Module 2 / 4 · 1 h
Understanding how it works
Prediction, training data, limits: understand in order to trust better… and to doubt better.
Log in to save your progress. Log in
1 / 4 · 12 min
Predicting the next word
A language model does not “know” in the human sense: at each step it computes which piece of text is the most likely given what precedes. These pieces are called tokens (a word, part of a word, a sign).
By chaining thousands of predictions, the model produces fluent and often relevant sentences. But fluent does not mean accurate.
2 / 4 · 10 min
Training data
The model learns from very large volumes of text. Three practical consequences follow:
- Its knowledge has a cut-off date: it may ignore recent news.
- It reflects the biases in its data (stereotypes, over-representation of some languages or cultures).
- It does not always cite its sources, and may invent some.
3 / 4 · 12 min
Limits: hallucinations and errors
A hallucination is a plausible but false answer, delivered with confidence. It is more frequent when the question is specialised, ambiguous or about precise facts (dates, figures, references, quotations).
Habit to adopt: for any important fact, cross-check with a reliable source before using it.
4 / 4
An example of hallucination
Example files
question.txt
Cite three scientific articles from 2021 on the impact of remote work, with full references.
reponse-plausible-mais-fausse.txt
1. Dupont, M. (2021). Remote work and productivity. Management Review, 45(2), 112-130.
(This reference is made up: it looks credible but does not exist.)
Never enter personal or confidential data into a public generative AI tool.
Quick test: the limits
Pass mark: 60%. Tick every right answer.
Log in to save your progress.
Module quiz
Pass mark: 70%. Tick every right answer.
Log in to save your progress.