It'd be easy enough to train a Lora for an LLM to introduce errors. But i can still tell when talking to an AI for more than 20 seconds