Artificial intelligence has acquired a vocabulary of its own. This reference provides straightforward explanations of the terms you are most likely to encounter when working with, testing or discussing AI systems.
Search for a term or select a letter to browse the glossary.
A measure of how often a model produces the correct result. In classification, it is normally the proportion of all predictions that are correct. Accuracy can be misleading when classes are badly imbalanced.
The ability of an AI-based system to respond appropriately to changes in its environment, data or operating conditions.
An AI system capable of pursuing a goal by deciding what actions to take, often using software tools, external information or other systems with a degree of autonomy.
A hypothetical form of AI able to perform intellectual tasks across a broad range of domains at a level comparable with, or exceeding, human general intelligence. No generally accepted AGI system currently exists.
The broad field concerned with creating computer systems capable of tasks associated with intelligence, including learning, reasoning, perception, language processing, prediction and decision-making.
An AI system developed by DeepMind for predicting the three-dimensional structures of proteins from their amino-acid sequences. Its success demonstrated the potential of AI to accelerate scientific discovery.
DeepMind's Go-playing AI. In 2016 it defeated champion Lee Sedol four games to one, a landmark achievement because Go had long been regarded as exceptionally difficult for computers because of its enormous search space.
A DeepMind reinforcement-learning system that learned games including chess, Go and shogi largely through playing against itself rather than learning directly from human games.
An input deliberately or accidentally altered in a way that causes an AI model to make an incorrect prediction even though the change may be difficult for a human to notice.
Testing that deliberately challenges an AI system with hostile, unusual or manipulated inputs to expose weaknesses in robustness, security or safety.
Rules devised by science-fiction writer Isaac Asimov governing the behaviour of fictional robots. They became an enduring thought experiment about conflicting objectives, machine autonomy and the difficulty of specifying safe behaviour.
A vehicle capable of performing some or all driving tasks using technologies such as computer vision, sensors, machine learning, localisation and automated decision-making.
A systematic tendency for an AI system to produce outcomes that favour or disadvantage particular results, populations or situations. Bias may originate in data, algorithms, assumptions, human decisions or deployment conditions.
A description applied to systems whose internal reasoning is difficult for people to inspect or understand, even when their inputs and outputs can be observed.
A research project launched in Switzerland in 2005 that used detailed computer modelling to investigate the structure and operation of biological neural circuits.
A conversational generative-AI service introduced by OpenAI in 2022. Its rapid public adoption brought large language models and generative AI into widespread public and commercial awareness.
A thought experiment proposed by philosopher John Searle questioning whether manipulating symbols according to rules is sufficient to constitute genuine understanding.
A machine-learning task in which an input is assigned to one of a set of categories, such as identifying an email as spam or legitimate.
An unsupervised-learning technique that groups data items according to similarities discovered within the data rather than predefined labels.
A collaborative robot designed to operate alongside people, often in manufacturing or logistics, rather than being physically isolated from human workers.
AI techniques that allow computer systems to analyse and interpret images and video. Applications include medical imaging, facial recognition, manufacturing inspection and autonomous driving.
A table showing how predicted classifications compare with actual classifications, typically separating true positives, true negatives, false positives and false negatives.
The amount of information a language model can consider during a single interaction, normally measured in tokens. Material outside the available context cannot simply be assumed to remain available to the model.
A neural-network architecture particularly associated with processing images and spatial information. CNNs played an important role in the rapid development of modern computer vision.
A change over time in the characteristics of operational data. If the data encountered after deployment differs significantly from training data, model performance may deteriorate.
Adding meaningful labels or expected outcomes to data so that it can be used for activities such as supervised machine learning.
Deliberately contaminating training data in an attempt to manipulate the behaviour of the resulting model.
An organised collection of data used for purposes such as training, validating, testing or evaluating an AI system.
An AI research organisation founded in London in 2010 and later acquired by Google. Its work includes AlphaGo, AlphaZero and AlphaFold.
IBM's specialist chess-playing computer. In 1997 it defeated world chess champion Garry Kasparov in a six-game match, becoming an important milestone in the public history of artificial intelligence.
Machine learning using neural networks with multiple processing layers. It has been central to major advances in speech, vision and language processing.
Synthetic or manipulated media, commonly video, images or audio, that uses AI to create a convincing representation of something that did not occur as presented.
An early natural-language program created by Joseph Weizenbaum at MIT in the 1960s. Its famous DOCTOR script mimicked a psychotherapist, demonstrating how readily people can attribute understanding to relatively simple conversational software.
