The labellers in Nairobi who taught ChatGPT to refuse
On 18 January 2023 TIME published an investigation: the toxic-content filter for ChatGPT was trained on text labelled by workers at the contractor Sama in Nairobi for $1.32 to $2 an hour. OpenAI paid Sama $12.50 for that same hour.
Why it matters
Until then the work that makes a model safe was known only to exist. Here contracts, payslips and rates appeared for the first time - and it became visible that "alignment" has an address, a shift and an hourly rate.
From November 2021 OpenAI sent Sama tens of thousands of text snippets describing violence, suicide, incest and child sexual abuse, so that a detector could be trained on labelled examples. Three contracts worth about $200,000 in total, around three dozen workers in three teams. The rates: a junior labeller had a basic salary of 21,000 Kenyan shillings ($170) a month plus a monthly bonus of around $70 for the nature of the material, which came to at least $1.32 an hour after tax and up to $1.44 when targets were exceeded; a quality analyst could reach $2. Kenya has no universal minimum wage; the minimum for a receptionist in Nairobi at the time was $1.52 an hour. The two sides give different numbers, and the record does not blend them. Employees told TIME they were expected to label 150 to 250 passages per nine-hour shift; Sama replied that the figure was 70 passages, and that take-home pay ran from $1.46 to $3.74 an hour, declining to say which roles reach the top of that range. OpenAI said it set no productivity targets at all. Sama cancelled all its work for OpenAI in February 2022, eight months earlier than planned.