Skip to content

Glossary

Plain-English definitions of the terms used across both courses. Use the search box, or your browser's find (Ctrl/Cmd + F), to jump to a term. The class where a term is introduced is shown as L (AI Literacy) or E (AI Engineering) plus the class number.

A

A/B test, shadow mode, canary release (E7)
Ways to compare or roll out a new version safely on live traffic: split users between versions; run the new version silently alongside the old; release it to a small share first.
A2A (Agent2Agent) (E6)
An open protocol that lets agents from different teams or vendors delegate tasks to each other. Each agent publishes an Agent Card describing what it can do.
Agent (L7, E6)
A language model running in a loop that chooses its own next action (usually a tool call) until the job is done, a limit is hit or a person stops it.
Agent-computer interface (ACI) (E5)
The names, descriptions, schemas, results and error messages through which a model uses a system. Good tool design is good ACI design.
Agent framework (E6)
A library that provides agent loops, state, checkpoints, approval gates and tracing — e.g. LangGraph, OpenAI Agents SDK, Claude Agent SDK, Microsoft Agent Framework, CrewAI (as of 2026).
Agent loop (E6)
The cycle an agent repeats: observe → think → act → observe again.
Agentic RAG (E4)
A model that decides when and what to search, possibly searching several times.
AI / code / human split (L1, E1)
Giving each step of a task one owner. AI proposes, code checks, humans decide.
AI gateway (E3)
A central service between applications and model providers that handles keys, logging, quotas and cost.
AI governance (L8, E8)
The laws, policies, standards and bodies that shape how AI is built and used.
AI winter (L1)
A period when excitement and funding for AI research fell after promises weren't met.
API key (E0)
A secret string that identifies you to a model provider. Treat it like a password.
Artificial intelligence (AI) (L1)
The broad field of making computers do tasks that normally need human intelligence.
Attention (L2, E1)
The transformer's way of working out which other parts of the text matter most for each token.
Automation bias (L6)
The human tendency to over-trust a computer's output, especially when it sounds confident.
Autoregressive generation (E1)
Producing text one token at a time, adding each chosen token before predicting the next.

B

Back-translation (L5)
Translating a translation back into the original language to check that the meaning survived.
Base model (L2, E1)
A model after pre-training only — a text continuer, not yet a helpful assistant.
Bias (L6)
A systematic skew in AI output for or against certain groups, places, languages or viewpoints.
BM25 (E4)
The standard method for scoring how well a document matches a query's keywords.
Budget (E6)
A hard limit, enforced in code, on an AI system's steps, tokens, cost or time.
Byte-pair encoding (BPE) (E1)
A common way to build a tokenizer's vocabulary by repeatedly merging frequent character pairs.

C

Cascade (E6)
Trying a cheap model first and escalating to a stronger one only when a check fails.
Chain of thought (L3, E2)
Getting the AI to work through a problem in visible steps before giving its final answer.
Checkpoint (E6)
A saved snapshot of a run's state after a step, so it can resume without redoing earlier work.
Chunk, chunking, overlap (L4, E4)
A short piece of a document cut out so it can be searched; the act of cutting documents up; words repeated between neighbouring pieces so meaning isn't cut in half. Chunking can be fixed-size, recursive (following the document's structure) or semantic (where the meaning changes).
Circuit breaker (E7)
Stops calling a failing service for a while and fails fast instead of waiting.
Citation / citation accuracy (L4, E4)
A reference showing which source a statement came from / whether each cited source actually supports the claim attached to it.
Closed (proprietary) model (L2, E1)
A model available only through its maker's service.
Coding agent (E3)
An AI that plans, edits files and runs commands such as tests to complete a programming task.
Coding assistant (E3)
An AI tool that suggests or drafts code while you work.
Cohen's kappa (E7)
A measure of agreement between two graders that corrects for agreement expected by chance.
Computational thinking (L1, E1)
Breaking a problem down so that people and machines can carry it out reliably: decomposition, pattern recognition, abstraction and writing clear steps.
Context (L4)
Everything the model can see when it answers: instructions, conversation, pasted text, files and search results.
Context engineering (E4)
Deciding what goes into the model's context window, in what order and within what budget.
Context isolation (E6)
Giving each sub-agent only the information its own step needs.
Context rot / lost in the middle (L4, E4)
Answers get worse as context grows; information buried in the middle of a long context is used less reliably than information at the start or end.
Context window (L2, E1)
The maximum amount of text, measured in tokens, a model can take into account at once.
Cosine similarity (E1, E4)
A measure of how closely two embeddings point in the same direction — i.e. how similar in meaning two texts are.
Cost tracking (E3)
Counting input and output tokens, pricing them, and refusing further calls past a budget.
Custom instructions (L3)
Your own standing instructions (or project settings) that apply to every chat in that space.

