The dirty secret of Indian women watching hours of abusive content to train AI

On a veranda in rural Jharkhand, a young woman balances her laptop on a mud ledge built into the wall of her family home, positioning herself in one of the few places where the mobile signal remains stable. Around her, the rhythms of village life unfold in familiar cadences, with utensils clinking in the kitchen, footsteps crossing the courtyard, and low voices drifting through the house. On her screen, however, an altogether different reality plays out, one saturated with brutality, sexual violence and humiliation. She is required to watch until the end, even when instinct demands she look away. This is not an aberration in the global technology economy. It is its concealed foundation. The woman, as reported by The Guardian, Monsumi Murmu, is part of a vast but largely invisible workforce in India tasked with content moderation and data annotation for major global technology platforms. Her role is to review and classify images, videos and text flagged by automated systems as potential violations of platform rules. In practical terms, this means exposure to graphic depictions of violence, fatal accidents, child sexual abuse material, and extreme pornography, sometimes up to eight hundred items in a single day. These judgments are then fed back into machine learning systems, refining algorithms that underpin artificial intelligence tools marketed across the world as efficient, intelligent and transformative. The moral paradox is stark. The celebrated breakthroughs of contemporary artificial intelligence depend not merely on computing power, but on the psychological endurance of poorly protected human beings. India has become central to this global AI supply chain. According to estimates cited by the National Association of Software and Service Companies, the industry body commonly known as Nasscom, approximately seventy thousand individuals were working in data annotation as early as 2021, within a domestic market valued at roughly two hundred and fifty million dollars. Around sixty per cent of revenue in this segment derives from clients in the United States, while only a fraction originates within India itself. The geopolitical asymmetry is unambiguous. Wealthy jurisdictions externalise the most psychologically corrosive tasks of digital governance to lower cost labour markets, capturing the profits while exporting the trauma. The regulatory framework in India has not kept pace with the realities of this work. The Occupational Safety, Health and Working Conditions Code 2020 consolidates several earlier statutes governing workplace safety, yet its practical enforcement in the context of remote digital labour remains limited. The legislation was drafted with traditional industrial settings in mind, not dispersed home based moderators operating under non disclosure agreements. Psychological injury is not robustly recognised as an occupational hazard in most sectors, and compensation mechanisms for mental harm remain underdeveloped. Although the Employees Compensation Act 1923 provides for compensation in cases of personal injury arising out of and in the course of employment, its application to chronic psychological trauma linked to content moderation is legally ambiguous and rarely tested. This regulatory vacuum is compounded by the contractual architecture of the global technology industry. Most Indian moderators are not directly employed by the multinational platforms whose content they review. Instead, they are engaged through third party vendors, subcontractors and outsourcing firms. These entities operate from smaller cities and semi rural hubs where rents are lower and first generation graduates seek stable income. Women form at least half of this workforce, and in some centres a majority. They are often recruited with assurances that the work is respectable, flexible and safe, particularly when performed from home. In communities where female labour force participation remains constrained by social norms, the promise of income without migration is compelling. Yet the legal reality is that these workers are typically engaged on fixed term contracts with stringent confidentiality clauses. Non disclosure agreements prevent them from discussing the nature of their work even with family members. Breach can lead to termination or potential legal action. From a labour law perspective, this arrangement fragments accountability. The multinational technology corporation retains control over platform standards and algorithmic design, while the intermediary vendor manages day to day employment relationships. When psychological harm occurs, responsibility becomes diffuse and contestable. The experience described by Murmu illustrates the cognitive toll. In the early months, she could not sleep, reporting that even when she closed her eyes she saw the screen loading. Images intruded into her dreams, and she experienced persistent anxiety. Over time, the shock gave way to emotional numbing. She describes a state of blankness rather than active distress. Academic research supports this trajectory. Studies of content moderators have documented intrusive thoughts, hyper vigilance, sleep disturbance and symptoms consistent with secondary traumatic stress. A study published in December of the previous year, including Indian workers, identified traumatic stress as the most pronounced psychological risk, even where limited workplace interventions were available. From a comparative legal perspective, it is instructive to consider developments elsewhere. In the United States, litigation against major social media companies by content moderators has alleged failure to provide adequate psychological safeguards. Settlements have included commitments to counselling services and enhanced screening protocols. In the European Union, the Digital Services Act imposes due diligence obligations on very large online platforms to assess and mitigate systemic risks, including risks to fundamental rights. While the Act focuses primarily on user protection and content governance, it has catalysed debate about the human cost of moderation. The United Kingdom Online Safety Act similarly creates duties of care in relation to harmful content, but it does not directly regulate the working conditions of those who review such content overseas. Consequently, the individuals who absorb the daily burden of digital sanitation often remain outside the effective reach of the legal systems that benefit from their labour. The situation also raises questions under international labour standards. India is a member of the International Labour Organization and has ratified several core conventions, including those concerning discrimination and forced labour. However, it has not ratified certain conventions addressing occupational safety and health in comprehensive form. The ILO has increasingly recognised psychosocial risks as legitimate occupational hazards, yet global enforcement mechanisms remain weak. When content moderation is framed as a low skill digital