artificial intelligence

In depth

The nested circles

The field is a set of nested circles, and that is what explains the confusion. Artificial intelligence is the widest: it includes approaches that learn nothing, such as systems built on hand-written rules, which still govern a great deal of software reputed to be intelligent. Inside it lies machine learning, where behaviour is no longer written but derived from examples. Inside that again, deep learning, which uses neural networks with many layers. And only inside that, generative AI, of which language models are one family. Saying “AI” while thinking of a language model is like saying “medicine” while thinking of an antibiotic.

The price of imprecision

This imprecision has concrete consequences. A regulation, a budget or a training plan that speaks of “AI” without saying which circle it means describes badly what it governs: the risks of a system that scores applications have nothing in common with those of a writing assistant, nor do the skills they call for. This is precisely why serious regulations classify uses rather than technologies, and why one and the same method can be innocuous there or strictly governed depending on what it is made to decide.

The AI effect

Finally, the word carries an old problem, known as the AI effect: as soon as a task is solved reliably, it stops being counted as artificial intelligence and becomes ordinary software again. Character recognition, machine translation and route planning have all followed that path. One practical conclusion follows: the word names less a stable category than a moving frontier, that of what still surprises us. To speak precisely about a system, it always pays to name what it does rather than the family it belongs to.