How do you know a human tutor is not confidently incorrect?

I know it sounds like a glib answer, but in reality the answer to both questions is you don't know for certain in either case. Honestly I would say both are incorrect for at least some portion of their responses, but you seem to be insinuating that humans are closer to 0% while the LLM's percent is higher. In my estimation the LLM is probably less often incorrect but that's just my opinion. I am sure a more rigorous study could be conducted though.

It's true that you can't know with a human tutor either. But ChatGPT is known to be confidently incorrect. If a human tutor had earned a similar reputation, one would be wise to distrust him as well.

This is also my own experience. In particular, ChatGPT has deeper knowledge on complex grammatical structures (like the 7 cases polish have) than I can expect from a regular tutor – my friend was not able to explain complex consonant clusters, while ChatGPT happily babbles out resources and explains just what I need for a particular case.

I know that as well as with any human tutor. The main difference is that I am motivated to check the answer, eg. by asking the woman who is my main reason to learn Polish.

Also, the value of a tutor that in and encouraging way corrects my verb conjugation of być (to be) for the 500th time has a lot more value than the risk of it being wrong on a corner case.

A human tutor would semi-arrogantly have told be off and asked me to look at the dictionary (Yes, I have studied English, German, Spanish, and French in school, I have had more language teacher than anything else)

Curious, when you went back to that woman and checked an answer with her, did she ever say "No that's definitely the wrong way to say it" or anything like that?

I'm not arguing in any way I'm legitimately curious, as someone who has dabbled in language learning, as well as trying to use ChatGPT as a tutor, how its track record has been. Hallucinations was definitely one of my concerns when asking it about language stuff.

She corrected usage. Eg. when to use mama and matka where it is super hard to intellectually learn how formal one should be.

But again, in english, when is it appropriate to use mom and mother? It really depends on the context, culture, and relations.

Also hallucinations are probably a good thing for language learning. especially when you have an interface, chat, where you can interrogate – and an AI that does not get offended when you question it.

I think hallucinations would be less worrisome, since words and their meanings are collectively probably the strongest suit for ChatGPT.

You never actually know that, and most people are. That's why there's so much "nous" use in French classes and so little "on" and so on. But the learn rate is so fast, that you get sufficient coverage to fix it up in actual conversation. My wife and I did a combo of another AI product (https://www.speak.com/) and ChatGPT to learn sufficient Spanish to get by in Chilean Patagonia and it did a pretty decent job (we just learned pronunciation through mimicry).

We're not experts as a result, obviously, but it helped us get around and stuff which is fun. It was pretty much a crash course started as we were landing in Santiago (which is where we entered the country).

> But the learn rate is so fast, that you get sufficient coverage to fix it up in actual conversation.

This should really be emphasised. Correctness might not be entirely on every time. But learning pace and engagement is so much higher, that it more than makes up for it.

Especially engagement is a point where ChatGPT wins over all human teachers.

Any tutor makes mistakes, it is left for the learner to figure out whether it is BS later.

Truth to be told, human WILL never match ChatGPT's patience and coverage when it comes to explain most insignificant details of one particular language, any time anywhere you wanted. It takes out the fear/embarrassment factor when it comes to a human tutor.