Inside Samasource's office in Nairobi, Kenya, where employees from low-income backgrounds train data used for AI; Samasource's clients include Google, Microsoft
so much “innovation” is powered by the hidden labor of the poor http://reallifemag.com/... http://twitter.com/... Tim Maughan / @timmaughan : 'When Artificial Intelligence works as intended, Silicon Valley types often say it's “like magic”. But it isn't magic. It's Brenda, a 26-year-old single mother who lives Kibera, Africa's largest slum, and perhaps the toughest neighbourhood on earth' http://www.bbc.com/... Hayden Lewis / @haydenrlewis : Excellent storytelling highlighting the widespread misconception that AI-driven products are sheer marvels of human ingenuity — rather than human labor. Left pondering: At what point does this iteration of the digital race to the bottom become oppressive rather than liberating? http://twitter.com/... Chris Hooks / @cd_hooks : really incredible statement from a silicon valley exec on why they don't pay outsourced kenyan labor more. looking forward to the wave of self-administered pay cuts in the bay area http://www.bbc.com/... http://twitter.com/...
Context & Ripple Effects
This 2018 BBC visit to Samasource's Nairobi office put faces on the annotation floor — workers like Brenda from Kibera labeling the data behind AI products sold by clients including Google and Microsoft — at a moment when the industry still marketed those products as 'magic'.
The years since have validated the piece's premise rather than dated it: TIME's investigation into Sama's contract with OpenAI found Kenyan workers paid $2/hour to label toxic content for ChatGPT, while reporting on Scale AI's 'tasker' underclass showed the hidden workforce now includes subject-matter experts, not just entry-level labelers.
First-order effects
- Samasource's low-income Nairobi employees perform the labeling work that directly determines the quality of AI products shipped by Google and Microsoft, making their wages and conditions a live reputational exposure for those clients.
Second-order effects
- The pay-and-power gap exposed here has already forced alternatives into the market: India's nonprofit Karya sells training data but redirects all profit to workers who keep ownership of what they create, positioning worker ownership as a competitive answer to the $2/hour contracting model.
Third-order effects
- If the pattern holds, AI's training-data supply chain becomes an audited tier-one supplier relationship like conflict minerals — yet the emergence of opaque middleman networks hiring Kenyan annotators via WhatsApp shows accountability can fragment faster than scrutiny can spread.
The trend: The human labor inside AI training is moving from invisible cost line to named supply-chain liability, with buyer scrutiny, worker-ownership models, and off-the-books intermediaries all competing to define its terms.