> Yeah, telling grammatically correct but factually wrong things, having biases, crashing into things, and whatnot.

If you don't think it's right more often, please go poke an older model. Especially look at the ability to stay on topic.

> An assistant which can't be checked for accuracy, or for telling the truth.

What do you mean "can't be checked"? And any kind of assistant can make mistakes.

And calling it an assistant was your idea.

Any kind of assistant can make mistakes. But a human assistant can be made to show their work and explain their reasoning to check their output. If ChatGPT says "this thing is totally legal" or "don't worry about that rash", how am I to validate it's "reasoning"? How do I know where it's drawing it's inference from?