A numerical representation of information such as words, documents or images that allows computational comparison of their characteristics or semantic similarity.
The study and application of principles concerning the responsible development and use of AI, including fairness, accountability, privacy, transparency, safety and human impact.
The extent to which people can obtain an understandable explanation for an AI system's output or behaviour.
Testing in which learning, test design and execution occur together. It can be particularly useful for AI systems whose range of possible behaviours makes complete scripted prediction difficult.
A classification performance measure combining precision and recall into a single value using their harmonic mean.
The degree to which an AI system avoids unjustified differences in treatment or outcomes between individuals or groups.
A measurable property or characteristic supplied to a machine-learning model as input.
Selecting, creating or transforming characteristics of raw data to produce useful inputs for machine-learning models.
Further training a pretrained model on additional data to adapt its behaviour to a particular domain, task or desired style of response.
A result in which a system incorrectly predicts that a condition is absent when it is actually present.
A result in which a system incorrectly predicts that a condition is present when it is actually absent.
A model trained on broad quantities of data that can subsequently be used or adapted for many different downstream tasks.
A machine-learning architecture in which two neural networks compete: a generator attempts to create convincing data while a discriminator attempts to distinguish generated examples from real ones.
AI designed to create new content such as text, software code, images, audio, video or other data in response to input.
Connecting an AI response to defined information or evidence so that its output is based more directly upon relevant source material.
A technical or procedural control intended to constrain AI behaviour, block inappropriate actions or reduce the likelihood of unsafe output.
The fictional AI controlling the spacecraft Discovery One in 2001: A Space Odyssey. HAL became one of popular culture's most influential depictions of the dangers created when machine objectives conflict with human intentions.
AI-generated information that appears plausible but is false, unsupported or invented.
An approach in which human judgement or approval remains part of an AI-driven process rather than allowing the system to operate entirely autonomously.
A very large labelled image dataset that became central to progress in computer vision. The ImageNet competitions helped demonstrate the power of deep neural networks during the early 2010s.
Using a trained model to produce a prediction, classification, generated response or other result from new input.
The degree to which a person can understand how an AI model works or how its internal relationships contribute to its outputs.
An attempt to persuade a generative-AI system to bypass behavioural restrictions or safeguards and produce output that would normally be blocked.
American computer scientist who coined the term "artificial intelligence" and helped organise the 1956 Dartmouth workshop. He also created the Lisp programming language and became one of the foundational figures of AI research.
Methods for representing information about the world in forms that computer systems can use for reasoning or decision-making.
A machine-learning model trained on very large quantities of language data to process and generate text and, increasingly, other forms of information.
Data supplied with known categories, values or expected outputs, commonly used to train supervised-learning models.
Nineteenth-century mathematician whose notes on Charles Babbage's Analytical Engine included what is often described as the first published computer algorithm. Her observations about whether machines could originate ideas later became relevant to debates about machine intelligence.
A branch of AI in which models learn patterns from data rather than relying entirely on rules explicitly programmed by developers.
A testing technique useful when exact expected outputs are difficult to determine. Instead of checking one absolute answer, testers check whether defined relationships between multiple inputs and outputs hold.
The computational representation created through training and used to make predictions, classifications, decisions or generated outputs.
Deterioration in a model's effectiveness over time as the real world changes from the conditions represented during development.
AI capable of processing or generating more than one type of information, such as combinations of text, images, audio and video.
AI techniques concerned with analysing, interpreting or generating human language.
A machine-learning structure consisting of interconnected processing units arranged in layers whose parameters are adjusted during training.
A technology company whose graphics processing units and associated computing platforms have become important infrastructure for training and operating modern AI models.
An AI research and technology organisation known for the GPT family of models and ChatGPT, among other AI systems.
The difficulty of determining the correct expected result for a test. It is particularly significant in AI because acceptable behaviour may be probabilistic or difficult for humans to calculate independently.
When a model learns its training data too closely and performs well on that data but poorly on new examples.
A US software company known for platforms that integrate and analyse large quantities of organisational data. Its products are used in government, defence and commercial environments and increasingly incorporate AI-driven capabilities.
Of the cases a classifier predicted as positive, the proportion that really were positive.
Input supplied to a generative-AI system to instruct, question or otherwise influence the output it produces.
The deliberate design and refinement of prompts to improve the usefulness, reliability or structure of generative-AI responses.
An attack in which instructions supplied directly or indirectly to a generative-AI system attempt to override its intended instructions or manipulate its behaviour.
A broad term for research combining quantum computing with artificial intelligence or machine learning. Much of the field remains experimental.
