I've tried ChatGPT on cases taken from the European Qualifying Examination for patent attorneys to see if ChatGPT can correctly identify how prior art document combine to attack a granted patent. (I used Paper C for this).<p>While it spouts lots of vaguely correct patent law back at you, at no point is there even the beginning of a valid attack. The program would definitely not get a passing grade.<p>To pass the exam, one would have to realize that elements in the prior art anticipate claim features, even if they are named differently. A lot of prior needs to be ignored to find the anticipating parts. I did not even use the full exam, just snippets, but this clearly is too difficult.