AI learning library
AI Glossary: 150+ AI Terms Explained in Plain English
A plain-English glossary of the AI terms that actually come up at work — from tokens and RAG to the EU AI Act. Compiled from Google's ML glossary, NIST and Regulation (EU) 2024/1689.
162 terms
No term matches that. Try a shorter word, or .
A
Action Space
The complete set of things an agent is permitted to do — its tools, APIs, and permissions.
Activation Function
The non-linear step applied at each unit, without which a deep network would collapse into a single linear one.
Adversarial Example
An input perturbed just enough to fool a model, while looking unchanged to a person.
Agent Memory
How an agent retains information across steps and sessions, beyond the context window.
Agentic AI
AI that acts — plans, uses tools, self-corrects — rather than only producing text.
Agentic Loop
The cycle an agent repeats: reason about what to do, act, observe the result, decide again.
AI Agent
Software that plans and takes actions toward a goal on your behalf, rather than only answering.
AI Alignment
The problem of making AI systems pursue what we actually want, including things we never thought to specify.
AI Governance
The internal structure of decisions, owners, and controls for how an organisation builds and uses AI.
AI Incident
An event where an AI system causes or nearly causes harm — increasingly something you must record and report.
AI Literacy
Under the EU AI Act: the skills and understanding staff need to use AI systems knowingly — and a legal obligation since February 2025.
AI Regulatory Sandbox
Under the EU AI Act: a supervised environment where innovators can develop and test AI systems under a regulator's eye.
AI Safety
The field concerned with preventing AI systems from causing harm, from everyday failures to systemic risks.
AI Slop
Low-quality AI-generated content produced for volume, not value.
AI System (legal definition)
Under the EU AI Act: a machine-based system that operates with some autonomy and infers outputs from inputs.
Algorithmic Bias
When a system produces systematically different outcomes for different groups, without justification.
API API
The interface your code uses to send a prompt to a model and get a response back.
API Key
The secret credential that authenticates and bills your API calls — and must never reach a browser.
Artificial General Intelligence AGI
A hypothetical system matching human breadth across essentially any cognitive task — no agreed definition, no agreed timeline.
Artificial Intelligence AI
Software that performs tasks we normally associate with human thinking — recognising, predicting, generating, deciding.
Attention Mechanism
The operation that lets a model weigh which parts of the input matter most for each part of the output.
Autonomy
How far a system acts without human involvement — a dial, not a switch.
B
Backpropagation
The algorithm that works out how much each weight contributed to the error, so gradient descent can fix it.
Batch Inference
Running many predictions together as a job instead of one at a time in real time.
Batch Size
How many examples the model processes before updating its weights once.
Benchmark
A standard test set used to compare models — useful for direction, unreliable as a guarantee.
Bias
Systematic skew in a model's outputs — a data and design property, not a moral accusation.
Biometric Categorisation
Under the EU AI Act: assigning people to categories based on their biometric data — restricted, and prohibited for sensitive attributes.
C
CE Marking
The mark declaring that a high-risk AI system conforms to applicable EU requirements.
Chain-of-Thought Prompting CoT
Asking the model to work through its reasoning step by step before answering.
Chunking
Splitting documents into passages small enough to embed and retrieve usefully.
Classification
Predicting which category something belongs to.
Computer Use
An agent operating a computer the way a person does — reading the screen, moving the cursor, typing.
Conformity Assessment
The procedure demonstrating that a high-risk AI system meets the Act's requirements before it goes on the market.
Content Credentials C2PA
Cryptographically signed metadata recording how a piece of media was created and edited.
Context Engineering
Deciding what information goes into the context window, in what order, and what gets left out.
Context Window
The maximum number of tokens a model can consider at once — prompt, documents, and its own answer combined.
Copilot
An AI assistant embedded in a tool that suggests while you keep control.
D
Data Poisoning
Corrupting training data so the resulting model misbehaves, often only on a specific trigger.
Data Provenance
A documented record of where data came from, how it was collected, and what may be done with it.
Deep Learning
Machine learning built on neural networks with many layers — the approach behind almost every modern AI system.
Deepfake
Synthetic audio or video that convincingly depicts a real person doing or saying something they did not.
Deployer
Under the EU AI Act: whoever uses an AI system under their own authority in a professional capacity.
Differential Privacy
A mathematical guarantee that a result barely changes whether or not any one person's data was included.
Diffusion Model
The architecture behind most AI image and video generation: start from noise, remove it step by step.
Distillation
Training a small model to imitate a large one, keeping most of the quality at a fraction of the cost.
Downstream Provider
Under the EU AI Act: a provider whose system integrates a general-purpose AI model made by someone else.
