> "pretrained models with guaranteed no IP issues"
The problem is that even if Shuttertock train model on the images they routinely sell it doesn't mean it's going to be free from IP issue. The subtleties are going to vary widely between jurisdiction, but it's not like Shutterstock was the author of most of the images in its bank: most if not all the pictures there have been taken by third party photographers, who (at least in some jurisdictions) keep the entirety of the copyright over their creation. So if court ends up ruling that images generated by ML models trained on copyrighted material is plagiarism, then Shutterstock will be plagiarizing their own suppliers.
Then maybe now their contract terms include a license to do so, but I'm pretty sure their old contracts did not, and even now, the said contract has to be ruled valid by a court.
In the US, where the big international players are, the legislator will probably shape the law in their favor (and against the photographers), but in many countries, where photographers are the actual business and where Shutterstock and others are just foreign service provider, the legislator can be tempted to protect the local businesses against Sutterstock.
It's going to be a mess in the next few decades (but IP has always been a mess, that's what happen when you artificially construct an exclusive “property right” out of something immaterial: intellectual property is trying to get a square peg in a round hole)