Investigation: starting in November 2021, OpenAI contracted Kenya-based Sama to label violent and toxic content, paying workers $2/hour, to improve ChatGPT
Content warning: this story contains descriptions of sexual abuse — ChatGPT was hailed as one 2022's most impressive technological … Mastodon: @karengregory@mastodon.scot , @parismarx@mastodon.online , and @mhoye@mastodon.social Tweets: @hkesvani , @erikmbaker , @nkulw , @pt , @bgurley , @spcohn , @billyperrigo , @hare_brain , @neilturkewitz , and @billyperrigo Mastodon: Karen Gregory / @karengregory@mastodon.scot : When you teach ChatGPT, teach the whole story, teach labour, teach students to see the materiality and social costs of these tools: — “Exclusive: OpenAI Used Kenyan Workers on Less Than $2 Per Hour to Make ChatGPT Less Toxic” … Paris Marx / @parismarx@mastodon.online : ChatGPT gets treated like technological magic, but that ignores the humans behind the curtain that make it function. — OpenAI paid Sama to hire Kenyan workers at $1.32 to $2 an hour to review “child sexual abuse, bestiality, murder, suicide, torture, self harm, and incest” content. … @mhoye@mastodon.social : There's no other way to say this: OpenAI is extractive-resource colonialism, and the resource being extracted is human's mental health. That fancy new tool you think is so shiny, that's a blood diamond. — https://twitter.com/... … Tweets: @hkesvani : Make no mistake that the ‘employment opportunities’ that come out of AI really amounts to outsourcing and underpaying people to make it be less racist https://twitter.com/... Erik Baker / @erikmbaker : Behind every miraculous tech innovation there is an exploited low-wage workforce https://twitter.com/... Noah Kulwin / @nkulw : Sam altman is a scuzzy idiot who will get himself and everyone around him into trouble https://twitter.com/... Parker / @pt : @bgurley @spcohn The irony is this angsty story has been written a dozen times on moderating social, which never ends. Once you train the AI, you can use that to do the objectionable work for humans. This should be an optimistic piece. The wage thing is just dumb. Not even worth commenting on. Bill Gurley / @bgurley : @spcohn @pt For Time to publish such an angry piece without mentioning this fact is really sad. Steven Cohn / @spcohn : @pt $2/hour = $16 per 8 hour day = ~$320/week, which is 53% above the Living Wage in Kenya https://take-profit.org/... https://twitter.com/... Billy Perrigo / @billyperrigo : Their 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. (6/8) https://t.co/zLXfgGjIwM Stephanie Hare / @hare_brain : Technology is not neutral https://twitter.com/... Neil Turkewitz / @neilturkewitz : “They're impressive, but ChatGPT & other generative models are not magic—they rely on massive supply chains of human labor & scraped data, much of which is unattributed & used without consent.” @agstrait 👀 @timnitGebru @emilymbender @Abebab @mer__edith @merbroussard @ubiquity75 https://twitter.com/... Billy Perrigo / @billyperrigo : The purpose of their work? Well, without a filter over the top, ChatGPT would spew racism and sexism, just like its predecessor GPT-3. These Kenyan workers were helping OpenAI build that filter. (3/8) https://twitter.com/...
Context & Ripple Effects
The TIME investigation pulls back the curtain on how ChatGPT got its safety layer: starting in November 2021, OpenAI outsourced the labeling of violent and sexual content to Kenya-based Sama at roughly $1.32–$2/hour, meaning the product hailed as 2022's breakthrough was built partly on low-paid exposure to graphic material. That reframes OpenAI's own Q&A on training ChatGPT, which discussed fixing toxicity as a technical problem without surfacing who did the fixing.
The story also lands mid-arc with other coverage of AI's hidden inputs: The Verge's later look at the vast "tasker" underclass hired by firms like Scale AI shows Sama was not an outlier but one instance of an annotation industry, while the Washington Post's analysis of Google's C4 dataset showed what happens when that filtering is skipped — training corpora laced with 4chan, Kiwi Farms, and Stormfront content.
First-order effects
- OpenAI faces immediate reputational damage over wages and moderator harm, with its safety-filter pipeline now publicly tied to $2/hour Kenyan labor rather than presented as purely algorithmic.
- Sama's role as OpenAI's content-labeling contractor is exposed, putting its other big-tech client relationships and pricing under scrutiny.
Second-order effects
- Rival labs and annotation vendors like Scale AI are forced to defend or reprice their own labeling workforces, since the investigation sets a public benchmark for what "AI safety" actually costs in labor terms.
- Buyers of AI safety services gain leverage to demand disclosure of working conditions and pay rates across the annotation supply chain, pressuring vendors to compete on more than price per label.
Third-order effects
- If the pattern holds, the mostly invisible annotation workforce documented in coverage of the "tasker" underclass becomes a named constituency — inviting labor regulation, disclosure requirements, or wage standards specific to AI data work.
- The episode sharpens a structural tension already visible in the C4 dataset analysis: labs can either expose workers to raw toxicity to filter it, or ship models trained on unfiltered extremist content — making human moderation cost a permanent line item that shapes where AI companies locate this work.
The trend: Frontier AI's safety claims increasingly rest on a global low-wage annotation workforce whose pay and conditions are becoming a reputational and regulatory liability for the labs that depend on it.