94 terms

Artificial Intelligence Glossary

The key vocabulary of Artificial Intelligence, each term defined in plain language. Start learning in the Artificial Intelligence subject, or drill these terms as flashcards.

Activation function: A function that decides a neuron’s output, letting networks capture non-linear patterns.
Adversarial example: An input altered slightly and deliberately to fool a neural network into a confident wrong answer.
AI agent: An AI system that plans and takes multi-step actions using tools to pursue a goal, not just answer one prompt.
AI governance: The policies, laws, and institutions that guide how AI is developed and used.
AI lifecycle: Define the problem, prepare data, train, evaluate, deploy, monitor, improve — a loop, not a line.
AI winter: A period of collapsed funding and attention following unmet expectations.
Algorithm: A procedure for solving a problem.
Algorithmic bias: Systematic unfairness in an AI system’s outputs, often reflecting bias in its data or design.
Alignment problem: The challenge of making powerful AI systems reliably pursue what humans actually intend.
AlphaFold: An AI system that predicts the 3D structure of proteins, a landmark scientific application of deep learning.
API: An interface through which software systems communicate.
API (application programming interface): A defined way for software to send requests to a service — such as an AI model — and receive responses.
Artificial general intelligence: A hypothetical system able to perform a broad range of intellectual tasks at roughly human level.
Artificial intelligence: The field of building systems that perform tasks associated with intelligence — not a single technology.
Artificial neuron: A basic computing unit that combines weighted inputs and passes the result through an activation to produce an output.
Artificial superintelligence: A theoretical system exceeding human ability across most intellectual domains.
Attention: A mechanism letting a model weigh which parts of the input matter most for each output.
Augmentation: Using AI to enhance and speed up human work rather than replace the human.
Automation: The use of machines or software to perform tasks previously done by people.
Autonomous vehicle: A vehicle that can sense its environment and drive with little or no human input.
Backpropagation: The method of sending prediction error backward through a network to adjust its weights.
Bias: A systematic tendency inherited from data, labelling, objectives, or deployment choices.
Classification: A supervised task that predicts a category, such as spam or not-spam.
Clinical decision support: AI or software that helps clinicians by flagging findings or suggesting options, while the clinician stays responsible.
Clustering: An unsupervised task that groups similar items together without predefined labels.
Confusion matrix: A table of true and false positives and negatives used to evaluate a classifier beyond raw accuracy.
Context: The information available to a model during a particular interaction.
Context window: The amount of text a language model can consider at once; anything beyond it is not available to the model.
Convolutional neural network: A network design well suited to images, detecting simple features and combining them into complex ones.
Data: Recorded information: text, images, audio, measurements, transactions.
Dataset: An organised collection of data.
Deep learning: A branch of machine learning built on multilayer neural networks.
Deepfake: Synthetic media that makes a real person appear to say or do something they never did.
Diffusion model: An image-generation approach that starts from random noise and iteratively refines it toward a prompt.
Discriminative model: A model that predicts a label or category for an input, as opposed to generating new content.
Distribution shift: Deployed conditions drifting from training conditions, degrading performance quietly.
Embedding: A numerical representation positioned so that related things sit near each other.
Evaluation: Testing a model on data it never saw in training; measures learning rather than memorisation.
Expert system: A 1980s approach encoding specialist knowledge as explicit rules; brittle and costly to maintain.
Facial recognition: AI that identifies or verifies a person from an image of their face.
Fairness (in AI): The goal that an AI system treat people equitably; it has several competing, sometimes incompatible, definitions.
Feature: An individual measurable input a model uses to make a prediction.
Fine-tuning: Further training an existing model for a task or domain; persists, unlike prompting.
Foundation model: A large, general-purpose AI model trained on broad data that can be adapted to many tasks.
Fraud detection: Using machine learning to spot unusual patterns that may indicate fraudulent activity.
Generalization: A model’s ability to perform well on new, unseen data — the real goal of machine learning.
