EXIN BCS AI Essentials key terms

The field of building computer systems that perform tasks normally requiring human intelligence, such as recognising patterns, understanding language, reasoning, or making decisions. It covers both narrow practical applications and long-term research ambitions.

A thought experiment proposed by Alan Turing in 1950 in which a human judge converses with an unseen machine and a hidden person; if the judge cannot reliably tell them apart, the machine is said to pass. It tests behaviour, not inner experience.

A thought experiment by John Searle arguing that manipulating symbols according to rules, however convincingly, does not produce genuine understanding. It challenges the claim that passing a behavioural test proves a machine has a mind.

The position or goal that a suitably programmed computer could have a mind, consciousness, and real understanding rather than merely simulating intelligent behaviour. Contrasts with weak AI, which treats the computer as a modelling tool.

The view that computers are useful for simulating or studying intelligence without claiming they possess minds. Also used loosely to describe systems built for one specific task rather than general reasoning.

Systems designed to perform one specific task or a small set of related tasks, such as face recognition, spam filtering, or route planning. Nearly all deployed AI today is narrow, with no general reasoning beyond its training.

A hypothetical system able to learn and perform any intellectual task a human can, transferring knowledge across domains. It does not exist today; claims about it are speculative and often debated.

A hypothetical intelligence far exceeding the best human minds across virtually all domains. It is a speculative future scenario used to discuss long-term risk and control problems, not a current technology.

A speculative future point at which machine intelligence improves itself so rapidly that technological and social change becomes unpredictable and beyond human foresight. The idea is philosophical, not an established scientific prediction.

An early approach, often called good old-fashioned AI, that represents knowledge as explicit symbols and rules and reasons by logical manipulation. It excels at well-defined problems but struggles with perception and messy real-world ambiguity.

A rule-based program that captures a specialist's knowledge as if-then rules and applies it to diagnose problems or give advice. Popular in the 1980s, it required costly knowledge engineering and could not learn from data.

The labour-intensive process of extracting expertise from human specialists and encoding it as rules or structured knowledge for a computer system. Its cost and fragility were key reasons symbolic AI fell out of favour.

A period of reduced funding and lost confidence in AI research, notably in the 1970s and late 1980s, after promises went unmet. It illustrates the cycle of hype and disillusionment around the field.

A 1956 summer research project at Dartmouth College widely credited with founding artificial intelligence as a named academic discipline. Its proposal assumed that every aspect of learning could in principle be described precisely enough for a machine to simulate.

An early artificial neuron developed by Frank Rosenblatt in the late 1950s that learns a simple decision boundary from labelled examples. Its limitations, famously critiqued by Minsky and Papert, helped trigger reduced interest in neural networks.

A computing model loosely inspired by connected brain cells, built from layers of simple units whose weighted connections are adjusted during training. Deep networks with many layers underpin modern image, speech, and language systems.

The branch of AI in which systems improve at a task by learning patterns from data rather than following hand-written rules. Performance is measured on new, unseen examples rather than the examples used for learning.

A learning approach that trains a model on labelled examples, each pairing input data with the correct answer, so it can predict labels for new inputs. Common uses include classification and regression.

A learning approach that finds structure in unlabelled data, such as grouping similar items or reducing dimensions. It is used for clustering customers or detecting unusual patterns without predefined answers.

A learning approach in which an agent learns by trial and error, taking actions and receiving rewards or penalties, to maximise long-term reward. It suits problems like game playing, robotics, and control.

A supervised task that assigns each input to one of a set of predefined categories, such as spam or not spam. The model learns from labelled examples and is judged by how often it labels new cases correctly.

A supervised task that predicts a continuous numeric value, such as a price, temperature, or delivery time, rather than a category. It fits a relationship between input features and the target number.

An unsupervised technique that groups data points so items within a group are more similar to each other than to items outside it. It is used for segmentation and exploration when no labels are available.

The examples used to fit a machine-learning model by adjusting its parameters. Its quality, size, and representativeness strongly shape what the model learns, including any biases it picks up.

A separate set of examples, not used during training, on which a model's performance is evaluated. Keeping it separate shows whether the model generalises or has merely memorised the training examples.

When a model learns the training examples too closely, including their noise, and performs poorly on new data. It typically arises from excessive model complexity or too little varied training data.

A model's ability to perform well on new, unseen examples rather than only the data it was trained on. It is the central goal of machine learning and the reason separate test data is used.

Machine learning using neural networks with many layers, which automatically learn increasingly abstract features from raw data. It drives modern speech recognition, image analysis, and large language models but needs large datasets and compute.

A deep-learning model trained on very large text collections to predict and generate language, enabling tasks such as summarising, translation, and question answering. Outputs can be fluent yet factually wrong, a behaviour called hallucination.

The branch of AI concerned with enabling computers to understand, interpret, and produce human language, in text or speech. Applications include translation, chatbots, sentiment analysis, and information extraction.

The branch of AI that extracts meaning from images and video, such as detecting objects, recognising faces, or reading text. It powers medical imaging tools, quality inspection, and autonomous vehicles.

A finite, step-by-step procedure for solving a problem or producing an output from inputs. In machine learning, the algorithm defines how a model is trained, while the model itself holds the learned parameters.

The trained artefact produced by a machine-learning algorithm, containing learned parameters that map inputs to outputs. It is what is deployed to make predictions on new data.

Systematic error or unfair skew in data or model outputs, often reflecting past human prejudice or unrepresentative samples. It can cause AI to disadvantage groups and raises fairness and legal concerns.

The degree to which the workings, data, and limitations of an AI system can be inspected and understood by relevant people. It supports accountability and trust, and is limited in complex opaque models.

The extent to which a person can understand why an AI system produced a particular output. It matters most in high-stakes settings such as credit, hiring, and healthcare, where decisions must be justified.

A system whose internal reasoning is not readily inspectable by users, even if its inputs and outputs are known. Many deep-learning models are described this way, creating challenges for justification and audit.

The principle that identifiable people or organisations remain responsible for an AI system's design, deployment, and effects. It requires clear lines of responsibility so harms can be traced and remedied.

The study and practice of ensuring AI systems behave as intended, avoid harm, and remain under appropriate human control. It spans technical reliability, security, testing, and governance of deployment.

The arrangement by which people monitor, review, and can intervene in or override an AI system's decisions. Meaningful oversight requires that humans have the information, authority, and time to act.

A design in which a person reviews or approves AI outputs as part of the workflow, rather than the system acting fully automatically. It is common in high-stakes decisions such as medical triage or content moderation.

The human tendency to over-trust automated suggestions and accept them uncritically, or to discount them too readily. It can undermine the value of human oversight and is a key concern in safety-critical AI use.

A confident but incorrect or fabricated output from a generative AI system, such as an invented citation or false fact. It arises because such models generate plausible text rather than verifying truth.

The framing of AI as a tool that complements and extends human capability rather than replacing it, with people and machines each contributing strengths. It emphasises collaboration, judgement, and shared decision-making.

A system that perceives its environment and takes actions to achieve goals, either physically, such as a robot, or digitally, such as software that books tasks. Autonomy level varies and raises control questions.