Exploitative labor to train GenAI models

Some GenAI tools have been trained using 'reinforcement learning through human feedback' (Fine-tuning). For this kind of training, human workers review a prompt and the generated output to give the model feedback about the accuracy and helpfulness of the output. Workers also have to check whether the output is appropriate.  

While this approach is not new - social media platforms are also known to employ humans to moderate their content - ChatGPT’s company OpenAI was criticised for outsourcing this practice to low-wage employees in Kenya. These employees must review toxic and explicit content to make the tools safer to use, but this often comes at the cost of their own mental well-being. More generally, companies are known to outsource AI-related work to employees in low-income countries in order to reduce costs. For example, a 2023 academic study shows how French start-ups have outsourced AI-related tasks to low-paid workers in Madagascar.

 

‘Exploitative labour to train generative AI models’, Microsoft Designer.