Gaming

These generative models are only capable of regurgitation. They're fed content, and they come up with statistical distributions of words and phraes in their training data to sample. But they're still just chopping up other people's copyrighted work and gluing it together.

This gets into generic "you can't copyright individual words" territory when dealing with very common topics, but when you get into niche topics, the training data gets sparse, and the distributions stop allowing for creating something new.

And they can't extrapolate beyond their data. All models break down once you leave the boundaries of the data.

So, any articles generated by these things about something new are either going to be trained on too small a data set to avoid lawsuits, or just be incredibly factually wrong.

And if they're publishing reviews of new games that are full of factual errors, they could get sued for defamation.

This is just stupidly risky on their part.