AI Glossary

A hypothetical AI that matches or exceeds human ability across essentially all cognitive tasks, rather than being good at just a few. It does not exist yet.

An AI system that can take a goal and pursue it over multiple steps, deciding what to do next and using tools, rather than answering a single question.

An agent asked to book a trip might search flights, compare prices and fill a form on its own.

A tool that claims to judge whether text or an image was made by AI. They are unreliable, producing both false accusations and easy misses.

Google's AI-generated summary that appears at the top of many search results, answering your query directly above the usual list of links.

The subtle giveaways that text or media may be AI-made — overly balanced phrasing, hedging, generic vividness, or too-smooth symmetry.

A suspicious fondness for em-dashes and phrases like 'it's important to note'.

The effort to make AI systems pursue what people actually want and value, and behave safely, rather than optimising for the wrong thing.

An interface that lets one program talk to another. AI providers offer APIs so developers can send prompts to a model from their own apps.

Tell-tale glitches in AI-generated images — mangled hands, garbled text, mismatched earrings or impossible reflections — that reveal the fake.

Six-fingered hands and warped background text are classic image artifacts.

The broad field of building software that performs tasks we normally associate with human intelligence, such as understanding language, recognising images, or making decisions.

A spam filter, a chess engine, and ChatGPT are all forms of AI, despite working very differently.

The mechanism that lets a transformer decide which earlier words matter most when interpreting or predicting the next one, capturing long-range context.

In 'the trophy did not fit in the case because it was too big', attention links 'it' to 'trophy'.

The algorithm that works out how much each parameter contributed to a model's error, so gradient descent knows which way to adjust it.

Processing many requests together, often more cheaply and slowly than one-at-a-time, for jobs that do not need an instant answer.

A generation strategy that keeps several candidate sequences alive at once and expands the most promising, aiming for a higher-quality overall output.

A standard test used to measure and compare how well models perform on a task, such as maths problems, coding, or general knowledge.

Labs often quote scores on benchmarks like MMLU or GPQA when announcing a new model.

Systematic unfairness in a model's outputs, usually inherited from skewed training data, that can disadvantage particular groups.

A hiring model that rates identical CVs differently based on the name at the top.

Prompting a model to reason step by step before giving its final answer, which often improves accuracy on maths and logic problems.

Adding 'Let's think step by step' can nudge a model to show its working.

A program you converse with in natural language. Modern chatbots are usually a friendly interface wrapped around a large language model.

Misleading media made with simple tricks like selective editing, mislabelling or slowing a clip down — no AI required, but just as deceptive.

A model whose job is to sort inputs into categories, such as 'spam or not', 'positive or negative', or which animal is in a photo.

Tamper-evident metadata attached to a file that records how it was made and whether AI was involved, helping people judge if media is trustworthy.

The C2PA standard adds a signed 'nutrition label' describing an image's origin.

Storing a reused chunk of prompt (like a long document or system prompt) so repeat requests are cheaper and faster than re-sending it every time.

The maximum amount of text, measured in tokens, a model can consider at once — including your prompt, any documents, the chat so far and its reply.

A 200,000-token context window fits a few long books at a time.

An AI assistant embedded inside a tool to help with a specific job — writing code, drafting documents, or filling spreadsheets — while you stay in control.

GitHub Copilot suggests lines of code as you type in your editor.

The question of what happens to the text, files and personal information you send to an AI service — whether it is stored, logged, or used for training.

Machine learning built on neural networks with many layers. The depth lets these systems learn very rich patterns, and it powers almost all modern AI.

Synthetic media — usually video or audio — that convincingly shows a real person saying or doing something they never did.

The technique behind most AI image generators. It learns to turn random noise into an image step by step, reversing a process that gradually added noise.

Type a prompt and a diffusion model 'denoises' it into a picture.

A model that learns to tell classes apart or predict a label from an input, rather than to generate new data. Classifiers are discriminative.

Training a smaller 'student' model to imitate a larger 'teacher', producing a lighter model that keeps much of the bigger one's ability.

A list of numbers that represents a piece of text (or image) as a point in space, so that things with similar meaning sit close together.

'Dog' and 'puppy' land near each other; 'dog' and 'bicycle' land far apart.

