I don’t know that GPT-3 is at the level of material damage but it’s clearly moving in that direction. Watching someone interact with it that has a good sense for composing the right kind of prompt is a spine tingling experience.
Given how poorly the PULSE demo that hallucinated fuzzy black folks into tan white people was received, it also makes PR sense not to release pretrained models that are obviously going to be easily tricked into saying terrible things. Exposing that as an API gives them the ability to police it a bit and study how the model behaves in the real world.
Or it could just be a giant sellout and cash grab.