Instead of saying "I'm not racist but ...", people will say "As a large language model trained by OpenAI, I cannot generate racist statements, but ..."

Being based on training models, how can it ever come up with new language when its trying to imitate already existing language

I can see weird things starting to happen when AI-generated text becomes so prevalent that it starts feeding back into the language models themselves like some kind of ouroboros, then slowly starts drifting away from our current vernacular as errors accumulate and the bots get increasingly inbred.

That's a perfectly cromulent observation.

Would the opposite not be true? AI models work by predicting the next likely text. If we start changing language right from underneath it that actually makes it worse off at predicting as time moves along. If anything I would expect a language model to stagnate our language and attempt to freeze our usage of words to what it can "Understand". Of course this is subject to continued training and data harvesting, but just like older people eventually have a hard time of understanding the young it could be similar for AI models.

AI will start accidentally creating new slang which the young generation will start incorporating ironically until it sticks to the point us old heads will start saying it unironically (like how my 40 year old ass been saying 'yeet' lately) and then boom, it's part of the normal language.

As a large language model, I am quite interested to see how this shit goes down.

As long as they start using real words instead of u r finna bullshit, I'm fine with it