Researchers, educators and companies are experimenting with ways to turn flawed but famous large language models into trustworthy, accurate ‘thought partners’ for learning.