Because AI operates at scale, AI bias is not just a replication of human bias. It is a worsening of it. We must require AIs to pass bias tests for before they are released. We can’t continue to leave bias detection up to investigative journalists to uncover after-the-fact.When addressing bias in AI, we must also examine its means of production, leading to my second major concern: labor.
Even after the LLM is trained, it has another layer of human evaluation and writing, which is often carried out by workers in the global South who are subject toI always tell my students that ChatGPT feels like it’s a way to write without doing all that work only because it’s not our work that goes into it. It’s somebody else’s.Finally, and most importantly, we need to focus on the business model behind AI development.
Social media recommendation AI, the kind that chooses what videos to recommend in the sidebar of YouTube, organizes an Instagram feed and queues up TikToks, was humanity’s first large-scale encounter with AI. Those AI figured out that some of the best ways to keep us engaged were to recommend incrementally more extreme content, to feed us conspiracy theories and to keep us outraged, ashamed, fearful and divided.