Machines that infer

Lesson 1 of 5 in What Is AI? From Rules to Reasoning Machines.

Start with a question that sounds trivial and is worth billions of euros: what counts as AI?

If a bank denies your loan using a hand-written rule — reject if income below €30,000 — no AI law in the world applies. If it denies you using a model that learned its own rejection logic from ten thousand past loans, several do. The entire architecture of AI governance rests on being able to tell these two apart.

Every clause is load-bearing, but one word carries the weight: infers. A conventional program follows logic a human wrote down. An AI system derives its own way of producing outputs from the inputs and data it was given. That single distinction — followed rules versus inferred behaviour — is what the EU AI Act, the 2023 revised OECD definition, and most modern statutes use to draw the legal boundary.

Three consequences follow, and they shape everything you will learn on this site:

1. Nobody wrote the behaviour down. The decision logic emerged from data, so there is no rulebook to open when something goes wrong — governance has to demand its own artifacts (documentation, logs, testing) instead.
2. The behaviour is statistical. The system is right usually, not always, and it fails in patterned ways — which is why bias and robustness sit at the core of every framework.
3. The behaviour can drift. A system that adapts after deployment keeps changing after the lawyers signed off. Governance therefore never ends at launch.

Key terms: AI system, inference, machine learning, autonomy spectrum

Interactive checkpoint quiz (2 questions) — open this page in a browser to take it.