Foundations of AI Systems, Their Use Cases and Impacts
Get the taxonomy right
The single most common source of missed points is confusing the nested categories. Artificial intelligence is the broad field of systems performing tasks associated with human cognition. Machine learning (ML) is a subset in which behaviour is learned from data rather than explicitly coded. Deep learning is a subset of ML using multi-layer neural networks. Generative AI is a subset of deep learning that produces new content (text, images, audio, code), typically via foundation models trained with self-supervision and adapted by fine-tuning, prompting or retrieval augmentation.
Learning paradigms and task types
Supervised learning uses labelled examples (classification, regression); unsupervised learning finds structure without labels (clustering, dimensionality reduction); self-supervised learning derives labels from the data itself (the basis of large language models); reinforcement learning optimises actions against a reward signal, including RLHF for alignment. Know the difference between training (fitting parameters) and inference (serving predictions), and between deterministic rule-based systems and probabilistic models.
Use cases and impacts
Candidates should map capabilities to sectors: fraud detection, credit scoring, medical imaging triage, predictive maintenance, recommendation, conversational agents, document review. Impacts are two-sided: efficiency, personalisation and discovery versus bias and allocative harm, opacity, privacy erosion, safety failures, automation-driven labour displacement, energy and water consumption, and error amplification at scale. Evaluation vocabulary matters: confusion matrix, precision, recall, F1, false positive versus false negative costs, overfitting, data drift, and human-in-the-loop oversight.
Worked example
A bank builds a model predicting loan default from ten years of labelled repayment history. That is supervised classification, not generative AI and not reinforcement learning. Because the output is a probability, the threshold is a policy choice: lowering it raises recall (fewer missed defaults) but lowers precision (more good applicants rejected) - so the governance question is who owns that trade-off, not merely how accurate the model is.
Trap: options that treat AI, ML, deep learning and generative AI as synonyms, or that assume more data always improves outcomes.