task rather than hazardous labour, it evades the moral urgency typically associated with dangerous industries such as mining or heavy manufacturing. The gendered dimension of this workforce is particularly troubling. Recruiters and managers often describe women as detail oriented and compliant, and more willing to accept home based work. This language reflects entrenched stereotypes that can intersect with caste and class hierarchies. A significant proportion of moderators come from Dalit and Adivasi communities, for whom digital employment may represent an upward economic shift compared with agricultural labour or mining. The narrative of gratitude can suppress dissent. When employment is scarce and socially transformative, workers may hesitate to articulate psychological harm or to demand structural reform. The case of Raina Singh, who transitioned from text based annotation to reviewing child sexual abuse material and later explicit pornography without clear prior disclosure, exposes another legal fault line. Informed consent is a foundational principle in employment relations, particularly where work involves exposure to disturbing material. If job descriptions are vague and substantive tasks are disclosed only after contracts are signed, the legitimacy of that consent is questionable. Under Indian contract law, misrepresentation or concealment of material facts can in theory render agreements voidable. In practice, however, individual workers rarely possess the resources to challenge multinational supply chains in court. International human rights law provides an additional lens. The United Nations Guiding Principles on Business and Human Rights articulate a corporate responsibility to respect human rights, including the right to just and favourable conditions of work under Article 23 of the Universal Declaration of Human Rights and the International Covenant on Economic, Social and Cultural Rights. Companies are expected to conduct human rights due diligence across their operations and supply chains. If psychological harm from foreseeable exposure to extreme content is not systematically assessed and mitigated, serious questions arise about compliance with these principles. While the Guiding Principles are not binding in the same manner as domestic statutes, they increasingly inform litigation, investor scrutiny and regulatory reform. The economic logic underpinning this system is transparent. Artificial intelligence models require vast quantities of labelled data to function effectively. Automated systems flag content, but human judgment remains essential in training and calibration. By locating this labour in rural and semi rural India, global firms benefit from lower wages, weaker enforcement of occupational health norms, and limited public scrutiny. The cost differential is significant. Monthly salaries of the equivalent of two hundred and sixty to three hundred and thirty pounds, while meaningful in local contexts, are a fraction of comparable wages in Europe or North America. The externalisation of risk is therefore not incidental but structural. The psychological consequences do not end when contracts expire. Singh reports lingering aversion and dissociation in intimate relationships long after leaving her role. Such delayed effects complicate traditional models of workplace injury, which often rely on clear temporal links between exposure and harm. If mental health deterioration manifests months or years later, evidentiary challenges increase. Without statutory recognition of psychological injury specific to digital content moderation, affected workers may struggle to obtain compensation or therapy funded by employers. Meanwhile, corporate narratives emphasise safety and purpose. Some managers frame the work as moral guardianship, suggesting that moderators perform a public service by removing harmful material. There is truth in this assertion, yet it cannot absolve companies of responsibility for the conditions under which that service is rendered. In sectors recognised as hazardous, such as emergency response or forensic investigation, extensive training, rotation policies and psychological support are standard. In the content moderation industry in India, only a minority of companies report providing structured mental health assistance, and even then the burden often rests on workers to seek help proactively. The broader geopolitical context further complicates accountability. As artificial intelligence becomes a strategic asset in international competition, states are reluctant to impose stringent regulations that might slow innovation. India has positioned itself as a key hub in the digital economy, promoting technology services exports and seeking to attract foreign investment. Regulatory assertiveness in relation to data annotation firms could be perceived as jeopardising competitiveness. This creates tension between economic policy objectives and labour protection imperatives. At a systemic level, the invisibility of these workers sustains the arrangement. Users in Europe, North America and elsewhere interact with sanitised platforms, rarely confronting the human cost embedded in each moderated feed. Investors celebrate efficiency gains driven by machine learning without examining the human inputs required to produce them. Policymakers debate artificial intelligence governance in terms of algorithmic bias and data privacy, while the psychological welfare of those who curate the training data remains peripheral. If the global community is serious about ethical artificial intelligence, then transparency regarding labour conditions must become integral to regulatory frameworks. Mandatory disclosure of supply chain practices, enforceable standards on psychosocial risk mitigation, and clear legal recognition of psychological injury are not radical propositions. They are logical extensions of existing occupational health principles into the digital age. Absent such reform, the current model will persist, in which young women in rural India absorb the trauma necessary to maintain the illusion of frictionless digital life. In the quiet moments after logging off, Murmu walks into the forest near her home and sits beneath the open sky, seeking stillness after hours of mediated brutality. She collects stones and paints geometric patterns on the walls, small acts of restoration in a system that offers little institutional care. Her coping strategies are deeply human. The industry that depends on her labour, however, remains disturbingly indifferent. The future of artificial intelligence is frequently described in terms of efficiency, scale and disruption. Rarely is it described in terms of the women who, through repeated exposure to the worst of humanity, train the machines that promise to transcend it. Until law and policy confront this dissonance directly, the digital revolution will continue to rest on a foundation of outsourced trauma and legally neglected harm.

News

Premium

Region

The dirty secret of Indian women watching hours of abusive content to train AI

Are you sure?