The people who made ChatGPT polite aren't on the timeline
TIME reported that Kenyan workers labelled the internet's worst text for OpenAI through Sama, taking home about $1.32 to $2 an hour.
In January 2023, TIME published an investigation into how OpenAI made ChatGPT less toxic. From November 2021, OpenAI sent tens of thousands of text snippets to Sama, an outsourcing firm with workers in Kenya, so that people could label descriptions of violence, hate speech and sexual abuse. TIME reported that those labellers took home between about $1.32 and $2 an hour.
The reason for the job is simple enough. GPT-3, the model ChatGPT was built on, had been trained on hundreds of billions of words scraped from the internet, and TIME says it was prone to blurting out violent, sexist and racist remarks. Nobody could clean that dataset by hand. So OpenAI built a second model to spot toxic text, and a model like that only learns from examples a human has already read and labelled. The detector went into ChatGPT to filter what it says before users see it.
TIME reviewed internal documents, including payslips, and spoke to four Sama employees. It reported three contracts worth about $200,000 in total, signed in late 2021. Around three dozen workers were split into teams for sexual abuse, hate speech and violence. Three of them said they were expected to get through 150 to 250 passages in a nine-hour shift. Junior staff earned a basic 21,000 Kenyan shillings a month, about $170, plus a bonus of around $70 for the explicit nature of the work.
All four employees said the work left them mentally scarred, and that the "wellness" sessions were unhelpful and rare. One told TIME he had recurring visions after a single passage. "That was torture," he said.
OpenAI and Sama both responded, and their answers are in the piece. Read them alongside the workers' accounts.
My reaction is about who gets remembered. Every AI history timeline has the same cast: Turing, the Dartmouth crew, Deep Blue, AlphaGo, ChatGPT. The people who read the worst writing on the internet so a chatbot would stop repeating it are on none of them. Yet without that work, ChatGPT would have been the same difficult product GPT-3 was.
This is an old story, and I'm going only on TIME's reporting. I haven't checked what has happened since early 2023.
If you use ChatGPT at work, the polite refusal you get when you ask it for something nasty is partly their work. It is worth knowing whose.