I don't doubt it. I've only been using it for workflow aids (copilot style stuff, and using it to generate unit tests to capture error handling conditions etc), and now we are piloting first generative text products but very human in the loop (customer data used to feed into a prompt but the output then feeds into an editor for a human being to proof and update before doing something with it). The amount of totally fake webinars hosted by totally fake people it has hallucinated is wild (the content and agendas and such sound great and are sensible but none of it exists!)

The consistent with the world type stuff could be built into the prompt engineering (e.g., tell the user about a map you have) and I think you could largely minimize hallucination but still have very realistic conversations

I can definitely imagine Newton posting this, "that fucker Leibiniz gets his name on the notation it's not fair!"