Every capable model rests on an enormous amount of human labour: labelling images, ranking responses, writing demonstrations, and reviewing material to decide what a system should refuse. A substantial share of that work is contracted to companies operating in East and West Africa, and it is one of the larger technology employment stories on the continent.
Why the work is there
Wage differences, a large pool of English-speaking graduates, timezone overlap with Europe, and a well-established business process outsourcing sector that already had the buildings and the management. The work arrives through layers of subcontracting, which is why the company whose model is being improved is often several steps removed from the people improving it.
The conditions question
Reporting on content moderation and annotation work has repeatedly raised the same issues: performance targets that leave little time per item, exposure to disturbing material as a routine part of the job, limited psychological support, and pay that looks reasonable locally and modest against the value created. These are labour questions rather than AI questions, and the layers of subcontracting make accountability genuinely difficult to place.
The part that could be different
Annotation is skilled work. People who do it develop a real understanding of how systems fail, what edge cases look like, and how to specify a task precisely — which is exactly the expertise needed to build evaluation sets and to design products that work. Whether that expertise becomes a career or stays a piece-rate task depends on how the contracts are written and by whom.
The continent supplying much of the labour behind these systems has a stronger claim than most to a say in how they are built. Making that claim requires the work being visible in the first place.
