Executive summary

The report emphasises that AI systems remain fundamentally reliant on a global workforce performing tasks such as data annotation, reinforcement learning from human feedback (RLHF), and red-teaming. While awareness of labour rights issues in the AI data supply chain is growing, technology companies have not yet sufficiently integrated these concerns into their broader human rights programmes, responsible sourcing policies, or AI governance efforts. This assessment identifies actions companies can take to strengthen governance and sourcing practices, grounded in the UN Guiding Principles on Business and Human Rights (UNGPs).

Introduction

The development of AI systems relies on large-scale human labour. Since the early 2000s, millions of workers, primarily in the Global South, have worked as data labelers and cleaners. While the industry has created jobs, it has also faced scrutiny for exploitative wages and poor working conditions. Beyond human rights responsibilities, companies have business incentives to promote fair conditions, as data quality and model performance are directly linked to workforce well-being. Emerging regulations, such as the Corporate Sustainability Due Diligence Directive (CSDDD), increasingly require companies to address these risks.

Methodology

BSR’s Human Rights Assessment (HRA) methodology identifies impacts against internationally recognised rights, including those codified in the International Bill of Human Rights and the Core Conventions of the International Labour Organisation (ILO). This sector-wide assessment involved engagement with worker representatives, unions, civil society organisations, and academics. Due to the opaque and distributed nature of the industry, BSR did not conduct a formal prioritisation of impacts, instead describing how risks manifest across different workforce models and geographies.

The AI data supply chain

The supply chain consists of workers, vendors (including managed service providers, digital labour platforms, and specialist vendors), and buyers (model developers, downstream developers, and deployers). Human input is integrated across the AI lifecycle, from data procurement and pre-training to evaluation, fine-tuning, and deployment monitoring. While automation is reshaping repetitive tasks—with estimates suggesting more than 60 per cent of data used for AI applications in 2024 was synthetic—demand for human input remains strong, especially for evaluation and contextual understanding. Vendors like Sama and Accenture typically use the Business Process Outsourcing (BPO) model, whereas platforms like Clickworker match tasks with distributed gig workers.

Key observations

The market is becoming increasingly stratified, with demand rising for domain-specific experts in fields such as law and medicine. Despite this, low-wage data work remains prevalent; for instance, at least 200,000 African workers are active in the sector. Structural drivers of risk include buyer concentration, which pressures vendors on cost and speed, and unit-based pricing models that incentivise excessive workloads. Market stratification is deepening global inequalities, as lower-wage, precarious work is concentrated in the Global South. Additionally, the perceived transience of data work often discourages long-term investment in worker development, even as industry leaders expect humans to remain involved for decades.

Human rights impacts

The assessment identifies eight salient risk categories. Workers face precarious conditions, economic insecurity, and inadequate wages. In the Global South, workers face dramatically lower pay; BPO workers in Nairobi reported receiving less than $200 a month, barely covering rent and transport. Footnote evidence indicates that wages in the Global South can be roughly 90 per cent lower for work of equal skill compared to Western markets. Occupational health risks include PTSD and cognitive strain from exposure to harmful content, such as graphic violence or sexual activities. Privacy risks arise from excessive surveillance, such as keystroke logging and screen recording. Furthermore, platform workers are more likely to face restrictions on access to remedy due to account deactivations by algorithms without an opportunity to appeal.

Recommendations

The report provides recommendations for direct buyers and downstream actors across four priority areas. First, companies must establish visibility by mapping their supply chains and using screening questionnaires for vendors. Second, they should apply existing responsible sourcing practices, such as translating supplier codes of conduct to the AI data context and using vendor scorecards. Third, companies should address salient risks by incorporating fair contracting expectations and progressing toward living wages. Fourth, they must address structural drivers by reviewing commercial requirements, such as aggressive throughput targets. For example, instead of treating Average Handling Time (AHT) as a hard target, it should be used as a planning metric that accounts for breaks and difficult cases. Industry-wide collaboration is encouraged to establish shared grievance mechanisms and raise standards collectively.

Annex: relevant resources

The report highlights several industry resources, including the Fairwork Certification, the Aapti Institute’s Building Better Workplaces for Data Workers, and the Partnership on AI’s (PAI) Responsible Sourcing Guidelines. These tools assist companies in implementing the recommended due diligence and oversight frameworks.