As someone who's been gaslit by ChatGPT and Claude on more than one occasion on something being novel that simply didn't even do a google search to see that it was in fact previously patented... this is dangerous and seems to be happening far more often even as models are getting better at "grounding".