A large language model answers by predicting, one piece at a time, the most likely continuation of the text it has been given. It does not look facts up in a database; it produces the most plausible-sounding answer from patterns learned in training plus whatever is in its context. When the patterns are strong and the context is good, the answer is accurate. When neither is, the model can produce a fluent, confident answer that is simply wrong. That is a hallucination.
Prediction, not lookup
A language model is trained to predict the next piece of text. Given "The capital of France is", it has learned that "Paris" is overwhelmingly likely. Scale that up across trillions of words and the model learns grammar, facts, styles, arguments and the shape of a good answer. When you ask a question, it is not retrieving a stored answer; it is generating the most probable response to your words.
This is why models are so fluent and also why they can be wrong without any sign of it. Fluency comes from the prediction being good at producing natural text. Truth is a separate matter.
What a hallucination is
A hallucination is an answer that is plausible, confident and false: an invented citation, a quote the author never wrote, a date that is slightly off, a product feature that does not exist. It happens most when the model is asked about something it has little reliable information on. Faced with a gap, a model fills it with what sounds right rather than saying "I do not know", because "sounds right" is exactly what it was trained to produce.
Why context changes the picture
The single most effective way to reduce hallucination is to put the right material in front of the model when it answers. If the actual chapter is in the context, the model does not have to guess what the author said; it can read it. This is the principle behind a digital twin.
When someone asks an AccessMe twin a question, the relevant passages of the creator's own material are placed in the model's context, and the twin is asked to answer from that material in the creator's voice. The twin is still a language model, so it is not infallible, but it is anchored: the answer is assembled from what the creator actually wrote or said rather than from a statistical impression of them.
What that means in practice
- Where your material is rich, the twin is reliable. Questions your books, talks or notes clearly address get grounded answers.
- Where your material is thin, the twin is weaker. A question you have never addressed is where a twin is most likely to generalize. The fix is to add material on that topic, or to accept that the twin will say it does not have your view on it.
- The twin is not a fact-checker for the world. It represents your point of view. It is not meant to be a substitute for professional advice, and every twin page says so.
How to work with a guesser
Treat a language model the way you would treat a brilliant, well-read assistant who never admits uncertainty: give it the source, ask specific questions, and check anything that matters. For a twin, the creator can read the conversations their audience chooses to share and add material where an answer was vague. Over time the twin gets more grounded because the material behind it gets more complete.
Common questions
Does a digital twin hallucinate?
It can, because it is built on a language model. Anchoring it to the creator's own material reduces hallucination substantially on topics that material covers, and the creator can keep improving it by adding material where answers were thin.
Why does an AI sound so confident when it is wrong?
Because fluency and truth are produced by different things. The model is trained to produce natural, convincing text; whether that text is accurate depends on what it learned and what it was shown. Confidence in the writing is not evidence of correctness.
Can a twin say 'I don't know'?
A well-built twin, asked about something its creator has not addressed, should say that it does not have the creator's view on it rather than invent one. Adding material on the topic is the way to turn that into a grounded answer.