Kenyan Data Workers and AI Safety Labeling

AI safety labels and evaluations may be produced by workers employed through outsourcing firms, staffing agencies, or digital labor platforms. In January 2023, TIME reported that Sama workers in Nairobi reviewed and labeled text descriptions of sexual abuse, violence, and hate speech under contracts connected to OpenAI safety work. TIME said the take-home pay it reviewed ranged from about $1.32 to $2 per hour depending on seniority and performance. Sama later gave TIME a different range—$1.46 to $3.74 after taxes—and said workers reviewed fewer passages per shift than the article reported. The figures are disputed and time-specific, not a measure of all Kenyan annotation work. TIME reported the work contributed to a tool for detecting toxic ChatGPT outputs. Labelers read disturbing text, apply categories and instructions, and provide feedback that can shape model behavior. This labor is part of AI safety infrastructure, but it can carry exposure and workload risks. Workers may be far from the product company’s offices and may be employed by a local contractor, creating questions about the roles of buyers, vendors, and employers. The case also sits within broader Kenyan disputes about content moderation, working conditions, union activity, and whether foreign platforms can be held accountable in Kenyan courts. Those cases involve particular parties and claims and should not be conflated with the separate OpenAI-related data-labeling contract. Allegations in lawsuits are not court findings unless a court has decided them. The history nonetheless highlights how supply-chain design can make labor conditions hard for AI developers and users to see. Responsible procurement should trace each task and contracting tier, disclose sensitive-content exposure accurately, verify pay and hours, assess mental-health safeguards, and give workers a confidential grievance route. Buyers should audit and remediate conditions rather than assuming that outsourcing transfers all responsibility. Reported conditions should be attributed to the source and period, especially when company responses differ.