1.2 What AI is: ML, deep learning & generative AI

Why definitions matter for governance

For a governance professional the point is not to build AI but to recognize it, because laws and standards attach obligations to specific defined terms. The EU AI Act, echoing the OECD, defines an AI system as 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. Notice the definition centers on inference — the system derives outputs rather than following only hard-coded rules — and on autonomy and adaptiveness. Someone who can tell whether a tool truly "infers" can tell whether a regulation even applies.

From artificial intelligence to machine learning

Artificial intelligence (AI) is a broad field of computer science concerned with building systems that perform tasks normally associated with human intelligence — perceiving, reasoning, learning, and deciding. Machine learning (ML) is the dominant subfield today. Instead of a programmer writing explicit rules, an ML system learns patterns from data: you feed the algorithm many examples, it adjusts internal settings to fit them, and the resulting model can then generalize to new, unseen inputs. The three classic paradigms are:

• Supervised learning — the model learns from labeled data (inputs paired with correct answers), e.g., emails tagged "spam" or "not spam." It powers classification and regression behind credit scoring, medical triage, and hiring screens — precisely the high-stakes uses governance worries about.
• Unsupervised learning — the model finds structure in unlabeled data, e.g., clustering customers into segments or detecting anomalies. There is no answer key, so evaluating correctness (and bias) is harder.
• Reinforcement learning — an agent learns by trial and error, receiving rewards or penalties for actions in an environment, e.g., game-playing or robotics. Poorly designed rewards can produce unexpected, gameable behavior.

Deep learning and neural networks

Deep learning is a subset of ML that uses artificial neural networks with many layers ("deep" refers to the number of layers). Loosely inspired by biological neurons, a neural network passes inputs through interconnected nodes whose connection strengths — the parameters, or weights — are adjusted during training. Deep networks excel at perception tasks (image recognition, speech, natural language) because they learn hierarchical features automatically rather than relying on hand-engineered rules. The trade-off, central to governance, is opacity: a network with billions of parameters is effectively a "black box," which is exactly why explainability and transparency obligations exist.

Keep this core vocabulary straight:

Note that bias has two meanings the exam distinguishes: a statistical property of a model, and a societal harm (discrimination). Governance is chiefly concerned with the second, though it often originates in the first.

Generative AI, foundation models, and LLMs

Generative AI refers to models that create new content — text, images, audio, code, video — rather than only classifying or predicting. The breakthrough enabler is the foundation model: a very large model trained on broad, internet-scale data that can be adapted to many downstream tasks. Large language models (LLMs) such as the GPT and Claude families are foundation models specialized for text; they are trained to predict the next token (word fragment), and through that simple objective at massive scale they acquire striking language ability.

Two concepts recur on the exam:

• Training vs. inference. Training is the expensive, periodic process of learning parameters from enormous datasets — raising data-provenance, copyright, and environmental concerns. Inference is each subsequent use of the model to answer a prompt — raising privacy (what users submit), accuracy, and cost concerns. Governance controls differ at each phase.
• Hallucination. LLMs can produce fluent, confident output that is factually wrong or entirely fabricated, because they generate statistically plausible text, not verified truth. Hallucination is a core reliability risk addressed through human review, grounding/retrieval, and use-case restrictions.

Narrow vs. general AI

Almost every system in production today is narrow (or "weak") AI: competent at a specific task or bounded set of tasks. Even a powerful LLM, though broad, is still narrow in that it lacks genuine understanding, goals, or reliable performance far outside its training distribution. Artificial general intelligence (AGI) — a hypothetical system matching or exceeding human ability across essentially all cognitive tasks — does not exist and may or may not be achievable; superintelligence goes further still. The exam expects you to know the distinction and to treat AGI as forward-looking, not a present governance reality. This matters because policy and "existential risk" debates (an ongoing-issues topic) hinge on these categories, while day-to-day governance concerns narrow systems. Ultimately, knowing whether a tool is supervised or generative, narrow or general, and in training or inference tells you which risks, which stakeholders, and which legal definitions come into play.