D

Data classification (L8, E8)
Labelling information by sensitivity (e.g. public, internal, confidential, strictly confidential) to decide who and which tools may use it.
Data minimisation (L8)
Using only the personal data strictly needed for the purpose.
Data sovereignty (L8, E8)
The principle that data is subject to the laws and control of where it comes from or whom it describes.
Deep learning (L1)
Machine learning that uses large neural networks with many layers.
Defence in depth (E8)
Several independent defences, so one failure does not become a breach.
Delimiters (E2)
Markers such as <report>…</report> that separate data from instructions in a prompt.
Demographically identifiable information (E8)
Group-level data that can put a community at risk even without names.
Diffusion model (E8)
The kind of model behind AI image generation; it builds an image by removing noise step by step.
Distillation (E8)
Training a small model to copy a large model's answers on a particular task.

E

Embedding (L2, E1, E4)
A list of numbers that represents the meaning of a piece of text, so similar meanings sit close together. Used to search by meaning.
Energy use (of AI) (L8)
The electricity — and often cooling water — that data centres use to train and run models.
Environment variable (E0, E3)
A named value set outside the code; the safe place for secrets such as API keys.
Error analysis (E7)
Reading failures and grouping them into categories before choosing what to measure.
Error compounding (E6)
Small per-step error rates multiplying across a chain: ten steps that are each 95% reliable succeed together only about 60% of the time.
EU AI Act (L8, E8)
The European Union's AI law, which sets rules by risk tier: unacceptable, high, limited and minimal risk, plus rules for general-purpose models.
Eval (E7)
A repeatable test of an AI system on a fixed set of cases, producing a score.
Evaluator-optimizer (E6)
One model call drafts, another critiques against criteria, and the loop repeats until approved or out of rounds.
Example leakage (E2)
A model copying content from a few-shot example into a real answer.
Exfiltration (E8)
Getting data out of a system without permission — for example through a link or image address.
Expert system (L1)
An early form of AI built from many hand-written "if… then…" rules.
Exponential backoff and jitter (E3, E5)
Retrying after increasingly long waits (1 s, 2 s, 4 s…) plus a small random delay, so many clients don't retry at the same moment.

F

Fact-checking routine (L6)
Mark the claims, trace them to sources, compare meaning, look for gaps, record what you checked.
FakeLLM (E0)
A scripted stand-in for a model, so tests run without an API key or network.
Fallback (E5, E7)
A planned alternative — another model, a reduced mode or a human — when the main path fails.
Few-shot prompting (L3, E2)
Including a few examples of the output you want so the AI copies their pattern.
Fine-tuning (L2, E8)
Further training a model on a smaller, targeted set of examples to change its style or narrow skills.
Function calling (tool use) (E5)
The model requests a named function with structured arguments; your code checks it, runs it and returns the result.

G

Generative AI (L1)
AI that creates new content such as text, images, audio or code.
GDPR (General Data Protection Regulation) (L8, E8)
The European Union's data-protection law, a common reference point for handling personal data.
Global Dialogue on AI Governance (L8)
A forum set up by the UN General Assembly for governments and stakeholders to discuss international cooperation on AI (established 2025).
Global Digital Compact (L8, E8)
The UN Member States' 2024 agreement on digital cooperation, including AI governance.
Golden dataset (golden set) (E4, E7)
A versioned set of realistic inputs with known correct answers or expectations, used for evaluation.
Graceful degradation (E7)
Delivering a reduced but clearly labelled result instead of nothing.
GraphRAG (E4)
RAG over a knowledge graph of entities and relationships extracted from a document library.
Grounding / grounded answer (L4, E4)
Making the AI answer only from specific trusted sources; an answer whose every claim is supported by them.
Guardrail (E5)
A check in code on inputs, tool calls or outputs that holds whatever the model does.

