Something about the phrase “clean data” deserves more examination. Not because it is technically wrong, but because it conceals with such efficiency what happened before the cleaning, before the labeling that turned raw text and images into something a model could learn from. Data sets described as ethically sourced and governance-ready have, in a growing number of documented cases, been assembled on the work of people in Kenya, the Philippines, Venezuela, and Romania: annotators paid a few dollars an hour to label images, filter graphic content, and classify the kinds of material that language models would prefer not to explain having ingested.
Most structured AI advisory and data consulting engagements tend to position their work downstream of the sourcing question. Companies offering data governance consulting services to enterprise clients typically begin their work at ingestion: lineage tracking, access controls, schema design, and retention rules. Who created the raw material upstream, at what wages and under what conditions, rarely surfaces in a scoping conversation. Which means it rarely surfaces at all.
The Anatomy of a Quiet Extraction
There is a word for this dynamic that makes boardrooms go quiet: colonialism. Not the flag-planting variety. The quieter version, where value moves reliably from one region toward another, the people doing foundational work receive very little of it back, and the parties benefiting rarely examine the mechanism closely. The AI Now Institute found that over 60% of the digital labor supporting AI annotation in Western commercial markets is performed in countries the World Bank classifies as lower-middle income, with median hourly pay below $3.20. The report does not use the word colonialism, but others have started to.
The structural logic is recognizable from earlier extraction economies. A technology company in San Francisco needs labeled training data at scale. A vendor in Manila or Nairobi provides it. That vendor pays workers wages that are legal, sometimes competitive by local standards, and entirely invisible to the end-client’s ethics report. By the time the finished data set arrives inside a product marketed as auditable and governance-ready, the labor behind it has passed through three or four contract layers. The procurement team approved a primary vendor. That vendor approved a subcontractor. The subcontractor hired the annotators.
What makes the pattern hard to interrupt is that models trained on this labor frequently end up inside the governance systems affecting the daily lives of the same communities whose annotation work produced the training data. Loan-approval systems deny credit in the same geographies where the annotation was done. Content moderation classifiers trained by workers in Lagos help determine which posts from Lagos get removed.
The circularity is striking. Also, almost certainly, not an accident. Researchers working in this space have converged on a term for the underlying dynamic: digital labor exploitation. It describes the extraction of cognitive work from people who remain systematically excluded from the economic benefits of the technology their labor helped produce. The annotation worker in Nairobi is not a peripheral figure in the AI story. She is, in the most concrete sense, foundational to it. The phrase sounds academic. The phenomenon is not.
Building an Ethical Wall That Actually Holds
The phrase “ethical AI” has appeared in enough white papers and keynote talks to have drained itself of meaning. What an actual ethical wall requires is procedural, not rhetorical. The work lives in procurement conversations and vendor contracts, often long before any model architecture is on the table. Policies adopted at the board level do not survive contact with a procurement process that has no mechanism to verify them.
A small number of firms operating in the data governance consulting services space have started treating supply chain ethics as a condition of engagement rather than a voluntary disclosure. The demands they make tend to be specific:
None of this is expensive in absolute terms. Some firms, N-iX among them, have built supplier ethics reviews into pre-project scoping as standard practice across their data governance work. The resistance tends to be cultural rather than logistical: a settled assumption, rarely examined, that data provenance is a sourcing department concern and not a governance one.
MIT Technology Rating found that fewer than 12% of companies with formal AI governance programs include any downstream labor verification for training data. The gap between governance policy as written and as practiced is, by that count, almost total.
The legal environment is tightening. The EU AI Act, now in phased enforcement, establishes disclosure obligations for certain categories of training data. According to the World Economic Forum, responsible sourcing standards will become a baseline expectation for AI suppliers in regulated markets by 2027. Not yet a mandate everywhere. But the direction is not ambiguous, and organizations treating a compliance deadline as the earliest possible prompt to act are probably miscalculating.
The firms building something durable are already asking the upstream question in scoping calls and vendor contracts. What those firms tend to discover: annotation completed by people with adequate time, stable and predictable pay, and feedback channels that actually get responses is more accurately labeled and more defensible under audit. Better conditions produce better data. The ethical argument and the accuracy argument converge. In most documented cases, they are the same argument.
Conclusion
Data colonialism is not a metaphor, as the workers are real, the wage gaps are documented, and the arrangement’s structural invisibility was built in, not left by oversight. What remains open is whether the enterprise AI governance sector will treat sourcing accountability as a substantive professional standard or keep it as decorative language in client presentations. The organizations asking the upstream question now will be better positioned for the regulatory environment ahead. They will also have built something closer to what their governance materials actually claim.