A chatbot like the ones everyone uses is doing one thing repeatedly: predicting what word comes next. Everything else follows from that.
The prediction machine
The model was trained on enormous amounts of text. During training it repeatedly saw a passage with the next word hidden and adjusted itself to guess better. Do that billions of times and it becomes very good at continuing text plausibly.
When you ask a question, the model treats your question as the start of a document and writes what would plausibly come next. That happens to look like an answer because in the training text, questions are followed by answers.
Why it seems to understand
Because predicting text well requires representing a great deal about how things relate. To continue a sentence about physics correctly, the model has to encode something about physics. Whether that counts as understanding is a philosophical argument; functionally it behaves like knowledge.
Why it makes things up
Because plausible and true are different. The model produces the most likely continuation, and a confident wrong answer is often more likely than an admission of ignorance, since training text rarely contains people saying they do not know.
This is not a bug that will be patched. It is a consequence of how the system works, mitigated rather than removed.
What the training cutoff means
The model knows what was in its training data and nothing after. Ask about recent events and it will answer from a world that ended at some point in the past, unless it has a search tool.
The practical takeaway
Excellent at language tasks. Unreliable about facts. Verify anything that matters.