Building Data Pipelines That Create Jobs

The labor underneath the model

Every large language model that exists today is built on human labor. Before a single token is generated, thousands of people have written text, labelled data, ranked outputs, corrected errors, and reviewed quality. The model is the visible product of an invisible workforce.

The AI data labour market is estimated to employ more than one million workers globally, the majority in the Global South. Kenya, Nigeria, Ghana, Uganda, and the Philippines are major hubs for data annotation, content moderation, and reinforcement learning from human feedback. Workers label images, rank chatbot responses, transcribe audio, and filter toxic content for pennies per task.

The economics of this market are deeply extractive. A data annotator in Nairobi earns an average of $1.50 to $3.00 per hour for work that directly improves the capabilities of models whose market capitalisation exceeds one trillion dollars. The annotator's labour becomes the model's intelligence, but the annotator owns none of it. The value flows in one direction: out of the continent.

What extraction looks like on the ground

The human cost of this model is well documented. A 2024 investigation by Time magazine revealed that OpenAI employed Kenyan workers earning less than $2 per hour to label toxic content for ChatGPT, including graphic descriptions of violence and abuse. Workers reported symptoms consistent with secondary trauma. Their labour made the model safer for Western users while exposing the workers to psychological harm with no ownership stake in the resulting product.

This is not unique to OpenAI. Every major AI company relies on a global supply chain of data labour that concentrates value at the top and distributes risk at the bottom. African workers are overrepresented in the lowest-paid, most psychologically demanding tiers of this supply chain.

The problem is structural, not incidental. The current model treats data as a raw material to be extracted and data workers as interchangeable labour. When the dataset is complete, the workers are dismissed and the value of the dataset accrues entirely to the company that commissioned it.

A different design

Kora Lab's data pipelines are designed on a different premise: the dataset is the asset, and the contributor is a co-owner.

Our approach has three components:

**Co-ownership agreements.** Every contributor to a Kora Lab dataset signs an agreement that grants them a proportional stake in the dataset's value. If the dataset is licensed to third parties, if it is used to train a commercial model, or if it contributes to a research publication, the contributors share in the resulting revenue or recognition.

**Fair baseline compensation.** Contributors receive a base rate that exceeds local living wages, not the global minimum. Our target is $8 to $12 per hour for annotation and curation work in African countries, benchmarked to comparable skilled remote work rather than to the local poverty line.

**Career pathways.** Data work at Kora Lab is designed as an entry point to the AI economy, not a terminal occupation. Contributors receive training in machine learning fundamentals, dataset design, and quality assurance. The goal is that every contributor who completes twelve months of data work is qualified to apply for a junior research or engineering role in the lab.

What this enables

This model turns a structural weakness into a strategic asset. Africa has the youngest population in the world, with a median age of 19. The continent produces millions of university graduates every year, many in STEM fields, but lacks the employment infrastructure to absorb them into the AI economy.

Data pipelines that pay fairly, train intentionally, and share ownership create a new economic sector. They transform AI data work from a low-wage extractive industry into a skilled knowledge profession that retains value on the continent.

For Kora Lab, this model also produces better data. Contributors who own a stake in the dataset produce higher-quality annotations, catch edge cases that hourly workers would ignore, and stay with the project long enough to develop expertise in specific language domains. A dataset built by co-owners is more accurate, more complete, and more culturally grounded than one built by anonymous contractors.

The larger vision

If Kora Lab's model succeeds at scale, it creates a precedent that extends beyond our own datasets. It demonstrates that there is an alternative to the extractive data labour model, one in which the people who build the datasets share in the value they create.

A generation of African contributors participating in the AI economy as builders, not just as workers. That is what we are designing for.