“But then Meta is using those academic non-commercial datasets to train a model, presumably for future commercial use in their products.”
All 10.7 million clips in one of Make-A-Video’s training sets were scraped from Shutterstock, watermarks included. Another 10 million came from more than 3 million YouTube videos, collected by Microsoft Research Asia. The academics who built the datasets cited a non-commercial research exception. Meta then picked up the result. The university takes the legal risk and the company takes the product.