A specific web address you send requests to when using an API, such as the URL that accepts your prompt and returns a model's reply.

How well we can understand and articulate why a model produced a particular output. Big neural networks are notoriously hard to explain.

Including a handful of examples in your prompt to show the model the pattern you want before asking it to continue.

Give two sample question-and-answer pairs, then pose your real question.

To adapt an existing model to your own data or style, rather than building one from scratch — a common way teams customise AI cheaply.

Further training a pretrained model on a smaller, focused dataset so it specialises — for a tone of voice, a domain, or a particular task.

A hospital might fine-tune a general model on medical notes to make it better at clinical language.

A large model trained broadly on general data that can then be adapted to many downstream tasks, rather than being built for a single job.

One of the most capable, cutting-edge models available at a given moment, typically from the labs pushing the state of the art.

A structured way for a model to request that your code run a specific function with specific arguments, so it can act rather than just describe.

The model returns 'get_weather(city: London)' and your app runs it.

An older image-generation approach where two networks compete: one creates fakes and the other tries to spot them, pushing the faker to improve.

AI that creates new content — text, images, audio, video or code — rather than just classifying or predicting existing data.

ChatGPT writing an email or Midjourney producing an image are generative AI.

Short for Generative Pre-trained Transformer. It names OpenAI's family of language models and, more loosely, the transformer approach they popularised.

The optimisation method that trains neural networks: it nudges each parameter a little in the direction that reduces error, repeated millions of times.

Anchoring a model's answer to specific, trusted source material you supply, so it responds from real documents rather than memory alone.

The rules, filters and checks placed around a model to keep it from producing harmful, unsafe or off-policy outputs.

When a model states something false, made-up, or unsupported with total confidence. It is a fundamental limitation, not an occasional glitch.

A model inventing a fake but plausible-looking citation or court case.

A model's ability to pick up a new task from examples and instructions in the prompt alone, without any change to its trained parameters.

Actually running a trained model to get an answer. Every time you send a prompt and get a reply, that is one inference.

A prompt crafted to trick a model into bypassing its safety rules and doing something it is supposed to refuse.

The date after which a model was not trained on new information. Without extra tools, it will not know about events past that point.

A model with an early-2025 cutoff will not know who won a match played in 2026.

A very large neural network trained on huge amounts of text to predict the next token, which lets it write, summarise, translate and answer questions.

GPT, Claude and Gemini are all large language models.

The delay between sending a request and getting a response. Lower latency means the model feels snappier to use.

The problem that once deepfakes exist, people can dismiss genuine evidence as 'probably fake', so real footage loses its power to convince.

The raw, unnormalised score a model assigns to each possible next token before those scores are turned into probabilities.

A branch of AI where a system learns patterns from examples rather than being programmed with explicit rules. You show it data, and it figures out the rules itself.

Instead of coding what a cat looks like, you show a model thousands of cat photos until it learns.

An open standard for connecting AI assistants to external tools and data sources in a consistent way, so the same connector works across apps.

In AI, the trained system itself — the network plus all the numbers it learned. When you 'use GPT', you are sending input to a model and reading its output.

The degradation that can happen when models are trained too heavily on other models' output, gradually losing quality and diversity.

Able to work with more than one type of data at once — for example understanding text and images together, or generating audio from a written prompt.

You can show a multimodal model a photo of your fridge and ask what you can cook.

A model made of layers of simple units, loosely inspired by neurons, that pass numbers between them. Adjusting the connections during training is how it learns.

A single unit in a neural network that takes numbers in, combines them, and passes a number out. Networks stack millions of these into layers.

A model whose trained parameters are freely published, so anyone can download, run and adapt it on their own hardware.

Meta's Llama and Mistral models are distributed as open weights.

When a model memorises its training data instead of learning general patterns, so it performs well on examples it has seen but poorly on new ones.

One of the adjustable numbers inside a neural network. Large models have billions of them, and their learned values encode everything the model 'knows'.

A measure of how surprised a language model is by a piece of text. Lower perplexity means the text was more predictable to the model.

The first, biggest training stage, where a model learns general patterns of language from vast amounts of text before any task-specific tuning.

The input you give a model — your question, instruction, or the text you want it to work with. The reply is shaped heavily by how you phrase it.

