By- Aashita Sarin
Abstract
Artificial intelligence systems rely on large volumes of human-annotated data yet this labour is becoming increasingly fragmented. AI developers, outsourcing vendors, and digital platforms often perform different functions within the same production chain, separating the entity that commissions labour from the one that contracts with or oversees the worker. Who, then, is the employer responsible for statutory obligations? This article examines how Indian judicial tests for employer identification apply to AI data annotation. It argues that while existing legal principles remain applicable, fragmentation across multiple entities causes traditional factors to point toward different participants, making identification of the legally responsible employer considerably more complex.
Introduction
Artificial intelligence is rarely discussed as a labour question. Public debate tends to focus on increasingly capable models, larger datasets, and computing power, while the human labour required to train these systems remains largely invisible. However, no AI model reaches deployment without data annotation, the labour of labelling, classifying, correcting, and evaluating data so machine learning systems can recognise patterns and produce meaningful outputs. In India, this work has grown into an industry projected to exceed US$7 billion by 2030, yet it has received comparatively little attention in labour law scholarship.
One reason lies in how annotation work is organised. It is rarely performed through a direct relationship between the worker and the company that ultimately benefits from the labour. AI developers commonly obtain annotated datasets through outsourcing vendors, who may in turn, recruit workers through digital platforms or subcontracting arrangements. For the annotator, however, these commercial arrangements are largely invisible; what changes is not the work performed, but the legal relationship through which it is organised.
Indian labour law has confronted comparable questions in industries such as manufacturing and tailoring, where work is commonly organised through contractors and intermediaries. Comparable questions have yet to be examined in the context of AI production. More fundamentally, the question arises before worker classification: once an employment relationship exists, who bears the statutory obligations attached to it? Courts identifying an employer look beyond contractual labels to the substance of the relationship. Where engagement, supervision, remuneration, and commercial benefit are dispersed across different entities, those factors may no longer point to the same participant.
Fragmented Labour, Fragmented Responsibility
Every AI model rests on extensive human labour. To train a system to recognise images, understand language or generate text, it has to be trained on data labelled, classified, corrected, and evaluated by human annotators.
Many annotators never contract with the developer at all. Companies commonly obtain annotated datasets from specialised vendors, who recruit workers through digital platforms or subcontracting. Traditionally, a single employer recruits workers, directs how work is performed, pays wages, and derives the commercial benefit from the labour. AI production can divide those functions: a developer determines annotation guidelines, a vendor hires and pays the annotators, and a platform allocates tasks. The employer has not disappeared from the relationship; its functions have been split across multiple entities. Identifying the employer responsible for the statutory obligations imposed on employers under labour law becomes considerably more difficult once the relevant factors no longer converge on a single entity.
Identifying the Legally Responsible Employer
The Code on Social Security, 2020 defines an “employer” under section 2(27), but does not address how responsibility should be allocated where recruitment, supervision, payment, and commercial benefit are divided across multiple entities. The question is, therefore, left to judicial principles, under which Indian labour law does not identify an employer by contractual label alone; whether a worker is called a freelancer, consultant, or independent contractor is not determinative. Courts instead examine the substance of the relationship through judicial tests developed when employer functions were ordinarily exercised by a single entity. AI data annotation presents a markedly different arrangement.
The control test, from Dharangadhara Chemical Works Ltd. v. State of Saurashtra, asks which entity directs how work is performed. In annotation, project instructions may originate with the developer, while contractual oversight sits with a vendor and task allocation with a platform. A similar problem arises under the integration test, articulated in Silver Jubilee Tailoring House v. Chief Inspector of Shops, and reaffirmed in Hussainbhai v. Alath Factory Thezhilali Union. An annotator’s work is indispensable to a developer’s business, yet the annotator may contract only with an intermediary. Integration points one way. Contract points another.
Courts increasingly address this through a multifactor approach, from Workmen of Nilgiri Co-operative Marketing Society v. State of Tamil Nadu, and reaffirmed in General Manager, U.P. Cooperative Bank Ltd. v. Achchey Lal (2025). Rather than treating any single factor as conclusive, courts weigh supervision, pay, contracts, and economic reality together, an approach well suited to complex relationships. In annotation, fragmentation means these factors no longer converge on a single participant.
Existing principles remain applicable, what has changed is the organisation of the work to which they must now be applied.
The Consequences of an Uncertain Employer
Research from SOMO illustrates that the problem is not merely theoretical: Amazon, Google, Meta, Microsoft, and Nvidia collectively rely on at least thirty intermediary firms for annotation, three of which declined in 2025 to disclose their vendors. The International Labour Organization has similarly observed that emerging digital technologies are reshaping work organisation and creating new challenges for the application of existing labour law in India. Where developers contract through multiple intermediaries, who further subcontract through platforms, the relationship between the entity benefiting from the work and the entity engaging the worker becomes hard to trace. Employer identification determines who bears statutory obligations, including wages under the Payment of Wages Act, 1936, social security contributions, maternity benefits, gratuity, occupational safety, and termination liability. These rights are only as effective as the ability to identify who they attach to.
Such arrangements let each participant define its role narrowly. A developer may characterise itself as a purchaser of services, a vendor as an independent contractor, a platform a mere intermediary. These labels govern relationships between businesses, but do not decide labour law obligations, which are a matter of substance, not the terminology parties use.
More importantly, this question is conceptually distinct from worker classification. Worker classification asks whether an individual is an employee or an independent contractor. Employer identification assumes that an employment relationship exists and instead asks which entity bears the legal obligations attached to that relationship. In conventional employment, these questions often converge because a single employer recruits, supervises, pays, and benefits from the worker’s labour. AI annotation disrupts that assumption by distributing those functions across developers, vendors, and platforms. The legal challenge therefore lies not in recognising employment itself, but in determining which participant should be treated as the employer for the purposes of labour law.
Conclusion
Data annotation demonstrates that the organisation of AI labour has changed as significantly as the technologies it supports. Functions once held by a single employer, are now distributed across developers, vendors, and platforms, without the employment relationship itself disappearing. Indian labour law already provides tests for identifying an employer, and they remain applicable; the difficulty lies in applying them where factors no longer converge on one entity. The architecture of AI has not made labour invisible. It has made responsibility difficult to locate.
About the author
Aashita Sarin is a second-year undergraduate psychology student at O.P. Jindal Global University. She is interested in artificial intelligence, large language models, and how technology influences the way people think, behave, and make decisions. Her work focuses on the ethical, cognitive, and social impact of AI, especially the role of intelligent systems in everyday life and public decision-making.