I don't think this is about expecting AI to successfully fact-check observations, let alone do its own research.

I think it is more about using AI to analyze research papers 'as written', focusing on the methodology of experiments, the soundness of any math or source code used for data analysis, cited sources, and the validity of the argumentation supporting the final conclusions.

I think that responsible use of AI in this way could be very valuable during research as well as peer review.