What AI is, and why it needs governing
The exam uses the OECD definition, which the EU AI Act also builds on: an AI system is a machine-based system that, for explicit or implicit objectives, infers from the input it receives how to generate outputs such as predictions, content, recommendations or decisions that can influence physical or virtual environments. Systems vary in autonomy and in how much they adapt after deployment.
The key word is infers. Traditional software follows rules a programmer wrote. Machine learning (ML) learns patterns from data, so its behavior depends on that data and its outputs are probabilistic.
Types you must recognize
| Type | What it does | Example |
|---|---|---|
| Rules-based / symbolic | Follows hand-written logic | Expert system for tax rules |
| Supervised learning | Learns from labeled examples | Fraud classifier trained on labeled transactions |
| Unsupervised learning | Finds structure in unlabeled data | Customer segmentation by clustering |
| Reinforcement learning | Learns by reward and penalty | Robot control, game playing |
| Deep learning | Many-layered neural networks | Image recognition, speech |
| Generative AI / foundation models | Produces new content; trained on broad data and adaptable | Large language models, image generators |
| Agentic AI | Plans and takes actions with tools toward a goal | An assistant that books travel |
Also know narrow AI (one task, which is everything in production today) versus general AI (hypothetical human-level breadth).
Why AI is "socio-technical"
AI risk does not live in the model alone. It comes from the model plus its data, the people who build and use it, and the context it is deployed in. The same face-matching model is low risk unlocking a phone and high risk identifying protest attendees. Expect exam answers that fix a problem through people and process, not only through code.
What makes AI different to govern
- Opacity: complex models are hard to explain.
- Scale and speed: one flawed model can affect millions of decisions.
- Data dependence: bias and errors in data become bias and errors in outputs.
- Probabilistic outputs: the system will be wrong some of the time, by design.
- Change over time: performance drifts as the world changes.
- Autonomy and emergent behavior: especially in generative and agentic systems.
Check your understanding
Which word in the OECD definition best separates AI from conventional software?
A bank's model learns to flag fraud from millions of transactions already labeled 'fraud' or 'legitimate'. This is: