EXIN BCS AI Essentials flashcards

33 cards. Read the term and the definition here. The deck at the top of this page is optional.

Turing Test

A 1950 behavioural test: a human interrogator holds text conversations with a machine and a hidden person, and the machine passes if the interrogator cannot reliably identify it.

Chinese Room argument

Searle's thought experiment in which a person follows rules to produce fluent Chinese replies without understanding Chinese, arguing that symbol manipulation alone is not comprehension.

Strong AI

The claim that a suitably programmed computer can possess a genuine mind and real understanding, not merely simulate intelligent behaviour.

Weak AI

The position that machines can only simulate or model intelligent behaviour, and that any apparent understanding is an artefact of rule-following.

Narrow AI

Systems that perform a single task or narrow domain at high competence, such as spam filtering or game playing, without general transfer to other tasks.

AI winter

A period of collapsed funding and reduced interest in AI after inflated promises failed to deliver, notably around 1974-1980 and 1987-1993.

Dartmouth workshop 1956

The founding event of AI as a field, where McCarthy, Minsky, Shannon and others proposed that every aspect of learning could be described precisely enough for a machine to simulate it.

Perceptron

Rosenblatt's 1958 early neural model: it computes a weighted sum of inputs and applies a threshold, learning weights from labelled examples; it cannot solve non-linear problems like XOR.

Algorithmic bias

Systematic unfair error in a model's outputs, typically caused by unrepresentative or historically skewed training data, producing worse outcomes for some groups.

Black-box model

A model whose internal logic is not interpretable by humans, so outputs cannot be explained by tracing the reasoning; common in deep neural networks.

Automation bias

The human tendency to over-trust automated suggestions and stop checking them, which erodes vigilance and lets system errors pass uncorrected.

Adversarial example

An input deliberately perturbed in a way imperceptible to humans that causes a trained model to misclassify it with high confidence.

Data drift

Degradation of model performance because live input data gradually diverges from the distribution the model was trained on, requiring monitoring and retraining.

EU AI Act risk tiers

A regulatory classification of AI uses into unacceptable (prohibited), high-risk (conformity assessment), limited-risk (transparency obligations) and minimal-risk categories.

Accountability gap

The problem of assigning legal and moral responsibility for harm caused by an autonomous system among its developer, deployer, operator and user.

Human in the loop

A control design in which a person reviews and approves each consequential system decision before it takes effect.

Supervised learning

A machine-learning approach that learns a mapping from inputs to outputs using labelled example pairs, then predicts outputs for new unseen inputs.

Unsupervised learning

Machine learning on unlabelled data that seeks structure such as clusters, groupings or reduced dimensions, without any target output to predict.

Reinforcement learning

An agent learns a policy by taking actions in an environment and receiving rewards, balancing exploration of new actions against exploitation of known good ones.

Overfitting

A model fits training data including its noise, giving low training error but poor generalisation and high error on new test data.

Underfitting

A model too simple or insufficiently trained to capture the underlying pattern, producing high error on both training and test data.

Bias-variance trade-off

The tension where simple models err from high bias and complex models err from high variance; the aim is a complexity that minimises total error.

Precision

Of all cases a model flagged as positive, the proportion that truly are positive: TP divided by (TP + FP).

Recall

Of all cases that truly are positive, the proportion the model correctly flagged: TP divided by (TP + FN); also called sensitivity.

F1 score

The harmonic mean of precision and recall, giving a single balanced metric when classes are imbalanced and both error types matter.

Gradient descent

An optimisation method that iteratively adjusts model parameters in the direction that reduces the loss, using the gradient computed by backpropagation.

k-means clustering

An unsupervised algorithm that partitions data into k groups by iteratively assigning points to the nearest centroid and recomputing centroids.

Train-test split

Holding out a disjoint portion of labelled data to estimate generalisation performance, keeping it untouched during model selection and tuning.

Explainable AI

Methods such as feature importance, SHAP, LIME and counterfactuals that make a model's predictions interpretable to users, auditors and regulators.

Human-on-the-loop

A design where the system acts automatically and a human monitors it with the authority to intervene or stop it, suited to high-volume recoverable decisions.

Augmentation principle

Allocating tasks by comparative strength: machines handle scale, speed and pattern detection while humans retain judgement, ethics and accountability.

Model card

A short document describing a model's intended use, training data, performance across groups, and known limitations for deployers and auditors.

AI governance

The policies and controls covering risk assessment, data provenance, bias testing, human oversight, drift monitoring and audit across a system's lifecycle.