AI Glossary - The AI Compass

A

AGI (Artificial General Intelligence)

Hypothetical AI that could handle any intellectual task a human can. It does not exist today; current systems are "narrow".

AI Agent

An AI that does not just chat but can use tools, browse, and carry out multi-step tasks toward a goal. Because it can act, it needs clear limits and human oversight.

AI Literacy

A basic understanding of how AI works, what it can and cannot do, and its risks. The EU AI Act now requires it for staff who deploy AI.

Algorithmic Bias

When an AI produces unfair results because its training data reflected human prejudice or gaps - "garbage in, garbage out".

Alignment Problem

The challenge of making an AI's goals and behaviour match human values, a problem that grows harder as systems become more capable.

B

Bias Mitigation

Techniques used during data selection and training to reduce unfair or stereotyped outputs.

Black Box

A model whose internal decision-making is hard to explain, even for its creators. Common in deep learning.

C

Chain-of-Thought

Prompting or training an AI to reason step-by-step, which improves accuracy on maths and logic problems.

Computer Vision

The field of AI that interprets images and video, recognising objects, faces, and text.

Content Provenance (C2PA)

Tamper-evident "Content Credentials" attached to media that record how it was created or edited, helping flag AI-generated content.

Context Engineering

Supplying a model the right documents, examples, memory, and tools for a task - increasingly more important than a single clever prompt.

Context Window

How much text an AI can "see" at once. Anything beyond it is forgotten, so long chats lose their earliest parts.

Copyright (AI-generated work)

In the US, purely AI-generated work is not copyrightable; protection depends on meaningful human authorship.

D

Data Poisoning

An attack that feeds bad examples into training data to sabotage or manipulate a model's behaviour.

Data Privacy

Protecting personal or sensitive information. Public AI tools may log or train on your inputs, so avoid pasting secrets.

Deepfake

AI-generated media that mimics a real person's face or voice, creating realistic but fake video or audio.

E

Embedding

A representation of text (or other data) as a list of numbers that captures meaning, letting computers compare how similar things are.

EU AI Act

The world's first comprehensive AI law. It ranks AI by risk, bans some uses, and is phasing in (2025-2027); from August 2026 it adds transparency rules such as labeling AI-generated content.

F

F.A.T.E. Framework

A practical ethics checklist: Fairness, Accountability, Transparency, and Ethics.

Federated Learning

Training that happens on local devices so raw personal data stays put; only the learned updates are shared.

Fine-tuning

Taking a general pre-trained model and training it further on specialised data, such as medical or legal text.

Foundation Model

A large, broadly capable model (like GPT, Claude, or Gemini) that can be adapted to many downstream tasks.

G

Generative AI

AI that creates new content - text, images, audio, or code - rather than only analysing existing data.

GPU

A graphics processor good at doing many calculations in parallel, which is exactly what training and running AI models requires.

H

Hallucination

When an AI states false information confidently. Because models predict likely text, they do not truly "know" facts.

Human-in-the-Loop

Keeping a person in the decision chain to review or approve an AI's output - vital for high-stakes uses.

J

Jailbreaking

Using tricks or role-play prompts to bypass an AI's safety filters. Attempting it can violate a service's terms.

L

Large Language Model (LLM)

A model trained on huge amounts of text that predicts the next likely word; the engine behind chatbots.

M

Machine Learning

Teaching computers from examples rather than from hand-written rules.

Model Collapse

Quality and diversity degrading when models are repeatedly trained on AI-generated content.

Model Context Protocol (MCP)

An open standard, introduced by Anthropic, that gives AI agents a consistent, auditable way to connect to external tools and data.

Multimodal AI

Models that handle several input and output types - text, images, audio, sometimes video - not just text.

N

Narrow AI

AI that is good at one specific task (spam filtering, chess) and nothing else. Almost all AI today is narrow.

Natural Language Processing (NLP)

The field of AI focused on understanding and generating human language.

Neural Network

A mathematical model loosely inspired by the brain, built from layered connections whose strengths are learned.

O

OCR (Optical Character Recognition)

AI that turns text inside an image or scan into editable, searchable text.

Overfitting

When a model memorises its training data so closely that it fails on new, real-world inputs.

P

Parameter

An internal value learned during training - the "connection strengths" inside a model. Frontier models have billions to trillions.

Predictive Analytics

Using patterns in past data to forecast likely future outcomes, such as customer churn.

Prompt Injection

A security attack where malicious text tricks an AI into ignoring its instructions or safety rules.

R

RAG (Retrieval-Augmented Generation)

Connecting an AI to external data, like your own files, so it can answer from accurate, up-to-date sources.

Reasoning Model

A model that spends extra computation working through steps before answering, improving hard logic and maths at the cost of speed.

Red Teaming

Experts deliberately attacking a model to find vulnerabilities and safety flaws before release.

RLHF

Reinforcement Learning from Human Feedback - a training method where humans rank responses to teach a model to be more helpful and safe.

S

Sentiment Analysis

Detecting emotional tone (positive or negative) in text, used to scan reviews or social media.

Shadow AI

Unauthorised use of AI tools by employees, which can expose company data and create compliance risks.

Singularity

A hypothetical point where AI surpasses human control and change becomes uncontrollable. Speculative, not current reality.

Supervised Learning

Training with labelled examples - inputs paired with the correct answers.

Sycophancy

A tendency for chatbots to agree with the user even when the user is wrong, sometimes a side effect of feedback-based training.

Synthetic Data

Realistic but artificial data generated to train models when real data is scarce or private.

T

Temperature

A setting that controls randomness: high is more creative and varied, low is more focused and deterministic.

Token

A chunk of text (often part of a word) that a model reads or writes. Roughly 1,000 tokens is about 750 words.

Tool Poisoning

An attack where a malicious tool or web page feeds hidden instructions to an AI agent to hijack its behaviour.

Turing Test

A historical idea: can a machine's responses be indistinguishable from a human's? Modern AI is judged with more detailed benchmarks.

U

Uncanny Valley

The discomfort people feel when something looks almost - but not quite - human.

Unsupervised Learning

Finding patterns in data that has no labels or answer key.

V

Vector Database

A database that stores meaning as numbers (embeddings) so you can search by concept rather than exact keywords.

W

Watermarking

Hidden or embedded markers that help identify content as AI-generated.

Z

Zero-Shot Learning

A model performing a task it was not explicitly trained on, just by following instructions.