As a counter-anecdote: I've had such consistently bad results from its AI summaries that I've changed over to the no-AI page entirely - well over half have been very obviously incorrect or answering something completely different but spelled similarly (but not the same, especially true for chemical names), and the links it includes as references often directly oppose its summary.

I suspect the model might just be too small? Which makes sense when exposing it for free, but it's still a bad experience.