An approach in which a generative-AI system retrieves relevant external information and supplies it as context when generating a response.
Of all the genuinely positive cases, the proportion that a classifier correctly identifies as positive.
Deliberately attacking or challenging a system from an adversarial perspective to discover vulnerabilities, unsafe behaviours, unexpected responses or ways of defeating safeguards.
Machine learning in which an agent learns behaviour through interaction with an environment, using rewards or penalties as feedback.
A machine-learning task concerned with predicting numerical values, such as demand, temperature, price or duration.
An approach to developing and using AI that considers issues such as safety, fairness, transparency, privacy, accountability and human oversight.
The ability of an AI system to continue behaving acceptably when faced with noise, unusual inputs, environmental changes or other challenging conditions.
A road vehicle using sensors, perception systems, automated decision-making and control software to perform driving functions with reduced or potentially no direct human control.
A hypothetical future point at which technological or machine intelligence advances so rapidly that subsequent developments become extremely difficult for humans to predict.
The fictional military AI in the Terminator franchise that becomes self-aware and turns against humanity. Skynet has become shorthand in popular discussion for fears about uncontrolled autonomous AI.
A language model designed with substantially fewer parameters or lower computational requirements than very large models, potentially making it easier to run locally or for specialised purposes.
Machine learning using examples for which the expected output is known, allowing a model to learn relationships between inputs and labelled results.
Artificially generated data designed to reproduce useful characteristics of real-world data. It can supplement scarce data, create unusual test cases or reduce some privacy concerns.
A generation setting commonly used to influence how varied or predictable a language model's token selection is. Higher values generally permit more variation.
Data kept separate from model training and used to provide an independent assessment of the resulting model's performance.
A mechanism or source used to determine the expected result of a test so that the observed outcome can be judged correct or incorrect.
A unit of information processed by a language model. A token may represent a word, part of a word, punctuation or another fragment depending on the tokenisation method used.
The process of converting text or other input into tokens that a model can process.
Data used during the process of creating or adjusting a machine-learning model.
A neural-network architecture introduced in 2017 that uses attention mechanisms to process relationships within sequences. Transformers underpin most modern large language models.
The degree to which relevant information about an AI system's purpose, operation, data, limitations and behaviour is available and understandable.
British mathematician and computing pioneer whose 1950 paper Computing Machinery and Intelligence posed the question "Can machines think?" and introduced what became known as the Turing Test.
A test derived from Alan Turing's "imitation game" in which a human evaluator converses through text and attempts to distinguish a machine from a human participant.
When a model has failed to learn sufficient structure from its training data and consequently performs poorly even on relatively familiar examples.
Machine learning in which the supplied training data does not contain predefined correct labels and the model attempts to discover patterns or structure for itself.
Data separated from the main training data and used during model development to evaluate choices and tune model configuration.
A database designed to store and efficiently search numerical vector representations such as embeddings. Vector databases are commonly used in semantic search and RAG systems.
An IBM question-answering system that gained widespread attention in 2011 when it defeated leading human contestants on the US television quiz show Jeopardy!.
A weapon system capable of performing some targeting or engagement functions without direct human intervention. Such systems raise significant technical, ethical, legal and political questions.
A broad field concerned with techniques that help people understand or obtain explanations for the behaviour and outputs of AI systems.
Computer scientist whose research made major contributions to neural networks and convolutional networks, particularly their application to computer vision.
Asking a generative model to perform a task without first supplying worked examples of the desired task.
A period in which enthusiasm, investment and research funding for artificial intelligence declined sharply after earlier expectations proved unrealistic. Two major AI winters are generally associated with the 1970s and the late 1980s/early 1990s.
The challenge of ensuring that an AI system's goals and behaviour remain consistent with intended human objectives and values, particularly as systems become more capable or autonomous.
An AI company founded in 2021 by former OpenAI employees. It develops the Claude family of AI models and has placed particular emphasis on AI safety and controllability.
A technique that allows a neural network to give different importance to different parts of its input when producing an output. Attention is a fundamental component of transformer models.
The use of technology to perform activities with reduced human intervention. Automation does not necessarily involve AI: a conventional program following fixed rules can be highly automated without learning or intelligent decision-making.
Ridley Scott's 1982 science-fiction film, based on a Philip K. Dick novel, explores artificial beings known as replicants and questions what distinguishes human beings from sophisticated artificial life.
A software program that performs automated tasks. Some bots use AI, while others simply execute predefined rules. Chatbots are bots specifically designed for conversational interaction.
