Yes, actual typos. LLMs don’t perform spell checks. Yes, they can “spell check” your inputs, but that isn’t what is actually happening. It’s all predictive text, and if they’ve learned to predict that the appropriate word in this context in 98% of cases is Minnesota, but since their dataset includes real human errors, it’s not unrealistic that then the LLM could also conclude that the appropriate word in 2% of cases is “Minneosta” instead, which means given enough output variables the misspelling will appear.

Again, I have personally had LLMs generate output with typos, not just factual errors.

Edit: cleaned up a couple very human typos