Is everything an LLM tells you fact-checked?

No, not automatically. When an LLM answers you, it isn't looking anything up or checking a fact against a source in real time — it's generating text based on statistical patterns it learned during training. Think of it like a very well-read friend answering off the top of their head: usually right, because they've absorbed a huge amount of information, but perfectly capable of stating something wrong with total confidence.

Why do LLMs get facts wrong?

The reason is baked into how they work. An LLM produces the next most plausible chunk of text, not the most verified one. When its training data was fuzzy, outdated, or simply didn't cover a specific fact, it fills the gap with something that sounds right anyway.

This is what people mean by hallucination: an answer that reads as completely credible but is factually wrong, with no built-in mechanism to catch the error. There's no fact-checker quietly running in the background unless you specifically turn one on.

What tools add real fact-checking?

On its own, an LLM can't verify itself. But some tools bolt a verification layer on top of the model, and that changes the picture:

• Web search — the tool looks up live pages and grounds the answer in what it finds.
• Retrieval-augmented generation (RAG) — it pulls relevant documents from a trusted source and answers from those.
• Citations — it links the specific sources behind a claim so you can check them yourself.

These help a lot, but they're separate systems added around the model, not something the LLM does by default. And even grounded answers can misread their sources, so the link matters more than the confident tone.

What should you always double-check?

Treat unverified LLM output the way you'd treat that smart friend's off-the-cuff answer. It's great for brainstorming, drafting, and general explanations. But verify anything that actually matters — specific numbers, dates, and statistics; legal or medical claims; and especially citations, since LLMs have been known to invent sources that sound real but don't exist.

The rule of thumb: the higher the stakes, the less you should trust an unchecked answer.