“A process called reinforcement learning from human feedback is used right now in every state-of-the-art model,” to fine-tune its responses, Baum says. Most AI companies aim to create systems that appear neutral. If the humans steering the AI see an uptick of right-wing content but judge it to be unsafe or wrong, they could undo any attempt to feed the machine a certain perspective.
OpenAI spokesperson Kayla Wood says that in pursuit of AI models that “deeply represent all cultures,…
Read the full article here