AI Governance Institute
← News
Research2026-10-26

AI Training Labor Map Links 13 Vendors and 31 Countries, Exposing Supply Chain Gap

What happened

Cunyan Ma, an undergraduate at Brown University, published an interactive power map as part of ongoing research into AI labor supply chains, as reported in Brown student maps the hidden global labor powering AI. The map links 13 data labeling companies to 141 named corporate customers across 31 countries. It makes visible the annotation workers who produce the labeled datasets used to train commercial AI models. Ma drew on existing public databases and original data collection, arguing that the opacity of these labor relationships is structural rather than incidental. The research frames the concentration of low-wage data work in the Global South as a form of AI colonialism. It raises direct questions about whether corporate AI supply chain disclosures accurately reflect these dependencies.

Why it matters

  • ·Enterprise AI governance programs typically conduct due diligence on direct model vendors but rarely extend that review to data annotation subcontractors. This map names 141 customer organizations, creating a public record that regulators, investors, and litigants can reference when assessing whether supply chain disclosures are adequate.
  • ·The EU AI Act (Regulation (EU) 2024/1689) requires providers of high-risk AI systems to document training data practices, including third-party data sources. If annotation labor conditions are not documented, compliance teams face a gap between what the regulation expects and what the vendor's data sheet actually shows.
  • ·ESG and investor disclosure programs that reference responsible AI sourcing now face a credibility test. A published, citable map of labor conditions in the annotation supply chain makes it harder to claim that human rights risk assessments have been completed without extending them to data labeling vendors.

Governance controls affected

DGC-001Training Data ProvenancePRC-001AI Vendor Due DiligencePRC-005AI Procurement Risk AssessmentBRD-004AI Governance ESG and Investor DisclosurePRC-003Third-Party AI Model Evaluation

What to do now

  • ☐Ask your AI model and platform vendors to identify which data annotation companies they used or contracted, and in which countries those workers operate.
  • ☐Review your vendor due diligence questionnaires to confirm they ask about subcontracted data labeling, not only direct data sources.
  • ☐Cross-reference your current AI vendors against the 141 customer organizations named in the Brown University map and flag any matches for deeper review.
  • ☐Check whether your ESG or human rights disclosure program covers data annotation labor in the AI supply chain, and update it if it does not.
  • ☐Confirm that your training data provenance documentation, for any models your organization trains or fine-tunes, captures the identity and location of annotation providers, not just the dataset name.

What to watch next

The EU AI Act's technical documentation requirements for high-risk systems are becoming more specific as enforcement guidance develops. Labor conditions in the training data supply chain are a foreseeable area of examiner interest. Human rights due diligence laws in Germany, France, and Norway already require large companies to assess supplier labor conditions, and AI annotation vendors are not clearly excluded from that scope. Compliance teams should also watch for investor-side pressure. As ESG frameworks mature, the absence of annotation labor disclosures in responsible AI reporting may attract the same scrutiny that opaque garment or mineral supply chains have faced in other sectors.

Stay ahead of stories like this

Get every Global AI governance development like this one, plus the rest of the week's developments. Every Thursday.

Free every Thursday. Unsubscribe anytime.

Related Coverage

Corporate Policy2026-10-23

Meta's Revised Terms Still Permit AI Training, Legal Experts Warn

Meta updated its business account terms in October 2026 to remove explicit references to using scraped content for AI 'training,' following publisher pushback. Legal experts quoted by Nieman Lab argue the remaining 'develop' language is broad enough to cover model training. They add that robots.txt preferences carry no contractual force under the revised terms.

Research2026-10-25

National AI Evaluation Programs Are Expanding Into Agentic Risk

A Brookings Institution report finds that national AI evaluation programs in Singapore, Japan, the UK, and elsewhere are moving beyond basic safety benchmarks. The new scope covers multilingual behavior, agentic risk, sensitive data leakage, fraud, and cybersecurity. Enterprises whose vendor due diligence relies on earlier, narrower testing norms now face a shifting standard for what a tested and safe AI system actually means.

Research2026-10-25

Anthropic's Formal Verification Research Raises the Software Assurance Bar

Anthropic published research showing that Claude can automatically generate mathematical proofs of correctness for more than 100 real-world open-source software libraries. These include memory-safety guarantees, at a scale previously considered impractical.