E
Embedding
A list of numbers representing meaning, where similar things end up close together.
Emotion Recognition System
Under the EU AI Act: a system that identifies or infers emotions or intentions from biometric data — banned in workplaces and schools.
Epoch
One full pass of the training algorithm over the entire training dataset.
EU AI Act
The EU's regulation on artificial intelligence — the first broad, binding AI law, applying by risk tier.
Evaluation (Evals)
Systematic measurement of whether an AI system does what you need — your tests, not the vendor's.
Explainability XAI
Being able to give a human a meaningful account of why a system produced a particular output.
F
Fairness
The requirement that a system treat people equitably — with several mathematical definitions that provably conflict.
Feature
One input variable a model uses to make its prediction.
Few-Shot Prompting
Including several examples in the prompt so the model infers the pattern you want.
Fine-Tuning
Continuing training on your own examples to specialise a general model's behaviour.
Foundation Model
A large model trained broadly once, then adapted to many downstream tasks instead of being built per task.
Frontier Model
A model at or beyond the current state of the art in general capability — the ones regulators watch most closely.
Function Calling
The mechanism by which a model asks your code to run a specific function with specific arguments.
G
GAN (Generative Adversarial Network) GAN
Two networks trained against each other — one generating fakes, one detecting them.
General-Purpose AI Model GPAI
Under the EU AI Act: a model with significant generality that can perform a wide range of tasks and be built into many systems.
Generative AI GenAI
Models that produce new content — text, images, audio, video, code — rather than only classifying or scoring existing content.
GPU GPU
The parallel processor that made deep learning practical, and the scarcest resource in the industry.
Gradient Descent
The optimisation method that finds better parameters by repeatedly stepping downhill on the loss.
Guardrails
Controls that constrain what a model can be asked or allowed to output.
H
Hallucination
Fluent, confident output that is simply false — the defining failure mode of language models.
High-Risk AI System
Under the EU AI Act: a system in a listed sensitive use case, permitted but subject to the strictest obligations.
Human Oversight
Under the EU AI Act: the requirement that people can understand, monitor, override, and stop a high-risk system.
Human-in-the-Loop HITL
A design where a person reviews or approves before an AI action takes effect.
Hybrid Search
Combining keyword and vector search, because each fails where the other works.
Hyperparameter
A setting you choose before training, as opposed to a parameter the model learns during it.
I
Inference
Running a trained model to get an answer — everything that happens after training is done.
Inference Cost
What it costs to run a model in production, billed per token in and per token out.
Interpretability
Understanding how a model works internally, as opposed to just describing its outputs.
ISO/IEC 42001
The international standard for an AI management system — certifiable, unlike the NIST framework.
J
K
L
Label
The correct answer attached to a training example.
Large Language Model LLM
A model trained on huge amounts of text to predict the next token, which turns out to be enough to write, summarise, translate, and reason.
Latency
How long a request takes — the difference between an assistant that feels alive and one that feels broken.
Learning Rate
How big a step training takes each update — the hyperparameter most likely to ruin a run.
LLMOps
MLOps for systems built on language models — prompts, evals, cost, and safety instead of training runs.
LoRA (Low-Rank Adaptation)
A cheap fine-tuning method that trains a small add-on instead of updating the whole model.
Loss Function
The formula that scores how wrong a model's prediction is — the thing training tries to minimise.
M
Machine Learning ML
A way of building software where the system learns patterns from examples instead of following hand-written rules.
Meta-Prompting
Using a model to write, critique, or improve prompts for another model.
Mixture of Experts MoE
An architecture where only a fraction of the model's parameters activate for any given token.
MLOps
The practices for getting machine learning models into production and keeping them working.
Model
The trained artefact — the learned parameters plus the architecture — that turns an input into an output.
Model Card
A short standard document stating what a model is for, how it was evaluated, and where it fails.
Model Context Protocol MCP
An open standard for connecting AI applications to tools and data sources through one common interface.
Model Drift
Silent degradation as the world moves away from the data a model was trained on.
Multi-Agent System
Several specialised agents working together, usually with a coordinator dividing the work.
Multimodal AI
A model that handles more than one kind of input or output — text, images, audio, video.
N
Narrow AI
AI built for a specific task, which is every AI system that currently exists.
Negative Prompt
In image generation, a list of things you want kept out of the result.
Neural Network
A model made of layers of connected units that pass numbers forward, loosely inspired by neurons.
NIST AI Risk Management Framework AI RMF
A voluntary US framework for managing AI risk, organised around Govern, Map, Measure, Manage.
O
On-Device AI
Running a model on a phone or laptop instead of sending data to a server.
