human-in-the-loop systems reproduce both labour exploitation and ableist design, the question becomes: what would it look like to build human-in-the-loop systems oriented toward justice instead?

First, it would require radically reconceiving the labour relationship. Instead of workers like Nehma being hired as interchangeable data-labelling units, they would be recognised as knowledge producers and designers. They would be compensated fairly for their labour, at rates that recognise the cognitive and emotional demands of the work. They would have job security, benefits, and the possibility of advancement. They would not be treated as expendable, to be discarded when the data labelling phase is complete.

More radically, it would require centering the workers’ knowledge and preferences. If Nehma knows that caterpillars are not simply pests, that knowledge should be incorporated into how the AI system is designed. Rather than forcing her to suppress her knowledge to fit predetermined categories, the system should be redesigned to accommodate contextual knowledge. Perhaps the categories would be more nuanced, allowing for context-dependent classifications. Perhaps the system would include explanatory data allowing AI systems to learn not just that caterpillars are sometimes beneficial but why. Perhaps the system would include the labour and knowledge of farmers and agricultural workers as central to the design process, not as something to be extracted and then discarded.

Similarly, a justice-oriented human-in-the-loop system working on disability would not extract disabled people’s data and knowledge only to produce systems designed to identify and “correct” disability. Instead, it would involve disabled people as co-designers and decision-makers about what the system is designed to do and how it operates. It would centre disability justice principles: the leadership of disabled people, particularly disabled people of colour and disabled people from marginalised communities; the recognition that disability is a form of human difference, not a problem to be solved; and the commitment to building systems that support disabled people’s agency and self-determination.

This would mean fundamentally different human-in-the-loop systems. Rather than workers labelling data according to predetermined categories, disabled people would be involved in determining what categories matter and why. Rather than AI systems being trained to predict and identify disability, they might be trained to support disabled people’s needs, to accommodate disability in work and educational settings, to make information and opportunities more accessible. Rather than the output of human-in-the-loop systems being surveillance and control, it would be support and enablement.

Conclusion: Toward Disability-Centred AI and Labour Justice

The film Humans in the Loop offers a window into a world that usually remains hidden: the labour through which AI systems are built, and the human and epistemic costs of that labour. By combining the film’s narrative with critical analysis from disability studies, labour studies, and science and technology studies, we have argued that the mechanisms through which marginalised workers like Nehma are exploited are deeply connected to the mechanisms through which AI systems encode and perpetuate ableism and other forms of discrimination.

The key insight is that the problem is not simply that AI systems sometimes exclude disabled people or that workers are sometimes underpaid. Rather, the problem is more fundamental: it lies in how AI systems are currently conceived, designed, and developed. They are built on ableist assumptions about normalcy, standardisation, and the uniformity of human experience. They are built through processes that extract labour and knowledge from marginalised communities while excluding those communities from power and decision-making. They are built to serve the interests of powerful technology companies rather than the communities that will be affected by them.

Addressing these problems requires changes at multiple levels. At the level of individual technology use, disability-smart prompts and other tactical interventions can push back against biases. At the level of organisational practice, hiring diverse teams of developers including disabled people, centering disabled people in design processes, and collecting and centering disability-inclusive data can improve AI systems. At the policy level, regulations mandating inclusive design, accountability mechanisms for algorithmic harm, and protections for data workers can create structural changes.

But more fundamentally, addressing these problems requires a shift in how we think about AI and what it is for. Rather than thinking of AI as a tool to increase efficiency, eliminate variability, and standardise processes, we might think of AI as a tool that could support human flourishing, accommodate human diversity, and enable rather than constrain human agency. Rather than thinking of AI development as something that happens in labs and corporate offices and is then deployed to passive users, we might think of it as a fundamentally social and political process that should involve the people most affected by it.

For disabled people, this means centering disability justice in how we approach AI. It means insisting that disabled people are involved in decisions about AI systems that affect us, not simply as subjects to be helped or fixed but as experts and designers. It means asking not how AI can cure or correct disability, but how AI can support disabled people’s autonomy, participation, and flourishing as disabled people. It means recognising that the knowledge disabled people have about navigating the world, about accessibility, about designing for human diversity, is valuable and should shape how technology is built.