A term associated with intermediate reasoning used by language models when solving multi-step problems. Prompting techniques may encourage models to reason through a problem before providing an answer, although displayed explanations should not automatically be assumed to reveal the model's actual internal processing.
A family of generative-AI models and conversational assistants developed by Anthropic, capable of working with text, code, documents and other forms of information.
The use of software, increasingly including AI, to help healthcare professionals assess information, identify possible conditions, prioritise cases or select appropriate treatments. Such systems normally support rather than replace clinical judgement.
A name used by Microsoft and GitHub for AI assistants embedded within products and development environments. Depending on the product, Copilot can generate text, analyse information, assist with software development and perform other tasks.
A 1956 research workshop at Dartmouth College widely regarded as a founding event in artificial intelligence as an academic discipline. The proposal for the workshop used the term "artificial intelligence".
The android officer Data in Star Trek: The Next Generation explores questions about machine consciousness, emotion, rights and personhood. Unlike many fictional AIs, Data is primarily portrayed as seeking to understand humanity rather than threatening it.
Software that assists users through conversational or other interfaces, for example by answering questions, finding information, controlling devices or performing tasks.
The application of AI to activities such as identifying potential drug targets, predicting molecular properties, designing candidate compounds and analysing biological data. AI can help researchers explore possibilities more quickly, although laboratory and clinical validation remain essential.
An AI system that applies encoded specialist knowledge and rules to problems in a particular domain. Expert systems were one of the most commercially important forms of AI during the 1980s.
A 2014 science-fiction film centred on Ava, a highly sophisticated humanoid AI. The story explores manipulation, consciousness, intelligence, confinement and the difficulty of determining what an apparently intelligent machine actually intends.
Computer scientist known particularly for her work in computer vision and for leading the development of ImageNet, the large image dataset that played an important role in advances in visual recognition and deep learning.
A machine-learning approach in which models are trained across distributed devices or data sources without requiring all of the raw training data to be collected in one central location.
The use of analytical and machine-learning techniques to identify transactions or behaviour that differ from expected patterns and may indicate fraud. It is a long-established commercial application of machine learning.
Computer scientist whose research made major contributions to neural networks and deep learning. Hinton, Yoshua Bengio and Yann LeCun shared the 2018 ACM Turing Award for work that helped make deep neural networks a central technology in modern AI.
Google's family of multimodal generative-AI models and associated AI assistant. Gemini models can work across forms of information including text, images, audio, video and software code.
Generative Pre-trained Transformer: the name used for a family of language and multimodal models developed by OpenAI. GPT models helped demonstrate how large pretrained transformer models could perform a very wide range of language-related tasks.
A processor originally developed primarily for graphics calculations. Its ability to perform many calculations in parallel makes the GPU particularly useful for training and operating neural networks.
British computer scientist, neuroscientist and entrepreneur who co-founded DeepMind. He has been closely associated with systems including AlphaGo and AlphaFold and shared the 2024 Nobel Prize in Chemistry for work relating to protein structure prediction.
A 2013 science-fiction film in which a man develops a relationship with an AI operating system called Samantha. The film examines emotional attachment, consciousness and what relationships with increasingly capable artificial agents might mean.
An AI company and collaborative platform widely used for sharing and accessing machine-learning models, datasets and development tools. Its model hub has become important infrastructure within the open AI community.
The use of computing and AI to enhance human capability rather than replace human decision-makers. Examples include systems that help clinicians interpret scans or help analysts explore large datasets.
A technique for adapting large pretrained models by training a relatively small set of additional parameters rather than retraining all of the model's existing parameters.
A program developed in the 1950s by Allen Newell, Herbert Simon and Cliff Shaw that proved mathematical theorems. It is commonly regarded as one of the earliest artificial-intelligence programs.
The 1999 science-fiction film depicts intelligent machines controlling humanity within a simulated reality. Its imagery has become deeply embedded in popular discussions of AI, virtual reality and the distinction between simulated and physical experience.
The application of computer vision and machine learning to medical images such as X-rays, CT scans, MRI scans and pathology images to help detect, classify or measure abnormalities.
A generative-AI service best known for creating images from natural language prompts. It became one of the prominent systems associated with the rapid development of AI-generated imagery.
American computer scientist and one of the early leaders of artificial intelligence research. He co-founded the MIT Artificial Intelligence Laboratory and contributed to research into knowledge, cognition and machine intelligence.
Structured documentation describing an AI model, potentially including its intended uses, performance, limitations, training information and important ethical or operational considerations.
A model architecture containing multiple specialised components or "experts", with only selected experts being activated for a particular input. This can increase model capacity without requiring every part of the model to be used for every request.
