EXIN BCS AI Essentials cheat sheet

AI definition (McCarthy, 1956 Dartmouth workshop): the science and engineering of making intelligent machines, especially intelligent computer programs.

Turing Test (1950): a human interrogator chats blind with a machine and a person; the machine passes if the interrogator cannot reliably tell which is which. It tests behaviour, not inner consciousness.

Chinese Room (Searle, 1980): symbol manipulation following rules can mimic understanding without comprehension; supports the weak-AI position that syntax is not semantics.

Strong AI vs weak AI: strong AI claims a machine can have a mind and genuine understanding; weak AI claims machines only simulate intelligent behaviour.

Narrow AI performs one task domain (spam filters, chess engines); general AI would match human capability across domains; superintelligence would exceed it.

Laws of robotics (Asimov, fiction): a robot may not injure a human, must obey humans, and must protect its own existence, with priority in that order.

Historical milestones: perceptron (Rosenblatt, 1958), ELIZA chatbot (Weizenbaum, 1966), expert systems boom in the 1980s, deep-learning revival after 2012.

AI winters: funding collapses after over-promising, notably around 1974-1980 and 1987-1993; each followed a shift in method (symbolic to statistical).

Human vs machine strengths: humans excel at common sense, context, and transfer; machines excel at speed, scale, consistency, and handling high-dimensional data.

AI benefits: automation of repetitive work, 24/7 operation, pattern detection at scale, faster diagnosis, personalisation, and lower marginal cost per decision.

Bias arises when training data under-represents groups or encodes historical discrimination; the model reproduces and can amplify the skew.

Black-box problem: deep networks give outputs without interpretable reasons, which conflicts with duties to explain decisions (e.g. credit refusal, medical triage).

Accountability gap: when an autonomous system causes harm, responsibility must be assigned among developer, deployer, operator, and user.

Data privacy risk: models trained on personal data can memorise and leak it; GDPR principles include purpose limitation, minimisation, and a right to explanation.

Automation risk: task-level substitution of jobs, plus skill erosion when operators defer to the system (automation bias).

Adversarial and distributional risk: small crafted input changes can flip a classifier; performance degrades when live data drifts from training data.

EU AI Act risk tiers: unacceptable (banned), high-risk (conformity assessment), limited-risk (transparency duties), minimal-risk (unregulated).

Human oversight controls: keep a human in the loop for high-stakes decisions, log decisions, test for bias, and define a fallback to manual process.

Supervised learning maps labelled input-output pairs to a function; at inference it predicts labels for unseen inputs. Tasks: classification (discrete label) and regression (continuous value).

Unsupervised learning finds structure in unlabelled data: clustering (k-means, hierarchical), dimensionality reduction (PCA), and association rules.

Reinforcement learning: an agent takes actions in an environment, receives rewards, and learns a policy maximising expected cumulative reward; trade-off is exploration vs exploitation.

Training set, validation set, and test set are disjoint; the test set is touched once. Overfitting = low training error, high test error; underfitting = high error on both.

Bias-variance trade-off: a too-simple model underfits (high bias); a too-complex model overfits (high variance). Regularisation and more data reduce variance.

Neural network basics: layers of weighted sums plus a non-linear activation; loss is minimised by gradient descent with backpropagation. Deep = many layers.

Confusion matrix metrics: accuracy = (TP+TN)/total; precision = TP/(TP+FP); recall = TP/(TP+FN); F1 = harmonic mean of precision and recall.

Model types: decision trees split on feature thresholds; k-NN classifies by nearest labelled neighbours; SVM finds a maximum-margin separating boundary.

Data quality rule: garbage in, garbage out. Feature selection, labelling consistency, and handling missing values matter more than model choice in many projects.

Human-in-the-loop: the system proposes, the human disposes; required where errors are irreversible or legally significant.

Human-on-the-loop: the system acts and the human monitors with power to intervene; suits high-volume but recoverable decisions.

Complementarity principle: allocate tasks by comparative strength (machine for scale and speed, human for judgement and ethics) rather than automating whole jobs.

Explainable AI (XAI) methods: feature importance, LIME and SHAP local explanations, counterfactuals, and model cards documenting limits and intended use.

AI as augmentation: decision support, co-pilots, and triage tools raise throughput while the human retains accountability for the outcome.

Sociotechnical view: outcomes depend on process, incentives, training, and culture as much as on the algorithm; deployment is an organisational change.

Governance checklist: risk assessment, data provenance, bias testing, human oversight, monitoring for drift, incident reporting, and periodic audit.

AI literacy need: operators must know a model's confidence, failure modes, and data scope to judge when to override it.

Long-term open questions: goal alignment, value lock-in, concentration of capability, and whether general intelligence is achievable.