GlossaryFloor 1 · The Modela solid block on its own: the prediction machineFloor 1 · The Model
generative AI
No. 017 · v2026-08FR: IA générativeGenerative AI refers to the systems that produce new content, text, image, sound or code, rather than classifying or measuring what already exists. Like a draughtsman composing a scene never seen before, where a stamp always reproduces the same imprint.
What it is not
Generative AI is not artificial intelligence: it is one family of it, which appeared late and became visible very quickly. Fraud detection, recommendation, image recognition or demand forecasting are matters of machine learning without generating anything, and have been running in companies for decades. Nor is it a synonym for language model: producing an image, a voice or a molecular structure belongs to the same family without going through text. Using the word for everything that carries the AI label makes what already existed look new, and what is not universal look universal.
In depth
A regime, not a technique
What brings these systems together is not a single technique but a mode of operation: they learn the regularity of a set of examples, then draw from it new objects that resemble the set without figuring in it. Depending on the nature of what is produced, the architectures differ markedly, and a model that writes text has little in common, in its inner workings, with a model that synthesises an image. The common point lies in the gesture: manufacturing a plausible object rather than recognising an existing one. It is also what explains why the question of truthfulness arises everywhere in this family, whereas it arose differently for the models that merely classified.
What the confusion costs
The confusion of vocabulary has practical consequences inside an organisation. A forecasting or anomaly detection project has no need of a generative model: it calls for clean data, a measurable target and a specialised model, often smaller, faster and far easier to evaluate. Placing a generative model on a classification problem amounts to paying very dearly for a flexibility you have no use for, and to inheriting an uncertainty about the output that you did very well without. The useful reflex is to name the task before choosing the family, never the other way round.
Three questions
Three questions systematically accompany this family, and none of them is technical. Where the training data comes from and with what right of use: the point is not settled everywhere, and it commits the organisation that publishes the result as much as the supplier of the model. What a production nobody has checked is worth: the cost shifted onto the proofreading is real, and rarely counted in the gains announced. Finally, European regulation requires content produced or modified by an AI system to be signalled, which makes traceability a constraint of design and not a politeness added after the fact.
Relations where the neighbours live
- Often confused with
- Floor 1 · The Modela solid block on its own: the prediction machinelarge language model
- Related comparisons
- AI or Generative AI
Check 3 questions · click your answer
Level 1 · Recognise
A system spots suspicious transactions in a banking flow. Is this generative AI?
Level 2 · Distinguish
Which statement is accurate, about generative AI and language models?
Level 2 · Distinguish
You have to forecast next month’s volume of orders. What do you choose?
Lexigraph, "Generative AI", v2026-08, https://www.lexigraph.org/en/generative-ai/, CC BY 4.0.