AI designed to perform particular tasks or operate within particular domains rather than possessing general human-like intelligence. Virtually all deployed AI systems are forms of narrow AI.
The automated production of human-readable language from data, structured information or model representations.
An early artificial-neuron model developed by Frank Rosenblatt in the 1950s. The perceptron became an important milestone in the development of neural networks and machine learning.
A value within a machine-learning model that is adjusted during training. Modern neural networks may contain millions or billions of parameters.
Using sensor data and analytical models to predict when machinery or equipment is likely to require maintenance, allowing intervention before a failure occurs.
A system that predicts items a user may find relevant based on factors such as previous behaviour, similarities between users, item characteristics and contextual information. Recommendation systems are widely used in retail, entertainment and social media.
American psychologist and computer scientist best known for developing the perceptron, an early learning algorithm and artificial-neuron model that helped lay foundations for later neural networks.
The combination of robotics with AI capabilities such as perception, learning, planning and decision-making, enabling machines to respond more flexibly to their environment.
A technique for adapting model behaviour using human judgements about the quality or desirability of model outputs. It has been widely used in the development of conversational language models.
American researcher whose work spanned economics, psychology and computing. With Allen Newell and others he developed some of the earliest AI programs and made major contributions to theories of problem-solving and decision-making.
An influential early natural-language program developed by Terry Winograd around 1970. It could respond to commands and questions about a simplified virtual "blocks world", demonstrating impressive language behaviour within a tightly constrained domain.
An approach to artificial intelligence based on explicitly represented knowledge, symbols, rules and logical reasoning. It dominated much early AI research and contrasts with many modern data-driven machine-learning approaches.
Technology that converts spoken language into text or machine-readable commands. Modern speech-recognition systems commonly use deep neural networks.
Reusing knowledge learned by a model for one task or dataset to help solve another related task, reducing the need to train a new model entirely from scratch.
The process through which a machine-learning model adjusts its parameters using data so that it becomes better at performing the intended task.
The idea that people's emotional response to an artificial human-like character or robot may become strongly negative when it looks almost, but not quite, convincingly human.
A fictional test in Blade Runner used to distinguish replicants from humans by measuring responses to emotionally provocative questions. It provides an interesting fictional inversion of the Turing Test: the examiner is attempting to detect an artificial being that already appears convincingly human.
A multimodal AI model capable of working with both visual and language information, enabling tasks such as describing images, answering questions about diagrams or reasoning about visual content.
A 1983 film in which a teenager accidentally interacts with a military computer controlling nuclear-war simulations. Its fictional WOPR system eventually learns the futility of nuclear conflict by exploring possible outcomes — an early popular depiction of machine learning, simulation and automated military decision-making.
A numerical parameter within a neural network that determines the influence one signal has upon another. Training adjusts large numbers of weights to improve model performance.
Computer scientist whose research made major contributions to deep learning and neural networks. Bengio shared the 2018 ACM Turing Award with Geoffrey Hinton and Yann LeCun.
The principle that people or organisations remain responsible for the decisions, behaviour and consequences of AI systems. Deploying an automated system does not remove the need to establish who is answerable when something goes wrong.
A mathematical function applied within an artificial neuron that helps determine its output. Activation functions allow neural networks to learn complex, non-linear relationships.
European Union legislation establishing rules for the development and use of artificial intelligence. It uses a risk-based approach, imposing different requirements according to the nature and potential impact of an AI system.
The policies, responsibilities, processes and controls through which an organisation directs and oversees its use of artificial intelligence. Governance may cover issues including risk, ethics, data, security, compliance, monitoring and accountability.
The knowledge and skills needed to understand, use and critically evaluate AI systems. AI literacy includes recognising both what AI can do and where its outputs require caution or human judgement.
The field concerned with preventing AI systems from causing unacceptable harm, whether through technical failure, misuse, unintended behaviour or objectives that do not match human intentions.
The use of AI to identify suspicious activity, detect malware, analyse security events, prioritise threats and support incident response. AI can also be used offensively, creating an evolving contest between attackers and defenders.
The use of AI for activities such as adaptive learning, tutoring, feedback, content generation, assessment support and learning analytics. Its use also raises questions about accuracy, authorship, assessment integrity and appropriate human oversight.
Applications of AI in farming include crop monitoring, disease detection, yield prediction, autonomous machinery and precision application of water, fertiliser and pesticides.
A method widely used to train neural networks. The error produced by the network is propagated backwards through its layers so that weights can be adjusted to improve future predictions.
