Good point. But many are already directly training on output from GPT. Probably more efficient than copying the raw training data. Especially if it relies on this non-targeted approach.