How India's Labour Laws Fail To Protect Workers Behind AI Content Moderation

India's moderators power global AI giants and pay for it with their mental health, which the law has yet to catch up.

Every time an artificial intelligence (AI) system declines to generate violent content or correctly identifies a graphic image, it is doing so because a human being taught it to. That human is often a young woman in a small town in Jharkhand or Uttar Pradesh, working from a bedroom or a verandah, reviewing hundreds of disturbing videos and images a day for a contractor she will never meet, serving a tech company she may never have heard of. The computer programs (algorithms) that power the internet do not learn in a vacuum; they are meticulously trained, refined, and filtered by human intelligence. In India, a vast invisible workforce of data annotators (people who label images and videos to teach AI systems) and content moderators (people who review and filter harmful online content) serves as the human backbone of this global AI supply chain. This is the data annotation and content moderation industry, the unglamorous but essential underbelly of the global AI supply chain. By 2021, an estimated 70,000 workers in India were employed in data annotation alone, generating roughly 250 million dollars annually for the sector, with nearly sixty percent of that revenue flowing from the United States. The numbers have only grown since, and yet, these workers remain largely invisible to consumers, to policymakers, and disturbingly, to the law.

The invisibility is not accidental but structural. Workers are hired under vague labels like data processor or content reviewer. The job listings on YouTube and Telegram channels advertise the work as flexible, zero-investment, and safe work from home. It is only after contracts are signed that the actual nature of the work becomes apparent - classifying child sexual abuse material (CSAM), annotating footage of torture, and categorising pornographic content for hours on end.

The psychological toll of this work is devastating. Content moderators are routinely subjected to a barrage of digital toxicity. Some workers view up to 800 images and videos daily, frequently witnessing brutal acts of violence. Researchers classify content moderation as dangerous work comparable to lethal physical industries, with workers exhibiting acute traumatic stress, emotional numbing, dissociation (a mental state where one feels detached from reality), and severe sleep disturbances. This psychological trauma is deliberately obscured by strict Non-Disclosure Agreements. Workers are legally bound not to discuss the specifics of their tasks with anyone, including their families.

A Critique of the New Labour Codes

However, as India transitions into a new era of labour laws with the notification of the four consolidated Labour Codes, this digital frontline remains dangerously unprotected. While the government champions its vision for 'Safe and Trusted AI' at the 2026 AI Impact Summit, the legal structure fails to grasp the psychological hazards and contractual exploitation inherent in AI training labour. The government has hailed these reforms as a progressive step toward universalising minimum wages and social security, but a deeper examination reveals that the Codes remain anchored to the physical paradigms of the industrial age.

The Occupational Safety, Health and Working Conditions Code 2020 (OSH Code) subsumes laws like the Factories Act, 1948, to ensure workplaces are free from hazards. However, its conceptualisation of a hazard is insufficient, as under the erstwhile Act, specific manufacturing processes were designated as hazardous, triggering mandatory safety committees and exposure limits. The OSH Code retains these provisions but restricts their application exclusively to chemical, toxic, or physical manufacturing. The cognitive hazard of processing millions of data points containing extreme human suffering is entirely omitted. The Third Schedule of the Code lists recognised occupational diseases (e.g., asbestosis (a lung disease from asbestos exposure), lead poisoning) but entirely ignores psychological and psychosocial hazards (risks to mental and emotional health) such as Post-Traumatic Stress Disorder (PTSD) or secondary traumatic stress (trauma caused by repeatedly witnessing others' suffering). By failing to classify psychological trauma as an occupational disease, the Code strips content moderators of statutory rights to medical compensation and mandatory employer-funded psychiatric care.

The Code on Social Security 2020 represents a conceptual shift by formally defining gig workers, platform workers (workers hired through digital apps or websites), and aggregators (companies that connect workers to clients online). However, this inclusion is a double-edged sword as by bringing their status as gig or platform workers, the Code normalises the misclassification of data annotators as independent contractors rather than full-time employees, allowing AI firms to evade standard employer liabilities such as severance and strict working-hour limits. Although the Code provides for general health and maternity benefits, it lacks any explicit mandate for mental health coverage or trauma.

The Code on Wages 2019 universalises minimum wage protections, yet it struggles to regulate the piece-rate (paid per task completed, not by the hour) and task-based compensation models prevalent in digital labour. While physical arduousness is a recognised metric for wage calculation, the psychological arduousness of reviewing child exploitation material commands no statutory premium, leaving the mental destruction of workers entirely uncompensated.

Since AI training firms classify annotators as third-party contractors or gig workers, these individuals are systematically excluded from the definition of an employee under the Industrial Relations Code 2020. This denies them the legal right to form recognised trade unions or collective bargaining (negotiating as a group with employers). Furthermore, outsourcing vendors routinely bypass the requirements to establish certified Standing Orders (which govern fair termination and disciplinary proceedings), citing jurisdictional ambiguity and flexible workforce classifications.

The exploitation of data annotators is not merely a labour rights grievance but a major public interest issue. Content moderators are the immune system of the digital public sphere. Their labour ensures the safety of democratic discourse and the commercial viability of AI models. Recent regulatory moves have paradoxically worsened their plight. The February 2026 amendments to the Intermediary Rules, 2021, force platforms to adopt stricter due diligence and faster takedown timelines for deepfakes (AI-generated fake videos or images) and unlawful content. While protecting the public, this translates to higher quotas, faster viewing speeds, and increased daily exposure to graphic content for human moderators.

However, court rulings are beginning to recognise this gap. In a landmark 2025 case, Sukdeb Saha v. State of Andhra Pradesh, the Supreme Court ruled that mental health and emotional well-being falls under the Right to Life under Article 21 of the Constitution. While in this case it applied to educational institutions, this constitutional recognition of mental health creates a binding precedent that must be read into labour laws.

A Workable Path Forward

As India publishes its AI Governance Guidelines 2026 and positions itself as a leader for the Global South, it cannot champion 'Safe and Trusted AI' while ignoring the trauma of the workers building it. To align the Labour Codes with the realities of the digital economy and Article 21, the third Schedule of the OSH Code must be amended to explicitly classify PTSD, secondary traumatic stress, and chronic anxiety arising from digital exposure as occupational diseases. The definition of 'hazardous processes' must be amended to include high-volume content moderation of violent or sexually abusive material, triggering mandatory, employer-funded clinical psychiatric care, not mere wellness coaching. Deceptive hiring practices where workers are recruited for data annotation and later shifted to graphic content moderation must be penalised under the OSH Code. Furthermore, while companies may protect proprietary algorithms, NDAs must be statutorily voided to the extent that they prohibit workers from discussing their working conditions, seeking medical or psychological help, or reporting labour violations.

The Code on Wages must mandate a statutory 'hazard pay' premium for workers moderating violent, abusive, or sexually explicit content. Additionally, piece-rate compensation models must be subjected to algorithmic audits (automated checks on digital systems) to ensure quotas do not force workers to consume toxic content at psychologically destructive speeds just to earn a minimum wage. The principle of 'Principal Employer' liability, meaning the main company bears legal responsibility for violations by its contractors, must be strictly enforced. Multinational technology giants developing the AI must be held jointly and severally liable for the occupational health violations committed by their local outsourcing vendors in India.

Views are personal.