Artificial Intelligence and Emerging Technologies
Core concepts
Know the main learning paradigms: supervised (labelled training data, classification/regression), unsupervised (clustering, anomaly detection) and reinforcement (reward-driven agents). Models are developed on training data, tuned on validation data and judged on unseen test data. Overfitting means excellent training performance and poor generalisation; underfitting means the model is too simple to capture the pattern.
Governance concepts: model risk management (inventory, ownership, validation, monitoring, documentation), explainability versus black-box trade-offs, bias arising from unrepresentative or historically discriminatory data, data provenance and privacy, and human-in-the-loop oversight. Automation bias - over-trusting system output - is a recurring control failure.
Generative AI introduces specific risks: hallucination (fluent but fabricated output), prompt injection, leakage of confidential data into prompts, and copyright/IP exposure. Retrieval-augmented generation grounds responses in approved sources but does not eliminate error. Emerging technologies also include RPA (rule-based, brittle to interface change), blockchain (immutable distributed ledger - immutability is a risk as well as a benefit), IoT (expanded attack surface) and quantum computing (future cryptography risk).
Regulatory awareness: risk-tiered AI regulation (e.g. the EU AI Act's prohibited, high-risk, limited-risk and minimal-risk categories) and emerging AI assurance expectations.
Traps in questions
- Assuming AI removes the need for governance or documentation - it increases both.
- Judging a model on training accuracy alone.
- Treating RPA as 'AI' - it executes deterministic rules and fails on exceptions.
- Believing a confident, well-written generative output is verified output.
Worked example
A credit-scoring model achieves 99% accuracy on training data but 62% on held-out test data. This is classic overfitting: the model has memorised noise. It must not be deployed; remediate with more representative data, regularisation or simpler architecture, then revalidate on a fresh sample.
Rule of thumb: validate on data the model has never seen, and keep a human accountable for material decisions.