H

Hallucination (L2, E1)
A confident, fluent statement from an AI that is false or invented. A hallucinated citation is a reference that looks real but doesn't exist or doesn't say what is claimed.
Handoff / swarm (E6)
A multi-agent design where agents pass control directly to each other, with no central coordinator.
Handoff contract (E6)
A validated, structured format for what one agent passes to the next.
Harness (E5)
Everything wrapped around a model call — instructions, tools, the loop, validation, guardrails, permissions and reliability wrappers — that controls what the model sees and does.
Human approval point (gate) (L7, E6)
A step where a named person must approve before an AI system takes an irreversible or high-stakes action.
Human in the loop (L6, E7)
A person who reviews and approves AI work at agreed points before it has real effects. A rubber stamp is a checkpoint that approves without really checking.
Human sign-off (L5)
The final review and approval by an accountable person before anything is sent or published.
Hybrid search (E4)
Combining keyword (BM25) and meaning-based (vector) scores to rank search results.

I

ISO/IEC 42001 (L8, E8)
An international standard (2023) for an organization's AI management system.
Idempotent (E3, E5)
Safe to repeat: doing it twice has the same effect as doing it once.
Improvement loop (E7)
Run the eval, analyse the errors, change one thing, measure again.
Independent International Scientific Panel on AI (L8)
A panel of 40 experts, established by the UN General Assembly in 2025, that assesses AI's opportunities, risks and impacts.
Inference (L8)
Running a trained model to produce an answer; it uses energy each time.
Ingestion (E4)
Turning raw files into clean text with metadata, ready to be indexed for search.
Iteration (L3)
Improving an AI's answer through specific follow-up messages in the same conversation.

J

Jagged frontier (L1)
The uneven pattern of AI ability: strong at some tasks, weak at similar-looking ones.
Jailbreak (L8, E8)
A user's attempt to trick an AI into breaking its own safety rules.
JSON / JSON Schema (L3, E2)
A plain-text data format of labelled fields that programs can read / a standard description of the shape JSON data must have.
JSON mode vs structured-output (strict) mode (E2)
API settings that guarantee syntactically valid JSON / output that matches a given JSON Schema.

K

Kill switch (E8)
A configuration flag that turns off an AI feature or tool immediately.
Knowledge base / project (L4)
A saved space in an AI tool that holds documents and standing instructions reused across chats.
Knowledge cutoff (L2, L4, E1)
The date after which a model's training data contains no information.

L

Language coverage (L6)
How well an AI tool performs in a given language, which depends on how much of that language it learned from.
LangGraph (E6)
An agent framework in which you define the workflow explicitly as a graph of nodes, edges and shared state.
Large language model (LLM) (L1, E1)
A generative AI model trained on huge amounts of text to predict the next token — the engine behind chat tools such as ChatGPT, Copilot and Claude.
Least privilege (E5, E8)
Each component gets only the permissions its job needs.
Lethal trifecta (L8, E8)
The dangerous combination of access to private data, exposure to untrusted content, and the ability to send information out. Together they allow data theft. (Term coined by Simon Willison, 2025.)
LiteLLM (E0, E3)
An open-source Python library that calls many model providers through one interface.
LLM as a judge (grader) (E7)
A model that grades outputs against a rubric — trusted only after checking it against human grades.

M

Machine learning (ML) (L1)
A branch of AI where systems learn patterns from examples instead of following rules a person wrote.
MCP (Model Context Protocol) (L7, E5)
An open standard for connecting AI assistants to tools and data sources — a "USB port" for AI.
Memory (short-term, long-term, episodic, semantic) (E4)
Context kept within a session; kept across sessions; records of past events; distilled stable facts and preferences.
Messages API (E3)
The standard request format for calling a model: a list of messages, each with a role (system, user, assistant) and content.
Metadata filter (E4)
Limiting a search to documents with given attributes such as language, date, district or permissions.
Model (L1)
The trained system that is then used on new inputs.
Model deprecation (E8)
A provider retiring a model version. Handle it by pinning versions, tracking retirement dates and re-running evals.
Model routing (E6)
Choosing a different model for each step based on its difficulty and cost.
Multi-agent system (L7, E6)
A job split across several specialised agents, coordinated by an orchestrator or supervisor.
Multilingual embedding model (E4)
An embedding model that places text in many languages in one shared space, so a question in one language can find documents in another.
Multimodal model (L2, E1)
A model that handles more than one kind of input or output, such as text, images and audio.

