With the creation of an AI model like ChatGPT that is trained on vast amounts of data from the internet, there is always a risk that the AI will produce text containing explicit content. In order to prevent this, the AI model needs to be provided with examples of explicit content so that it can learn to filter out similar text in its outputs. This is where HITL training comes in. Developers look to data labeling firms to provide their AI model with extensive labeled datasets of explicit content that they can use for ML training. The issues arise when considering exactly what it means to label explicit content as a full-time worker.
The main concerns with creating ethical work conditions for this type of data labeling are making sure the workers are sufficiently compensated and provided with counseling services to mitigate the mental toll it takes to sift through explicit content on a regular basis. These concerns have been highlighted in the case of Sama, a self-acclaimed "ethical AI" data annotation firm that has come under fire for its unethical treatment of its data labeling workers. Sama provides outsourced data labeling and content moderation services to major companies like Google, Microsoft, Walmart, Facebook, and OpenAI. Sama employees providing content moderation services for Facebook out of Kenya have attested to being traumatized by the nature of their work while lacking the compensation to avoid mental healthcare services. In this case, investigators found the employees were making as little as $1.50 per hour, despite the difficulty of their work (1). Similar stories have come from the Kenyan Sama employees who provided data labeling services for openAI's ChatGPT model. These workers describe their daily task of labeling explicit text, sometimes detailing graphic situations like "child sexual abuse, bestiality, murder, suicide, torture, self harm, and incest", throughout a nine-hour shift, in order to bring home up to $2 per hour (2).
One potential measure to check for and prevent these unethical conditions, proposed by Caroline Sinders, is a wage calculator designed to determine if a data labeling gig worker is being fairly compensated for their work. While this tool doesn't account for the difficulty of specific data annotation jobs based on the content it covers (explicit v.s. non-explicit), it acts as a good jumping-off point for preventing wages that exploit these workers (3). It has even been stated by Sama employees "that they might be able to handle the trauma of the job – even take pride that they were sacrificing their own mental health to keep other people safe on social media – if only Sama and Facebook would treat them with respect, and pay them a salary that factors in their lasting trauma." (1). Paying fair wages is the first step to correcting and maintaining the ethics of this necessary occupation, with many more changes to come to ensure the security and mental health of these workers.
How can we prevent the exploitation of labor for content moderation?