James' World 2

Richard Mathenge felt he’d landed the perfect role when he started training OpenAI’s GPT model in 2021. After years of working in customer service in Nairobi, Kenya, he was finally involved in something that felt meaningful and held a future for him. But the position left him scarred. For nine hours per day, five days a week, Mathenge led a team that taught the A.I. model about explicit content. The goal was to train it so it could keep such things away from users. Today, it remains stuck with him.

While at work, Mathenge and his team repeatedly viewed explicit text and labeled it for the model. They could categorize the content, the provenance of which was unclear, as child sexual abuse material, erotic sexual content, illegal, nonsexual, or some other options. Much of what they read horrified them. One passage, Mathenge said, described a father having sex with an animal in front of his child; others involved scenes of child rape. Some were so offensive Mathenge refused to speak of them. “Unimaginable,” he told me.

The type of work Mathenge performed has been crucial for bots like ChatGPT and Google’s Bard to function and to feel so magical. But the human cost of the effort been widely overlooked. In a process called “Reinforcement Learning from Human Feedback,” or RLHF, bots become smarter as humans label content, teaching them how to optimize based on that feedback. A.I. leaders, including OpenAI’s Sam Altman, have praised the practice’s technical effectiveness, yet they rarely talk about the cost some humans pay to align the A.I. systems with our values. Mathenge and his colleagues were on the business end of that reality.

Mathenge earned a degree from Nairobi’s Africa Nazarene University in 2018 and quickly got to work in the city’s technology sector. In 2021, he applied for work with Sama, an A.I. annotation service that’s worked for companies like OpenAI. After Sama hired Mathenge, it put him to work labeling LiDAR images for self-driving cars. He’d review the images and pick out people, other vehicles, and objects, helping the models better understand what they encountered on the road.

When that project wrapped, Mathenge was transferred to work on OpenAI’s models. And there, he encountered the disturbing texts. OpenAI told me it believed it was paying its Sama contractors $12.50 per hour, but Mathenge says he and his colleagues earned approximately $1 per hour and sometimes less. Spending their days steeped in depictions of incest, bestiality, and other explicit scenes, the team began growing withdrawn.