It's marketing BS (edit, that may be too harsh)

They do reference "passGan" which I assume is this, from 2017: https://github.com/brannondorsey/PassGAN

Paper is here https://arxiv.org/abs/1709.00440

They use "AI" to generate passwords that are more likely based on actual choices, like P@s5w0rd kind of stuff. The paper that is, I have no idea about the company.

It isn't. You could perhaps make an argument that it's more likely to try "likely" passwords first, which means it will find theoretically strong but predictable passwords faster but should actually take longer to crack unpredictable passwords.

And of course it's trained on password leaks... but so are the people who write password cracking algorithms.