N

Neural network (L1)
A model made of many simple connected units whose adjustable weights are tuned during training.
Next-token prediction (L2, E1)
How a language model writes: give a probability to every possible next token, pick one, add it, repeat.
NIST AI Risk Management Framework (L8, E8)
A voluntary US framework (2023) for identifying and managing AI risks, widely used worldwide.
Non-determinism (E1)
Getting different outputs from the same input.
"Not stated" (L3, E2)
The explicit value used when a source doesn't give a piece of information, instead of guessing.

O

Observability (E7)
Being able to see what a system did and why, through traces, logs and metrics.
OECD AI Principles (L8, E8)
Intergovernmental principles for trustworthy AI, adopted in 2019 and updated in 2024.
Omission (L6)
An important fact, group or caveat from the source that an AI's output left out.
Open-weight model (L2, E1)
A model whose trained weights are published, so organizations can run it themselves.
OpenTelemetry (E7)
The open standard for recording traces and metrics.
Orchestration layer (L7, E6)
The part of an AI system that plans the work, routes it to agents or tools, keeps state and enforces budgets, stop rules and human approval.
Orchestrator-workers (E6)
A central model decides the sub-tasks at run time, hands them to worker models and combines the results.
OWASP Top 10 for LLM Applications (E8)
The standard checklist of security risks for applications built on language models.

P

p95 latency (E7)
The time within which 95% of requests finish.
Parallelization (E6)
Running several model calls at once — on different sub-tasks (sectioning) or on the same task to compare answers (voting).
Parameters (weights) (L1)
The billions of numbers inside a model that are adjusted during training.
pass@k / pass^k (E7)
Succeeds at least once in k tries / succeeds all k times.
Personal data (L8)
Any information that relates to an identified or identifiable person.
PII redaction (E5)
Masking personal data such as names, phone numbers and email addresses.
Pinned dependencies / lock file (E3)
Recording exact package versions so every install is identical.
Placeholder (L5)
A marker in [square brackets] where information is missing, used instead of letting the AI invent it.
Pre-training (L2, E1)
The first and largest training stage, in which a model learns to predict text from an enormous collection of documents.
Prompt (L3)
Everything you give an AI tool: instructions, questions, pasted material and follow-ups.
Prompt caching (E4)
The provider reusing an identical prompt beginning, for lower cost and faster replies.
Prompt chaining (E6)
Fixed steps run in sequence, each using the previous step's output, often with a check between.
Prompt engineering (L3)
The skill of writing instructions that get useful output from an AI.
Prompt injection (L8, E8)
Hiding instructions in content an AI will read so it follows them as if they came from its user. Indirect injection hides them in documents, emails or web pages the system processes.
Prompt template / version / test (L3, E2)
A saved prompt with named slots / an identifier for a prompt's wording, logged with every call / an automated test of how a prompt is built or the logic around the model.
Pydantic (E2)
A Python library that declares data shapes as classes, validates data and generates JSON Schema.

Q

Quantization (E8)
Storing a model's numbers with less precision so it is smaller and faster.
Query translation (E4)
Searching with translations of the question into other languages and merging the results.

R

RAG (Retrieval-Augmented Generation) (L4, E4)
Searching a document collection first, then giving the most relevant passages to the model so it answers from them, with citations — an "open-book exam".
Rate limit (E3)
A provider's cap on requests or tokens per minute; exceeding it returns error 429.
ReAct (E6)
A pattern where the model alternates short reasoning ("thought") with tool use ("action") and reads each result ("observation").
Reasoning model (L2, E1)
A model trained to work through a problem internally before giving its answer.
Recall@k (E4)
The share of questions whose correct document appears in the top k search results.
Reciprocal rank fusion (RRF) (E4)
Merging ranked lists by adding 1 ÷ (60 + rank) from each list.
Red-teaming (E8)
Attacking your own system — or one you are authorised to test — to find weaknesses first.
Regression test (E7)
An eval run automatically to catch things that used to work and no longer do.
Re-identification (L8)
Working out who a person is from data that had names removed.
Repair retry (E2)
Re-asking the model with its invalid reply and the exact validation errors.
Re-ranking (cross-encoder) (E4)
Rescoring the top search results with a slower model that reads question and passage together.
RLHF (reinforcement learning from human feedback) (L2, E1)
Adjusting a model towards answers that people rate as better.
Routing (E6)
A first step that classifies a request and sends it to the right specialised path.
Rubric (L8, E7)
A scoring guide with a concrete description for each level.