A dataset in which relevant classes or categories are represented in suitable proportions. Severe imbalance can cause apparently strong overall performance while important minority cases are handled poorly.
A group of training examples processed together during one step of model training. Batch size can affect training speed, memory use and learning behaviour.
A situation in which some classes occur much more frequently than others in a dataset. This is important because a model can achieve high apparent accuracy simply by favouring the common class.
A change over time in the relationship between input data and the outcome being predicted. Unlike simple data drift, the meaning of the relationship itself has changed.
A measure of how thoroughly selected aspects of an AI system have been exercised by testing. Traditional code coverage may still be useful, but AI testing can also consider data, model behaviour, scenarios, neuron activation or other forms of coverage.
A technique for estimating model performance by repeatedly training and evaluating using different portions of the available data. K-fold cross-validation is a common form.
Increasing the useful variety of training data by creating modified versions of existing examples, such as rotating or cropping images or applying other realistic transformations.
A situation in which information unavailable during genuine operation improperly influences model training or evaluation. Leakage can make a model appear far more effective during development than it will be in real use.
The sequence of processes through which data is collected, transferred, cleaned, transformed, stored and supplied to an AI or machine-learning system.
The activities required to make raw data suitable for use, potentially including cleaning, labelling, transformation, feature selection, balancing and splitting into training, validation and test datasets.
The degree to which data is suitable for its intended purpose. Relevant characteristics can include accuracy, completeness, consistency, validity, representativeness and timeliness.
A difference between the statistical characteristics of data used during model development and data encountered elsewhere, particularly during real-world operation.
The property whereby the same starting conditions and inputs produce the same result. Some AI systems are deliberately probabilistic, so repeated execution may legitimately produce different outputs.
One complete pass through the available training dataset during model training. Neural networks commonly train over multiple epochs.
The systematic assessment of a model using appropriate data, metrics and other evidence to determine how effectively it satisfies its intended objectives.
The extent to which generated statements are factually correct. Fluency and confidence are not evidence of factuality, so generative-AI outputs may require independent verification.
The degree to which generated output is supported by and remains consistent with the source material or context on which it is supposed to be based.
Computer-vision technology used to identify or verify people from facial images. Applications range from device authentication to policing, making accuracy, demographic performance, privacy and governance particularly important.
An optimisation technique used during machine learning to adjust model parameters in directions intended to reduce prediction error or another defined loss.
Assessment of AI outputs by people rather than entirely by automated metrics. Human evaluation can be important when characteristics such as usefulness, appropriateness or quality are difficult to capture numerically.
A configuration value selected for the learning process rather than learned directly from the training data. Examples can include learning rate, batch size and aspects of model architecture.
Data in which some relevant outcomes or classes are much less common than others. Fraud, disease and equipment failure are examples where the cases of greatest interest may also be the rarest.
Testing concerned with the data supplied to an AI system, including its format, validity, quality, range, representativeness and handling of unusual, missing or erroneous values.
A pretrained model further adapted using examples of instructions and desirable responses so that it becomes better at following user requests.
A training hyperparameter controlling how large an adjustment is made to model parameters during optimisation. A poor choice can make learning excessively slow or unstable.
The use of a language model to evaluate outputs produced by an AI system according to defined criteria. It can make large-scale evaluation practical, but the evaluating model can itself introduce bias, inconsistency and error.
A mathematical measure of how poorly a model is performing against a training objective. Training attempts to adjust model parameters so that this loss is reduced.
Testing focused on whether a machine-learning model performs adequately for its intended purpose, including its predictive performance, robustness, behaviour on important subgroups and response to unusual conditions.
A standard task, dataset or collection of evaluations used to compare the capabilities or performance of AI models. Benchmark results are useful indicators but do not guarantee equivalent performance in a different real-world context.
Potential degradation that can occur when successive models are trained heavily on generated rather than original data, causing important characteristics or less common parts of the original distribution to be progressively lost.
Ongoing observation of an AI model after deployment to identify changes in performance, data, behaviour, errors or other indicators that may require investigation or retraining.
Behaviour in which identical or apparently identical inputs can produce different outputs. This complicates traditional testing because a single fixed expected result may not be appropriate.
Technology for converting text contained in images, scans or photographs into machine-readable characters. Modern OCR frequently incorporates machine-learning and computer-vision techniques.
The ability to understand the operational state and behaviour of an AI system through information such as inputs, outputs, performance measures, traces, logs and changes in data.
Behaviour involving probabilities rather than fixed outcomes. Many AI systems produce likelihoods or select from several possible outputs, which changes how their correctness should be evaluated and tested.
