LLMs generate one token at a time, so anytime they happen to generate some garbage for any reason, you get "garbage in" for further tokens for as long as it stays in the context window.
Not totally sure, but based on other folks' comments a number of external sources of noise are possible, e.g. from other conversations or from hitting a weird spot in the training data.
Don’t present this as some kind of anomaly unique to AI, the concept of “garbage in garbage out” is all that applies here.