# Understanding a language model without being an IT specialist URL : https://skillrung.com/en/courses/artificial-intelligence/what-a-language-model-does Topic: Artificial intelligence · Duration: 20-30 min · Updated on 2026-09-15 ## In brief A language model consults no source of truth. It writes one word after another, each time keeping the most likely continuation of what came before. Two consequences hold together: it is right most of the time, and its mistakes come out with the same confidence. Three families of AI go by the same name; the one that generates is the only one dealt with here. ## Programme - 1. Knowing what you are dealing with before you use it : Identify which family of AI is meant here and what it produces - 2. What a model really does : Explain in three sentences what a language model produces and by what mechanism - 3. Your action plan : Turn this mechanism into two habits you can use on your very next request 1 content sections, each with its quiz. Three level quizzes: Beginner, Intermediate, Expert. ## Free preview (first 4 slides) ### 1. The same confidence, whether it is right or wrong A well-turned answer is not, for all that, a correct answer. The text comes out with the same assurance in both cases, and nothing in its form tells you which of the two you have in front of you. As long as you do not know how that text was produced, you cannot decide what it is worth in your structure. This course gives you that mechanism, in plain terms, and what it changes for you. ### 2. By the end of this course, you will be able to - Say in plain terms what a language model produces - Tell apart the three families that go by the same name - Name the four words you will meet everywhere afterwards ### 3. The four points we are going to look at Four points, in the order this course takes them. - The family that generates, and the other two - How text is cut into pieces, and what gets counted - The element of chance that makes two answers differ - The four words: token, context, prompt, hallucination ### 4. Where do you stand? One question before we start. It does not measure knowledge: it places what you already expect from an assistant. _Self-assessment of 1 questions in the player._ ## Access The first four slides of every course and the Beginner-level quiz are free, with no account. From the fifth slide onwards a subscription is required: €29.99/month incl. VAT, no commitment, cancellable online. skillrung is not a French training body certified under the Qualiopi scheme; its content is not eligible for the French CPF, OPCO or DPC schemes. ## Frequently asked questions ### Does a language model check its answers anywhere? No, no source of truth is consulted. Each word is chosen as the most likely continuation of what has already been written. From that come the two sides of the result: an answer that is right most of the time, and a possible mistake, stated with the same confidence. The assurance of the text therefore never tells you anything about whether it is correct. ### Why do two people get different answers to the same question? Because at every word several continuations are possible, and that choice involves an element of chance. Two answers that differ are not a defect: they are the mechanism itself. Keep what comes up both times; what came out only once is a lead for you to check yourself, not a conclusion. ### What is a token, and how does it concern me? Text does not reach the model as words: it is cut into pieces called tokens. A common word takes one, a long word several, a page a few hundred. It is the unit that gets billed when AI is plugged into your tools; with a per-person subscription, you never see it go by. A document you upload is therefore cut up, then counted. ## Sources - Règlement (UE) 2024/1689 établissant des règles harmonisées concernant l'intelligence artificielle : EUR-Lex : Journal officiel de l'Union européenne, 2024-07-12, https://eur-lex.europa.eu/eli/reg/2024/1689/oj?locale=fr - Recommandations de sécurité pour un système d'IA générative : ANSSI, 2024-04-29, https://messervices.cyber.gouv.fr/guides/recommandations-de-securite-pour-un-systeme-dia-generative - Les fiches pratiques IA : CNIL, 2026-09-11, https://www.cnil.fr/fr/les-fiches-pratiques-ia - Mata v. Avianca, Inc., opinion and order on sanctions, No. 22-cv-1461 (S.D.N.Y.) : Justia : US District Court, Southern District of New York, 2023-06-22, https://law.justia.com/cases/federal/district-courts/new-york/nysdce/1:2022cv01461/575368/54/