AI provenance should include the people who labelled the data
AI buyers now ask where a model is hosted, which evaluations it passed, and whether its training data has documented sources. That record can still omit the people who made the model usable.
Data workers clean training sets, label images, transcribe audio, rank model answers, and inspect material that automated filters cannot interpret. Their decisions affect model behavior and safety. By the time a customer meets the model through an API, that work has usually vanished.
In Challenging the Myth of AI Autonomy, published in 2026, International Labour Organization researchers Uma Rani and Morgan Williams describe two connected workforces. "Algorithmic workers" design, tune, and debug systems. "Data workers" label, clean, and expand the data used across the lifecycle. Product launches celebrate the first group. Companies often outsource the second through business process outsourcing firms or digital platforms.
A provenance record that ends with datasets, weights, and data centres is missing a supplier. AI buyers also need to know how the human judgment was produced.
A label is a decision
"Data labeling" sounds mechanical. Much of it is judgment under a deadline.
Workers decide whether an image contains a pedestrian, whether a response follows a policy, or whether a sentence is hateful rather than satirical. Annotators rank model outputs for reinforcement learning from human feedback. Content moderators handle context that keyword filters miss.
The ILO paper connects the conditions of this work to technical quality. Weak training, speed targets, and low pay can put errors into the pipeline. In medical imaging, poorly specified labels or workers without the right expertise can cause "data cascades." An early defect then spreads through development and weakens the deployed system.
Vendor reviews usually put labor conduct in one questionnaire and model accuracy in another. The product does not respect that administrative split.
Degrees do not guarantee decent data work
The paper draws on ILO surveys conducted in India and Kenya in 2022 and 2023. About 55% to 56% of surveyed data workers in India and up to 50% in Kenya had studied science, technology, engineering, or mathematics. Many still performed repetitive validation, annotation, or moderation that did not use those qualifications.
A contract did not necessarily bring stability. BPO workers often had formal contracts while platform workers were commonly treated as independent contractors. Short client projects, unexplained deductions, and arbitrary dismissals still left workers insecure.
Nor was BPO work automatically better paid. Indian workers in the surveys earned US$3.90 per hour on microtask platforms and US$2.20 per hour in BPO roles. Kenyan workers in both sectors earned US$1.10 per hour. Many lacked benefits and social security. Some content moderators were expected to process a task every 50 seconds while reviewing graphic material.
Those numbers come from particular surveys, countries, and years. They are evidence of a procurement problem, not a global wage table. A supplier's promise to follow local law cannot tell a buyer whether workers have the time, training, and protection needed to make reliable decisions.
A human checkpoint can move the risk
Documentation often uses "human in the loop" as a safety claim. The phrase does not identify the worker, their training, their power to reject a bad instruction, or the penalty for missing a quota.
A reviewer paid by the task has an incentive to move quickly. A moderator restricted by a non-disclosure agreement may have trouble obtaining independent mental-health care. A medical annotator who lacks domain training may have no safe way to record uncertainty. The checkpoint can protect the product by transferring strain to a person.
The ILO and its partners put this supply-chain view at the centre of the INDL-9 conference on AI Supply Chains, scheduled for 9 to 11 September 2026. Its program covers worker health, responsible sourcing, vendor audits, traceability, and social dialogue. That list describes procurement work, not a separate conversation about AI ethics.
Put labor questions in the contract
An AI buyer should know which data tasks people perform and where they perform them. The contract should identify whether those workers are employees, BPO staff, or platform contractors. Effective hourly pay matters, including paid waiting time and the process for appealing an error score.
For domain-sensitive data, the supplier should document worker qualifications and escalation rules. Workers need a way to mark ambiguity without losing pay. Content moderation contracts should describe exposure limits, access to independent care, and protection after a project ends.
Subcontractors need the same scrutiny. Buyers should know which firms can access user data or evaluation prompts, and whether an independent audit can test the supplier's account.
These fields belong in procurement, security review, and evaluation design. A provider reached through api.ish.chat, another gateway, or a direct API remains part of the same labor chain. Routing changes access, not the history of training and evaluation.
Governments can set a stronger baseline. The World Bank's World Development Report 2026 recommends using public purchasing power to test, evaluate, and procure AI for public services. Labor disclosures can sit beside performance and security requirements. Public procurement already shapes accessibility and supply-chain conduct in other markets.
Follow the judgment, not only the file
BLOGish has argued that source provenance should survive edits and that platform workers need meaningful human review of account termination. Data work connects those concerns.
When a person's judgment changes a training example, safety policy, benchmark label, or model preference, the conditions around that judgment are part of the evidence. Buyers do not need workers' names. They need traceability for the work, its safeguards, and the supplier's claims.
An AI system can document every dataset and still hide the people who cleaned it. It can pass a safety evaluation assembled under quotas that punish uncertainty. In both cases, the missing labor record makes the technical evidence harder to trust.