[pfisherman]

My experience building AI powered tools for scientists has been the opposite.

Other commenters have pointed out the difficulties users with limited domaine expertise having in parsing the output and distinguishing what is useful, novel, interesting from what is trivial or incorrect.

This slows down the iterative process of learning how to use the tool as well as refining queries because they don’t have a big enough database of ground truth knowledge to calibrate against.

That being said these can work well as educational tools to rapidly get non-experts to a point where they can participate in conversations with experts. In that regard one might say that the bulk of the value is in the non-expert users because they are far more prevalent than experts.

The biggest advantage I saw for the expert researchers was that they knew (1) which big questions to ask, and (2) how to break down those big questions into smaller, more precise questions.