Breaking a hard task into a sequence of prompts, where the output of one becomes the input to the next, for more reliable multi-step results.

The craft of writing prompts that reliably get good results: being specific, giving context, showing examples, and telling the model what format you want.

An attack where malicious instructions are hidden in content a model reads — a web page or document — to hijack its behaviour.

A web page containing hidden text that tells an AI browsing it to leak the user's data.

A reusable prompt with blanks you fill in, so you can apply the same reliable structure to many different inputs.

'Summarise the following in three bullet points: {text}'.

Verifying where a piece of media came from — its original source, date and any content-credentials metadata — instead of trusting it at face value.

Shrinking a model by storing its numbers at lower precision, so it uses less memory and runs faster, usually with only a small quality trade-off.

A technique where the system first fetches relevant documents, then feeds them to the model so it answers from that material instead of memory alone.

A support bot that looks up your help articles before replying uses RAG.

A cap on how many requests or tokens you can send to an API in a given time, used by providers to keep the service stable and fair.

Deliberately probing an AI system for weaknesses and harmful behaviours before release, by trying to make it fail on purpose.

Training where a system learns by trial and error, receiving rewards for good outcomes and penalties for bad ones, and adjusting its behaviour to earn more reward.

A second pass that reorders search results so the most relevant ones rise to the top, improving what gets fed to a model in RAG.

Uploading a picture to a search engine to find where else it appears online, a quick way to check if a 'photo' is old, staged or fabricated.

A tuning method where humans rank model responses and the model is trained to prefer the kinds people rate highly, making it more helpful and polite.

The step where a model actually chooses the next token from its predicted probabilities, rather than always taking the single most likely one.

Searching by meaning rather than exact keywords, using embeddings so a query for 'cheap laptops' can match 'budget notebooks'.

A piece of text that tells the model to stop generating when it produces it, used to keep responses tidy or end them at a natural point.

Delivering a model's response token by token as it is generated, so words appear gradually instead of after a long wait.

The way ChatGPT types out its answer live is streaming.

A hypothetical AI far surpassing the best human minds in virtually every field. It is a subject of serious debate about long-term safety, not a current product.

Training on data where each example comes with the correct answer, so the model learns to map inputs to known labels.

Teaching a model to flag spam using emails already labelled 'spam' or 'not spam'.

Any content generated or heavily altered by AI, including images, video, audio and text, as opposed to something captured or written by a person.

A behind-the-scenes instruction that sets a model's role, tone and rules for a whole conversation, separate from what you type in each message.

'You are a concise assistant that always answers in British English.'

A setting that controls randomness in a model's output. Low temperature makes it focused and predictable; high temperature makes it more varied and creative.

How much work a model can handle over time, such as tokens generated or requests served per second. It matters when serving many users at once.

The basic unit a language model reads and writes — usually a short chunk of text like a common word, part of a word, or a punctuation mark. Models see tokens, not letters.

'Tokenisation' might split into 'token' + 'isation', two tokens.

A cap on how many tokens can be involved in a request or response. Exceeding it means text gets truncated or rejected.

The process of chopping text into tokens before a model reads it. The same text can produce different token counts on different models.

Giving a model access to external functions — search, a calculator, a calendar — that it can call to get information or take actions beyond text.

Another randomness control. The model only picks from the smallest set of next-token options whose probabilities add up to p, trimming off unlikely choices.

The process of feeding a model data and gradually adjusting its parameters so its predictions get better. It is expensive and done once, up front.

Reusing a model trained on one task as a starting point for another, so you need far less new data and compute than training from scratch.

The neural-network architecture behind almost every modern language model. Its key trick, attention, lets it weigh how much each word relates to every other.

Training on data with no labels, where the model finds structure or patterns on its own, such as grouping similar items together.

Simply a list of numbers. In AI, embeddings are vectors, and comparing how close two vectors are is how systems measure similarity of meaning.

A database built to store embeddings and quickly find the ones most similar to a query, which is the retrieval half of most RAG systems.

Embedding a hidden, detectable signal in AI-generated content so it can later be identified as machine-made, even after light editing.

Another word for a model's learned parameters — the numbers on the connections between units. 'Open weights' means these numbers are published for anyone to use.

Asking a model to do a task with no examples in the prompt — just the instruction. Modern models handle many tasks zero-shot.