I do want to bother.

Say that I want my LLM to output professional fiction writing. For such content, humans would have editors to review the book for any grammatical and lexicographical errors. In addition to that, the first edition could potentially have millions of eyeballs checking for errors. Some of the readers would report their findings back to the publisher, who might have the editor correct the mistakes, then push the diff to the author, if things are properly arranged, who will reject and/or approve any changes for the second edition if there is enough popularity to warrant there be any.

The model is more likely to produce quality writing if the input resembles the desired output. Same logic. It's a machine designed to predict "What's the next most likely word?". If past words are full of spelling mistakes, misspelled words, which are also tokens in its vocabulary, are more likely to appear. In fact, big models tend to be able to 'understand' the mis-spellings.

I'm much more annoyed by models that do the opposite: where providers inject extra system prompt to create bad spelling even though my input has been wrung through my browser's spellchecker and thus, at least spelling-wise, has few mistakes.

At least, this is the case in theory, with the laziest training method, so that would be the most obvious explanation, see occam's razor. However, if a model is trained with pre-processing filters that correct common spelling mistakes in the target output but not in the input, then a model learns to 'be liberal in what you accept, be strict in what you emit'. In my opinion that would be the best of both worlds. The model would not emit grammatical or spelling errors, but it would still understand and accept them. These filters can be simple replace filters, using efficient indexing, so even trillions of tokens can be checked for mistakes.

Also... please don't assign these models significant intelligence. I checked something while writing this post, and I got this google "AI overview" of the results:

AI Overview

"Spelling-wise"

should not be hyphenated, as it is a compound word formed from "spelling" and "wise" used as a suffix meaning "in terms of". The hyphen is generally not used in this context, similar to words like "timewise" and "workwise". 

Rule of thumb

Do not hyphenate when "wise" is used as a suffix to mean "in terms of" or "with respect to".

Examples: ~~spelling-wise~~, moneywise, colorwise, otherwise, clockwise.

Do hyphenate when "wise" is part of a compound adjective meaning "smart" or "savvy".

Examples: street-wise, penny-wise. 

It doesn't just contradict itself (location marked with strikethrough), it's also just plain wrong, even when you just put together the top google results.

Real answer for English: It's complicated. If you 'made up' a compound word, then use a hyphen. If the compound word already exists: do not use a hyphen. Exceptions abound as well; where made-up words retain their hyphen, otherwise a few famous words have lost them. (See what I did there?)

This is in contrast to, for example, German, where if you concatenate some words into a new compound noun, that's an acceptable use of language and never hyphenated, creating stuff like Donaudampfschiffkapitän.

Language is hard, AI developers are lazy and move fast and break stuff.