I've been working on that since last year, and notably on GPT-3 and other models on textbooks specifically with a publisher since September - here's how we convinced people: https://www.youtube.com/watch?v=7V3VkNj2bag

The problem is the tech can be wrong and that destroys retention. You need to a) verify statements b) consider citing statements c) Identify out-of-domain or "base GPT model response" answer.

All I can say is it has been incredibly hard so far, but we've made noticeable progress recently, and hit recall of 98% for psychology. Having been in OPs shoes, I wouldn't underestimate how temporary traction or a UI doesn't make up for bad retention for these products... it's the core engineering and NLP focus that matters (that's our bet anyway!)