“Unreliability, including hallucinations and unexpected mistakes, is currently the single biggest obstacle for OpenAI's neural networks to be truly useful.”