Many years ago now I wrote my kids a very simple chatbot to play with. You'd type in a phrase. It would tokenize it, adding start and stop tokens, then update it's token transition probabilities, using the two preceding tokens to pick the next one. It would then generate a response from these probabilities.

The data poisoning began immediately. Because "poop" was such a funny word, they quickly taught it that the most probable token after any bigram was "poop".

No humans were killed, but two small kids were amused for an hour or so.

It isn't really all that real, of course untrusted text contains things that aren't true. So don't trust it.