TIME has released an exclusive report showing that OpenAI outsourced data labeling work to Kenyan workers earning less than $2 per hour in order to build its content filter for ChatGPT.
According to author Billy Perrigo, the data labelers were tasked with reading and labeling text pulled from the “darkest recesses of the internet” that “described situations in graphic detail like child sexual abuse, bestiality, murder, suicide, torture, self-harm, and incest.”
Perrigo says that the lowest paid among these workers earned a minimum of $1.32 after tax with a maximum of $1.44 after tax per hour, based on seniority and performance. A BBC report from 2018 estimated that around 75% of Sama’s workers in Kenya live in the Nairobi slum of Kibera, one of the largest slums in Africa. The area faces a high unemployment rate of 50% along with a lack of clean water and no sanitation facilities.
Behind the outsourced work was Sama (formerly Samasource), a San Francisco-based firm that markets itself as an “ethical AI” company, saying its mission is to connect low-income people to digital work. The company employs people in Kenya, Uganda, and India and has provided data labeling services for companies such as Google, Meta, and Microsoft.
ChatGPT is OpenAI’s chatbot which is an upgraded version of its large language model, GPT-3.5. The chatbot debuted in late November to great fanfare, and it saw over one million users less than a week after it was launched. ChatGPT’s use cases include digital content creation, writing, and debugging code. Like its GPT-3.5 predecessor, the system is prone to toxic content due to training data sourced from the open internet, and to mitigate this, OpenAI has developed content filters.
OpenAI confirmed that Sama’s employees in Kenya were contributing to its toxic content filtering for ChatGPT. “Our mission is to ensure artificial general intelligence benefits all of humanity, and we work hard to build safe and useful AI systems that limit bias and harmful content,” a spokesperson for the company told TIME. “Classifying and filtering harmful [text and images] is a necessary step in minimizing the amount of violent and sexual content included in training data and creating tools that can detect harmful content.”
TIME interviewed four Sama workers for the story who described feeling “mentally scarred” by the work they performed. Three workers told TIME they were expected to read between 150 and 250 snippets of explicit material each day, though Sama disputes this, saying they were only required to review 70 passages per day.
TIME asserts that the majority of Sama’s three teams of data labelers were paid $170 per month with bonuses of up to $70 and there were opportunities to earn additional commission based on accuracy and speed. An agent working nine-hour shifts could expect to take home a total of at least $1.32 per hour after tax, rising to as high as $1.44 per hour if they exceeded all their targets, according to the report. Perrigo’s story noted that there is no universal wage in Kenya, but that at the time these workers were employed, the minimum wage for a receptionist in Nairobi was $1.52 per hour.
Perrigo said in a tweet that these “working conditions reveal a darker side to the AI boom: that AI often relies on hidden, low-paid human workers who remain on the margins even as their work contributes to a multibillion-dollar industry.” Microsoft is currently in talks with OpenAI for a possible $10 billion investment with the company with $1 billion already invested. The proposed deal could make OpenAI worth $29 billion.
Media outlet Quartz reported on the TIME story, and Sama reached out to them to clarify the pay, saying it offers nearly double the compensation as other content moderation firms in east Africa: “Sama pays between Sh26,600 and Sh40,000 ($210 to $323) per month, which is more than double the minimum wage in Kenya and also well above the living wage. A comparative US wage would be between $30 and $45 per hour,” the company told Quartz, adding that it also offered workers personal welfare services such as counseling, meditation, prayer, nursing, gaming, and local artwork “and full meal services that are curated to support physical and mental wellbeing.” The workers interviewed by TIME asserted that these counseling services were hard to access due to productivity demands, and two said they were only offered group sessions after being denied 1:1 sessions by Sama management.
Sama ended its contract with OpenAI in February 2022, eight months earlier than planned. Another data labeling project had begun, this time for graphic images that sometimes dealt with illegal content, according to TIME. OpenAI released a statement chalking up the collection of illegal images to miscommunication.
Sama announced on Jan. 10 that it will be canceling all remaining work having to do with sensitive content, including a $3.9 million content moderation contract with Facebook. The company will instead focus on computer vision data annotation.