The secret to making AI chatbots sound smart and spew less toxic nonsense is to use a technique called reinforcement learning from human feedback, which uses input from people to improve the model’s answers. 

It relies on a small army of human data annotators who evaluate whether a string of text makes sense and sounds fluent and natural. They decide whether a response should be kept in the AI model’s database or removed. 

Even the most impressive AI chatbots require thousands of human work hours to behave in a way their creators want them to, and even then they do it unreliably. The work can be brutal and upsetting, as we will hear this week when the ACM Conference on Fairness, Accountability, and Transparency (FAccT) gets underway. It’s a conference that brings together research on things I like to write about, such as how to make AI systems more accountable and ethical.

One panel I am looking forward to is with AI ethics pioneer Timnit Gebru, who used to co-lead Google’s AI ethics department before being fired. Gebru will be speaking about how data workers in Ethiopia, Eritrea, and Kenya are exploited to clean up online hate and misinformation. Data annotators in Kenya, for example, were paid less than $2 an hour to sift through reams of unsettling content on violence and sexual abuse in order to make ChatGPT less toxic. These workers are now unionizing to gain better working conditions. 

In an MIT Technology Review series last year, we explored how AI is creating a new colonial world order, and data workers are bearing the brunt of it. Shining a light on exploitative labor practices around AI has become even more urgent and important with the rise of popular AI chatbots such as ChatGPT, Bing, and Bard and image-generating AI such as DALL-E 2 and Stable Diffusion. 

Data annotators are involved in every stage of AI development, from training models to verifying their outputs to offering feedback that makes it possible to fine-tune a model after it has been launched. They are often forced to work at an incredibly rapid pace to meet high targets and tight deadlines, says Srravya Chandhiramowuli, a PhD researcher studying labor practices in data work at City, University of London.

“This notion that you can build these large-scale systems without human intervention is an absolute fallacy,” says Chandhiramowuli.

Data annotators give AI models important context that they need to make decisions at scale and seem sophisticated.