There is a basic principle here. It is very difficult for AI to learn from human beings the things that human beings do not know that they know.
Read that sentence again.
No textbook, manual, algorithm, data, set, etc., successfully describes everything that a true domain expert knows. Domain experts often make decisions, create connections, and solve problems with insight that they would struggle to fully verbalize.
Text books and learning systems do not encode the knowledge itself, except on a superficial level. What they encode is a pathway which, when followed by a human being, leads to that human being acquiring knowledge and insight into the subject matter that is not reducible to the specifics of the text or learning system.
Anyone with deep domain expertise in an area, who interacts with an AI that has been “trained“ in that area, is likely to discover that the AI is a classic example of “book learning“ without “street smarts“ or real insight.
As more and more AI compute is dedicated to long-term direct observation of exactly how experts in an area operate, AI can probably learn the things that those human experts know, but do not know that they know in explicit terms. These are things that are sometimes called “intuition“ or “insight“.
Just as AI robotics can learn by observing how people move and how they physically do things, AI will have to shadow many experts in a discipline over a considerable length of time in order to learn what humans know that is not reducible to a text.
It can gradually happen. But an AI who has read and absorbed every text that exists on a topic is not the equal of a human domain expert yet.