Artificial Intelligence: Concepts, Techniques, and Limits
Artificial intelligence covers systems that perform tasks associated with human intelligence, such as recognising patterns, understanding language, and making decisions. The exam tests definitions, the relationship between AI, machine learning, and deep learning, and the practical limits of these systems. Questions often present an application and ask which technique it uses or why it may fail.
Categories and tests
- Narrow or weak AI performs a specific task; general or strong AI would match human capability across domains and does not currently exist.
- The Turing test assesses conversational indistinguishability from a human, not intelligence itself.
- Expert systems encode human rules in a knowledge base with an inference engine; they explain their reasoning but do not learn from data by themselves.
Machine learning
- Supervised learning trains on labelled examples to predict labels; unsupervised learning finds structure in unlabelled data, as in clustering; reinforcement learning learns from rewards.
- Features are input variables; overfitting means fitting training noise so that new data performs poorly.
- Data quality drives results: biased or unrepresentative samples produce biased outputs.
- Deep learning uses multilayer neural networks and excels at perception tasks such as vision and speech.
Generative systems and governance
- Large language models predict likely text and can produce confident but false statements, often called hallucination.
- Outputs require verification; human oversight, provenance, and disclosure matter in professional use.
- Automation replaces defined tasks; augmentation supports human judgement.
Worked example
A support team wants to route incoming tickets by topic. They have 20,000 past tickets already labelled by department. Reasoning: labelled outcomes exist, so supervised classification fits, with the department as the label and ticket text as features. Accuracy must be measured on held-out tickets, not the training set, because near-perfect training accuracy with poor new-ticket accuracy signals overfitting. If no labels existed, clustering would be the starting point, with humans naming the clusters.
Traps that cost marks
- Treating AI, machine learning, and deep learning as synonyms rather than nested categories.
- Assuming an expert system learns from new data automatically.
- Judging a model by training accuracy alone.
- Believing fluent generated text is necessarily accurate.
- Ignoring bias in training data when evaluating fairness.