Generative AI: AI that creates new content — text, images, audio, or code — rather than only classifying or predicting a label.
Gradient descent: An optimisation method that repeatedly nudges weights in the direction that most reduces error.
Guardrails: Rules and checks placed around an AI system to keep its behaviour safe and within bounds.
Hallucination: Plausible but unsupported or incorrect generated content; an output failure, not an intention.
Hidden layer: A layer between the input and output where a network builds internal representations of the data.
Human-in-the-loop: A design where a human reviews or approves an AI system’s outputs, keeping accountability with people.
Inference: Using a trained model to produce an output; the phase that runs constantly.
Intelligent tutoring system: Software that adapts instruction and feedback to an individual learner.
Label: The correct answer attached to a training example in supervised learning.
Large language model: A model trained on large text collections to process and generate language.
Levels of driving automation: A framework (SAE) describing how much a system drives versus the human, from none to full autonomy.
Liar's dividend: The benefit wrongdoers gain when the existence of deepfakes lets them dismiss real evidence as fake.
Limited-memory system: A system using recent or stored information within defined bounds; where most practical AI sits.
Loss function: A measure of how wrong a model’s predictions are, which training tries to minimise.
Machine learning: The branch of AI in which systems learn patterns from data rather than following written rules.
Minimum viable product (MVP): The smallest useful version of a product, released to learn from real users and iterate.
Model: A trained mathematical system that maps inputs to outputs.
Multimodal model: A model that handles more than one kind of input or output — for example text and images together.
Narrow AI: AI capable within defined tasks; everything in practical use today.
Neural network: A model of many simple units in layers that can learn very complex patterns, at the cost of interpretability.
No-code / low-code: Tools that let people build software or AI workflows with little or no traditional programming.
Open-weight model: A model whose trained parameters are published so anyone can run or adapt it, such as Meta’s Llama family.
Overfitting: When a model learns the noise in its training data and so performs well on it but poorly on new data.
Overreliance: Ceasing to check a system because it is usually right — the risk that grows with accuracy.
Parameter: An internal numerical value learned during training.
Precision: Of the cases a model flagged as positive, the fraction that truly were positive.
Prompt: The input or instructions supplied to a model at inference.
Prompt engineering: The practice of writing and refining prompts to steer a generative model toward better output.
Reactive system: A system responding to current input with no durable memory.
Recall: Of all the truly positive cases, the fraction the model successfully caught.
Recommendation system: An AI system that suggests items — products, videos, songs — based on patterns in user behaviour.
Red-teaming: Deliberately probing an AI system to find failures, biases, and harmful outputs before release.
Regression: A supervised task that predicts a number, such as a price or temperature.
Reinforcement learning: Learning through trial and error, where an agent takes actions and adjusts based on rewards and penalties.
Responsible AI: Developing and using AI with fairness, transparency, accountability, privacy, and human oversight.
Retrieval-augmented generation (RAG): Retrieving relevant documents and giving them to a model so it answers from trusted data (Lewis et al., 2020).
Reward hacking: When an AI optimises the literal specified goal in unintended, undesirable ways.
Supervised learning: Machine learning from labelled examples — inputs paired with the correct answers — used for classification and regression.
Surveillance: The systematic monitoring of people, greatly amplified by AI-driven data collection and recognition.
Synthetic media: Images, audio, or video produced or altered by AI rather than captured from reality.
Text-to-image: Generating an image from a written description.
Token: A unit of text processed by a language model, often a word fragment rather than a whole word.
Tool use / function calling: A model capability that lets it call external tools or functions to get data or take actions.
Training: Adjusting a model's parameters using data and an objective.
Transformer: The architecture introduced in 2017 that underpins modern language models.
Underfitting: When a model is too simple to capture the real pattern, performing poorly on both training and new data.
Unsupervised learning: Machine learning that finds structure in unlabelled data, most commonly by grouping similar items into clusters.
Weight: A tunable number controlling how strongly one unit influences another; learning means adjusting the weights.

← All subject glossaries