A quantitative measure used to assess some aspect of model performance. The appropriate metric depends upon the problem: accuracy alone, for example, may be dangerous when false negatives and false positives have very different consequences.
The extent to which a dataset adequately reflects the population, situations and conditions in which the resulting AI system is expected to operate.
A generative model designed or trained to devote additional computational effort to multi-step reasoning before producing an answer, particularly for problems involving mathematics, logic, planning or complex analysis.
Behaviour in which an AI system finds an unintended way to maximise a defined reward without achieving what its designers actually wanted. It illustrates the danger of specifying a measurable target that is only an imperfect proxy for the real objective.
Selecting a subset of a larger population or collection of data. Sampling methods matter because systematic omissions or distortions can produce datasets that do not represent real operation.
The proportion of genuinely positive cases correctly identified as positive. It is also commonly called recall or the true-positive rate.
The proportion of genuinely negative cases that a classifier correctly identifies as negative.
Testing in which conclusions are based on behaviour observed across a suitable collection of test cases rather than requiring every individual execution to produce one predetermined result.
A hypothetical form of artificial intelligence whose intellectual capabilities substantially exceed those of humans across most or all relevant domains. ASI remains a theoretical concept rather than an established technology.
The cut-off used to convert a model's score or probability into a classification decision. Changing the threshold can alter the balance between false positives and false negatives.
A quality concern involving generated content that is abusive, threatening, hateful or otherwise harmful. Toxicity evaluation is commonly included in safety testing of generative-AI systems.
A result in which a classifier correctly identifies that the condition being tested for is absent.
A result in which a classifier correctly identifies that the condition being tested for is present.
In machine learning, sensitivity to differences in the training data. Excessive variance can cause a model to fit its training examples too closely and generalise poorly to new data.
The ability of a model to perform a task or recognise a category without having been trained on labelled examples of that exact task or category. The term is related to, but broader than, zero-shot prompting.
AI designed to pursue objectives through sequences of actions rather than merely producing a single response. Agentic systems may plan, obtain information, use tools, evaluate results and decide what to do next with varying degrees of autonomy.
A neural network trained to encode input into a compressed representation and then reconstruct the original input. Autoencoders can be used for dimensionality reduction, representation learning and anomaly detection.
Identifying observations or patterns that differ significantly from normal behaviour. Applications include fraud detection, cybersecurity, manufacturing quality and equipment monitoring.
AI is used to analyse the enormous quantities of information produced by telescopes and space missions, helping researchers classify astronomical objects, detect unusual events and identify possible planets around other stars.
The balance between a model being too simple to capture important relationships and being so sensitive to its training data that it fails to generalise well. It is closely associated with underfitting and overfitting.
AI techniques are used alongside conventional scientific models to analyse environmental data, improve aspects of weather forecasting, model climate processes and identify patterns in large Earth-observation datasets.
A mathematical measure of the similarity between two vectors based on the angle between them. It is frequently used to compare embeddings in semantic search and retrieval systems.
A machine-learning model that makes predictions by following a series of branching decisions based on characteristics of the input data. Decision trees can be used for both classification and regression.
A generative-model approach that learns to create data by progressively reversing a noise-adding process. Diffusion models became particularly important in modern AI image generation.
A digital representation of a physical object, process or system that can be updated using real-world data. AI may be combined with digital twins for prediction, optimisation, simulation and maintenance.
Running AI processing on or close to the device where data is produced rather than sending everything to a remote data centre. This can reduce latency, network dependence and some privacy risks.
Behaviour or capability that becomes apparent in a complex system without having been explicitly programmed as a separate feature. The term is often used when discussing capabilities observed as AI models become larger or more complex.
Combining predictions from multiple models in order to produce a stronger overall result. Random forests, bagging and boosting are examples of ensemble approaches.
Supplying a generative model with a small number of examples showing the desired task or response pattern before asking it to perform the task itself.
A model's ability to perform effectively on new data rather than merely reproducing patterns memorised from its training examples.
An ensemble-learning technique that builds a sequence of models, commonly decision trees, with each model attempting to improve upon errors made by those built previously.
The use of technologies such as OCR, machine learning and language processing to classify documents and extract useful information from items such as forms, invoices, contracts and correspondence.
An unsupervised-learning algorithm that divides observations into a specified number of clusters by repeatedly assigning them to nearby cluster centres and updating those centres.
A form of recurrent neural network designed to retain useful information over longer sequences. LSTMs were widely used for language, speech and time-series tasks before transformers became dominant in many of those areas.
