Hallucination is pattern completion when the truth is not in the context or strongly in the weights.

Ask a chatbot for three research papers on some niche topic and you may get three beautifully formatted citations, with real-sounding authors and plausible journal names, for papers that never existed. It’s tempting to call that lying. It isn’t. It’s the model doing its normal job a little too well.

The idea

A model is trained to continue text that looks right (see the next-token lesson). When you use it, nobody is retraining it. It’s reading from frozen weights, the numbers it learned in training, and picking the next token, over and over.

At no point does it get a side channel that says “this fact is true.” It only gets: given this prompt, what token is statistically likely next?

So when you ask for a paper, a quote, or last quarter’s numbers, and the real answer is

not on the desk (in the context window), and
not burned into the weights with high confidence,

it still has to pick some next word. Fluency is cheap. Honesty is not a native output. The result is a citation that sounds like a real conference paper, by a real-looking author, that never existed.

Three flavors

Experts usually split hallucinations into three kinds:

Confabulation: filling a gap with a plausible story. This is the classic made-up citation.
Stale weights: it “knows” a 2023 fact and you asked about 2026.
Context conflict: your prompt said two different things and it blended them.

Each one has a different fix, which is why the distinction is useful.

An analogy: the confident tour guide

Imagine a tour guide who has given ten thousand tours and is never, ever allowed to say “I don’t know.” A tourist points at an unremarkable building and asks about its history. The guide has no idea. But they know what building histories sound like: built in the 1890s, a merchant family, a fire, a restoration. Out it comes, smooth and confident.

The guide isn’t trying to deceive anyone. They’re completing the pattern “tour guide answers question about building.” The fix isn’t to ask the guide to be more honest. It’s to hand them a plaque with the real history, or let them check a guidebook first.

For models, that plaque is grounding: putting the real source on the desk with retrieval (RAG), letting the model use search or a database, or requiring it to show where each claim came from.

Try this

Ask a chatbot: “Give me three academic papers about [a very narrow topic you know well], with authors and years.”
Search for each one.
Now ask again, but paste in a real article on the topic and say: “Using only the text above, list the sources it cites.”

Compare how many of the first list exist versus the second. You’ve just watched grounding work. If your chatbot has a web search mode, try the first question with search on and off.

Use it at dinner

“It’s not lying. Lying needs a goal. It’s completing a pattern. If truth isn’t in the context or strongly in the weights, the most likely continuation is still a fluent sentence, not a shrug. That’s why teams ground answers with retrieval, tools, or ‘show your source,’ instead of hoping the model grew a conscience.”

A quick extra: turning the “temperature” (randomness) down only half-fixes this. A less random model still picks a confident wrong answer if the right one isn’t available. Grounding is what changes the odds.

Sources

Philip Kiely, Inference Engineering (Baseten Books, 2026), on training versus inference. (Book, no link.)