Hallucinations
You know what a hallucination is, where it appears most and how to neutralise it.
ObjectiveBy the end of this lesson, you will be able to recognise situations prone to hallucinations and neutralise them with cited research and a neutral question.
A hallucination is a false statement presented with confidence by a model: a reference that does not exist, an invented figure, a quote attributed to the wrong person, a legal article that says something else. It is not a lie. It is the direct consequence of the prediction mechanism: when the model does not know, it still produces the most plausible continuation.
Where they appear most
- Precise references: article titles, page numbers, web addresses, legal articles.
- Rare details: the founding date of a small company, the name of a local manager.
- Niche topics, little present in the training texts.
- Questions that presuppose an answer: “Which study proved that…” pushes the model to find one.
Hallucinations are rare on well-documented topics and on rephrasing tasks. They become frequent once the model is asked for a precise fact it does not have in front of it.
Give me five studies that prove microlearning is more effective.
Look for academic studies that compare microlearning with longer formats. Cite each study with a link. If the results are mixed, say so.
Cited research replaces the model’s memory with verifiable sources, and the question no longer imposes the conclusion.
How Learnya reduces them
Whenever factual information is at stake, Learnya relies on sources rather than on the model’s memory. Searches on the web and in the academic literature are cited, and a verification agent can check that each source really says what is attributed to it. For your documents, the agent reads the file instead of remembering it.
These mechanisms greatly reduce the risk, without eliminating it. A source can be summarised badly, or a page can itself be wrong. That is why opening two or three sources remains the best habit for anything that will be published or passed on.