An open protocol for connecting AI applications to external tools and sources of context through a standard interface. It is intended to reduce the need for separate custom integrations between every model, application and external service.
The automated translation of text or speech from one language into another. Modern translation systems commonly use neural networks and transformer architectures.
A basic processing unit in a neural network. It combines incoming values using learned weights and normally applies an activation function before passing a result onward.
Providing a generative model with one example of the required task or response format before asking it to produce a new result.
An AI model whose trained parameter weights are made available for others to download or use. Open weights do not necessarily mean that the model's training data, source code or entire development process is open.
The number of adjustable parameters contained within a model. It is often used as a rough indication of model scale, although a larger parameter count does not by itself prove that one model is more capable or useful than another.
Reducing the numerical precision used to represent model parameters or calculations. Quantisation can reduce memory and computing requirements, making models easier to run on less powerful hardware.
An ensemble machine-learning method that combines predictions from many decision trees. Random forests are widely used for classification and regression and can perform well on structured data.
A neural-network architecture designed to process sequential information by allowing information from earlier processing steps to influence later ones.
A graph showing the relationship between the true-positive rate and false-positive rate of a binary classifier as its classification threshold is varied.
The area under a receiver operating characteristic curve. It provides one summary measure of a classifier's ability to distinguish between two classes across different thresholds.
Searching for information according to meaning rather than relying solely on exact keyword matches. Embeddings and vector similarity are commonly used to support semantic search.
Instructions supplied by an AI application to establish behaviour, role, constraints or context for a generative model before or alongside the user's own input.
The use of AI to analyse surveillance information, potentially including video, images, location or behavioural data. Applications can support security and public safety but also create significant questions concerning privacy, proportionality and civil liberties.
The ability of an AI model or application to request the use of an external function or service, such as searching a database, performing a calculation, retrieving a document or carrying out an authorised action.
A specialised processor developed by Google for accelerating machine-learning workloads, particularly calculations involving tensors and neural networks.
An ordered collection of numerical values. In modern AI, vectors are widely used to represent information such as words, documents, images or other objects in forms that models can process mathematically.
The fictional supercomputer in Douglas Adams' The Hitchhiker's Guide to the Galaxy, built to calculate the Answer to the Ultimate Question of Life, the Universe, and Everything. Its famously unhelpful answer — 42 — neatly illustrates the importance of asking the right question as well as obtaining an answer.
Mary Shelley's 1818 novel long predates computers, but its themes of creating an intelligent being, responsibility for one's creation and unintended consequences recur repeatedly in modern debates about artificial intelligence.
Karel Čapek's 1920 play introduced the word "robot" to a worldwide audience. Its artificial workers eventually rebel against humanity, establishing themes that have remained part of fictional portrayals of intelligent machines ever since.
A deep convolutional neural network that achieved a dramatic victory in the 2012 ImageNet image-recognition competition. Its success helped trigger the rapid adoption of deep learning for computer vision.
The 2017 research paper that introduced the transformer architecture. Transformers subsequently became the foundation for most modern large language models and many other generative-AI systems.
AI is used in manufacturing for activities including visual quality inspection, process optimisation, predictive maintenance, robotics, production planning and supply-chain management.
Applications include fraud detection, credit assessment, anti-money laundering, customer support, risk analysis and trading. Financial use of AI can have substantial consequences, making governance, explainability and bias important concerns.
Generative AI can assist with or produce writing, images, music, video, design and other creative material. Its rapid adoption has also generated debate about copyright, attribution, employment and the relationship between human and machine creativity.
AI can assist with requirements analysis, programming, code review, test generation, documentation, debugging and maintenance. Generated artefacts still require appropriate review and testing because plausible output is not necessarily correct output.
AI can support activities such as analysing documents, summarising stakeholder information, generating candidate requirements, exploring options and identifying patterns in business data. Human judgement remains essential for context, priorities and stakeholder needs.
An independent implementation or alternative mechanism used to produce results against which the system under test can be compared when a definitive test oracle is unavailable.
Testing in which the same inputs are supplied to two implementations, models or versions and their outputs are compared. Differences can reveal areas requiring further investigation even when the exact correct result is difficult to establish.
Comparing two versions of a system or feature with different groups of users or traffic to determine whether observed behaviour differs according to defined measures.
Testing that considers the distribution of inputs and outputs rather than examining only isolated examples. This is important where the overall statistical behaviour of an AI system matters.
The extent to which a model, experiment or evaluation can be recreated with sufficiently similar conditions and produce appropriately consistent results.