Frozen in time, and prone to guessing
An LLM like GPT-4o or Claude learns everything it knows during training, on a dataset that has a cutoff date. Ask it about an event from last week and it either says it doesn't know or, worse, invents a plausible-sounding answer. That invention is called hallucination, and it happens because the model's job is to predict the next likely word, not to check whether a fact is true. Nothing in the architecture distinguishes 'I remember this' from 'this pattern seems statistically likely to come next.'
This matters enormously the moment you put an LLM in front of real users. A support bot that confidently misquotes your refund policy, or a legal assistant that cites a court case that doesn't exist, isn't a minor bug, it's a trust-destroying failure. Fine-tuning the model on new data helps a little but is slow, expensive, and still doesn't stop hallucination on questions the model is unsure about.