S

Sampling, temperature, top-p (L2, E1)
Picking the next token from the model's probabilities; a setting for how focused (low) or varied (high) those choices are; sampling only among the most likely tokens whose probabilities add up to p.
Sandbox (E8)
An isolated environment with no network, credentials or broad file access, for running untrusted code.
Secrets manager (E3, E8)
A secure service that stores keys and passwords and hands them to applications.
Self-check / self-review (L5, L6)
Asking the AI to review its own output against the source. Useful, but not a substitute for human verification.
Self-consistency (voting) (E1)
Asking several times and accepting the majority answer only if enough answers agree.
Semantic check (E2)
Checking the content of output against the source — e.g. that a quoted figure is really in the report.
SitRep (situation report) (L1, E1)
A short, structured update on a crisis for decision-makers: what happened, where, who is affected, urgent needs and open questions. The running case study of both courses.
Skill (E5)
Packaged instructions and resources that an assistant loads only when a task needs them.
Slopsquatting (E3)
Publishing malicious packages under names that AI tools tend to invent.
Small language model (L2, E1)
A model with relatively few parameters — faster and cheaper, sometimes able to run on a laptop or phone.
Smoke test (E0)
The smallest check that a whole path — code, key, network, model — works.
Spec (specification) (L7, E3)
A short document describing what a system must do, must not do, and how you'll know it works. Spec-first development means writing the spec and tests before the code.
State (E6)
Everything a run needs to continue, explain or resume: inputs, plan, step outputs, decisions, counters and errors.
Stateless (API) (E3)
The model remembers nothing between calls; your code resends the conversation each time.
Stop reason (finish_reason) (E3)
Why the model stopped writing: finished, hit the length limit, called a tool, or was filtered.
Stop rule (E6)
The condition that ends a run: success, an exhausted budget, or a human saying stop.
Streaming (E3)
Receiving a model's answer in pieces as it is generated.
Structured output (L3, E2)
An AI answer in a fixed, predictable shape such as a table, form or JSON.
Success criteria (L7)
Measurable statements of what "working" means for a system.
Supervisor / hierarchical (E6)
Multi-agent designs where one coordinating agent calls specialists and merges their results / supervisors manage other supervisors.
System prompt (L3, E2)
Standing instructions given to the AI before the conversation starts, usually set by the tool's owners and hidden from users.

T

Test case (L7, E3)
A realistic input plus a description of the correct result. Tests as the contract means the automated tests define when a change is finished.
Threat model (E8)
A structured review of what a system protects, who might attack it, how, and what you'll do about it.
Timeout (E3)
The longest your code will wait for a reply before treating the call as failed.
Token / tokenizer (L2, E1)
A chunk of text — word, part of a word or punctuation — that a model reads and writes / the component that splits text into tokens. Limits and costs are counted in tokens, and some languages need more tokens for the same meaning.
Tool (L7, E5)
A specific action an AI system is allowed to request, such as searching a folder or drafting an email. Tool poisoning hides malicious instructions in a tool's description or output.
Tool-call loop (E5)
Ask the model, run the tool it requests, feed back the result or error, and repeat until a final answer or the step budget runs out.
Trace / span (E7)
The record of one request from start to finish / one timed step within it.
Traffic-light rule (L6)
Green (go, with a normal check), amber (care and a named checker), red (don't use a general AI tool).
Training (L1)
Adjusting a model using many examples until it performs a task well.
Trajectory eval (E7)
Checking the path of tool calls an agent took, not only its final answer.
Transformer (L2, E1)
The model design (2017) behind almost all modern large language models, built around attention.
Transient vs permanent error (E3)
A failure worth retrying after a wait vs one that will fail the same way again.
Trust boundary (E8)
The point where data crosses from a less trusted to a more trusted part of a system.

U – Z

UNESCO Recommendation on the Ethics of AI (L8, E8)
A global standard on AI ethics adopted by UNESCO's member states in 2021.
Terminology list (L5)
Agreed translations of key terms that the AI must use, checked against your organization's list or an official terminology database.
Validation (E2)
Checking output against a schema before code uses it.
Vector store / vector database (E4)
A store that holds chunks with their embeddings and metadata and quickly finds the most similar. Fast search uses an approximate nearest neighbour (ANN) index.
Vibe coding (E3)
Building software by accepting AI-generated code because it seems to work, without close review.
Virtual environment (venv) (E0)
A private folder of Python packages for one project.
Workflow (L7, E6)
Several model calls connected along a path your code defines in advance — as opposed to an agent, which chooses its own path.
Zero-shot prompting (L3, E2)
Asking the AI to do a task with instructions only, no examples.