For workers like Nehma, this means recognising the labour of data annotation and AI training as valuable knowledge work that should be compensated fairly and that should be involved in decision-making about what is being built. It means not simply extracting labour from marginalised communities but building relationships of genuine collaboration and mutual respect. It means recognising that workers’ knowledge, including their contextual, embodied, place-based knowledge, is crucial to building AI systems that actually serve human needs.

Bringing these together—disability justice and labour justice, the insights of Humans in the Loop and the frameworks developed by disability scholars and activists—we arrive at a vision of AI that is fundamentally different from what currently exists. It is an AI that is inclusive not as an afterthought but from the ground up. It is an AI built through genuinely collaborative processes that respect the knowledge and leadership of those most affected by it. It is an AI oriented toward justice rather than profit maximisation, toward supporting human flourishing rather than increasing efficiency, toward accommodating human diversity rather than standardising it away.

This vision might seem utopian given the current state of AI development, dominated by large technology corporations oriented toward profit and control. Yet Aranya Sahay’s film reminds us that another world is possible. By documenting the reality of data labelling work, by centring the humanity and knowledge of workers like Nehma, by asking difficult questions about what is being built and to what ends, the film opens space for imagining different possibilities. Disability justice movements, labour movements, and communities demanding accountability from technology companies are all working toward these different possibilities. The work is hard, progress is slow, and power is unevenly distributed. Yet the possibility of building AI systems oriented toward justice, inclusion, and genuine accessibility remains. This is the work of our time: to ensure that as AI systems increasingly shape our world, they are built in ways that centre rather than marginalise disabled people, that recognise and value rather than extract labour from marginalised communities, and that serve human flourishing rather than corporate profit.

References

• Aboulafia, A., Bogen, M., & Swenor, B. (2024). To Reduce Bias in AI, Start with Disability Data. Center for Democracy and Technology.

• Berne, P., et al. (2018). Sins Invalid: An Uncompromising Critique of Ableism and Racism. AK Press.

• Bragg, D., Safar, I., & Wobbrock, J. O. (2019). Facilitating Deaf-Sighted Collaboration in Online Spaces. In Proceedings of CSCW 2019.

• Crenshaw, K. (1991). Mapping the Margins: Intersectionality, Identity Politics, and Violence against Women of Color. Stanford Law Review, 43(6), 1241–1299.

• Erard, M. (2017). Talking about Sign Language Translation. Translation Studies, 10(1), 23–39.

• Foley, A. (2025). Disabling AI: Power, Exclusion, and Disability. Journal of Educational Psychology, 117(3), 234–251.

• Fricker, M. (2007). Epistemic Injustice: Power and the Ethics of Knowing. Oxford University Press.

• Garland-Thomson, R. (2012). The Case for Conserving Disability. Journal of Bioethical Inquiry, 9(3), 339–355.

• Gardner, K. B., et al. (2021). User Experiences of AI-Enabled Accessibility Tools. ACM Transactions on Computer-Human Interaction, 28(4), 1–28.

• Kafer, A. (2013). Feminist, Queer, Crip. Oxford University Press.

• Kamikubo, R., Ueoka, R., & Yamashita, N. (2022). Accessibility Challenges of AI-Based Hiring Systems. In Conference Proceedings of Accessible Design.

• McRuer, R. (2006). Crip Theory: Cultural Signs of Queerness and Disability. New York University Press.

• Mehrotra, K. (2023). Human Touch. FiftyTwo Magazine.

• Newman-Griffis, D., et al. (2023). Definition Drives Design: Disability Models and Mechanisms of Bias in AI Technologies. First Monday, 28(9).

• Noble, S. U. (2018). Algorithms of Oppression. New York University Press.

• Price, M. (2011). Mad at School: Rhetorics of Mental Disability and Academic Life. University of Michigan Press.

• Sahay, A. (Director). (2024). Humans in the Loop [Film]. Storiculture.

• Shakespeare, T., et al. (2009). Disability and Development: The Relevance of the Social Model in the Global South. The Lancet, 374(9704), 1821–1830.

• Shew, A. (2020). Ableism, Technoableism, and the Future of Accessibility. Disability Visibility Project.

• Stramondo, J. A. (2019). Impairment and Disability: Philosophical Perspectives. Philosophical Psychology, 32(4), 529–548.

• Swenor, B. K., et al. (2024). Disability Data in AI Development. Health Affairs, 43(6), 812–821.

• Williams, K., & Gilbert, P. (2019). Disability and the Brain: Neurodiversity and Neurological Conditions. Progress in Brain Research, 248, 395–421.