One-Shot Prompting
Giving exactly one worked example alongside the instruction.
Open-Weights Model
A model whose trained weights you can download and run yourself — not necessarily open source in the strict sense.
Orchestration
The layer that decides which model, tool, or agent handles each step, and in what order.
Overfitting
When a model memorises its training data instead of learning the pattern, and fails on anything new.
P
Parameter
One of the internal numbers a model adjusts during training; parameter count is a rough proxy for model capacity.
Post-Market Monitoring
Under the EU AI Act: providers must keep watching how their system performs in the real world after launch.
Precision and Recall
The two metrics that actually tell you whether a classifier works when one class is rare.
Prohibited AI Practice
Under the EU AI Act: uses of AI banned outright in the EU, in force since February 2025.
Prompt
The input you give a model — instructions, context, examples, and question, all as text.
Prompt Caching
Reusing the processed form of a repeated prompt prefix so you do not pay full price for it again.
Prompt Engineering
Designing and iterating on prompts to get reliable output — closer to spec-writing than to magic words.
Prompt Injection
Hiding instructions in content the model reads, so a third party effectively takes over the prompt.
Prompt Template
A reusable prompt with variable slots, so a working prompt becomes a repeatable asset.
Provider
Under the EU AI Act: whoever develops an AI system or GPAI model and puts it on the market under their own name or trademark.
Q
R
Rate Limit
The cap a provider puts on how many requests or tokens you can use per interval.
ReAct Pattern
An agent design that interleaves explicit reasoning with actions, one step at a time.
Reasoning Model
A model trained to spend extra compute working through a problem before answering.
Red Teaming
Deliberately attacking your own AI system to find failures before someone else does.
Regression
Predicting a number rather than a category.
Regularization
Any technique that deliberately constrains a model to stop it memorising the training set.
Reinforcement Learning RL
Learning by trial and error, guided by rewards rather than labelled answers.
Responsible AI
The practice of building and deploying AI with its effects on people deliberately considered.
Retrieval-Augmented Generation RAG
Fetching relevant documents first and putting them in the prompt, so the model answers from sources instead of memory.
RLHF (Reinforcement Learning from Human Feedback) RLHF
Training a model on human preference judgements so it becomes helpful and well-behaved rather than merely fluent.
Robustness
Whether a system keeps performing when inputs are noisy, unusual, or deliberately adversarial.
Role Prompting
Assigning the model a role to shape its vocabulary, depth, and framing.
S
SDK SDK
An official library that wraps a provider's API in your programming language.
Self-Supervised Learning
Training where the data generates its own labels — the trick that made large language models possible.
Shadow AI
Employees using AI tools that the organisation has not approved and cannot see.
Speech-to-Text STT
Turning spoken audio into written text.
Structured Output
Forcing the model to return data in a fixed schema so downstream code can parse it.
Supervised Learning
Learning from examples that come with the right answer attached.
Synthetic Data
Training or test data generated by a model rather than collected from the real world.
System Prompt
Standing instructions that set a model's role, rules, and tone for an entire conversation.
Systemic Risk
Under the EU AI Act: risk arising from the high-impact capabilities of the most capable general-purpose models.
T
Technical Documentation
Under the EU AI Act: the file describing how a system was built, tested, and controlled, kept ready for authorities.
Temperature
The setting that controls randomness in generation — low for consistency, high for variety.
Test Data
Data the model has never seen, used once at the end to estimate real-world performance.
Test-Time Compute
Spending more computation at answer time — thinking longer — to get better results without retraining.
Text-to-Image
Generating an image from a written description.
Text-to-Speech TTS
Generating spoken audio from written text.
Text-to-Video
Generating video clips from a written description.
Throughput
How much work a system handles per unit of time — tokens per second, requests per minute.
Token
The unit a model actually reads and writes — usually a word fragment, not a word.
Tokenization
Cutting text into the tokens a model can process — and a quiet source of odd model behaviour.
Tool Use
A model calling external software — search, code, APIs, databases — instead of answering from memory.
Top-p (Nucleus Sampling)
An alternative randomness control that samples only from the smallest set of tokens covering probability p.
TPU TPU
Google's custom accelerator chip, purpose-built for machine learning workloads.
Training Data
The examples a model learns from — the single biggest determinant of what it can and cannot do.
Transformer
The neural network architecture behind essentially every modern language model, built around attention.
Transparency
Making it clear what an AI system is, what it was trained on, and when someone is interacting with it.
Trustworthy AI
An umbrella for the properties an AI system needs to be relied on: valid, safe, secure, accountable, explainable, fair, privacy-enhanced.