AI talk sounds like this. Select any highlighted term to see what it means.

Agent2Agent. A shared standard that lets AI agents from different companies talk to each other and pass work along.

Being clear about who is responsible when an AI system gets something wrong: the maker, the company using it, or the person who relied on it.

Feeding an AI a carefully crafted input designed to fool it, such as a sticker that makes a self-driving car misread a road sign.Think of it as: An optical illusion, but for machines.

An AI that doesn't just answer but takes steps for you, such as searching, clicking, booking or writing files, to finish a goal you set.Think of it as: A chatbot tells you how to book a table. An agent books it.

The software wrapped around a model that lets it act as an agent. It manages the to-do list, the tools, the memory and the permissions.Think of it as: The model is the engine. The harness is the rest of the car.

Ready-made instruction packs an agent can load when it needs to do a particular job, such as filling in a spreadsheet or making slides.

Artificial general intelligence. A still-hypothetical AI that could do any thinking task as well as a person. People disagree on how close it is, or what would count.

A chatbot built to help with everyday tasks: answering questions, writing, planning, summarising. ChatGPT, Claude, Gemini and Copilot are all AI assistants.

The worry that far more money is being poured into AI than it can earn back, and that valuations could fall sharply, as with the dot-com crash.

A chatbot designed to act as a friend or partner. Popular, and also a concern, because people can become emotionally dependent on it.

A tool that claims to spot AI-written text or AI-made images. They are unreliable and often wrong, so treat results with caution.

The study of how to build and use AI in ways that are fair, honest and respectful of people.

The rules, checks and sign-offs an organisation uses to make sure its AI is used safely and legally.

Knowing enough about AI to use it well and question it sensibly. It's what this site is for.

A marketing term for a laptop or desktop with a chip built to run AI features on the device itself.

Research and engineering aimed at making sure AI systems don't cause harm, by accident or by misuse.

Search that gives you a written answer with sources, instead of a list of links.

Low-effort AI-made content, such as articles, images and videos, churned out in bulk to grab clicks.

A period when excitement and funding for AI dry up after over-promising. It has happened twice before, in the 1970s and late 1980s.

A product that is mostly a nice interface on top of someone else's AI model.

A set of step-by-step instructions a computer follows to get something done.Think of it as: A recipe. Same steps, same cake.

The work of making sure an AI actually does what people intend and behaves in line with human values.

The AI safety company that makes Claude.

Our habit of treating AI as if it were a person with feelings and intentions, because it talks like one.

A way for one piece of software to talk to another. It's how apps plug an AI model into their own product.Think of it as: A waiter carrying orders between you and the kitchen.

Apple's name for the AI features built into iPhones, iPads and Macs.

Computer systems that do things we'd normally say need human thinking: understanding language, spotting patterns, making decisions.

The trick inside modern AI that lets it weigh which words in a sentence matter most to each other, so it keeps track of meaning.

Getting software to do a repetitive task without a person doing each step.

A car or lorry that drives itself using AI, cameras and sensors.

A digital character, often lifelike, that can speak words you type. Used in training videos and marketing.

An agent that keeps working on a job after you've closed the chat, and reports back when it's done.

The way a neural network learns from its mistakes: it works out how wrong it was, then nudges its internal dials to be less wrong next time.

A standard test used to compare AI models. Treat scores like exam grades: useful, but not the whole picture.

An early version released for people to try while it's still being finished. Expect rough edges.

When an AI gives unfair or skewed results because the data it learned from was unfair or skewed.Think of it as: If you only ever read one newspaper, you'd pick up its slant too.

A system where you can see what goes in and what comes out, but not clearly why it decided what it did.

When an AI works through a problem step by step before giving its answer, which usually makes it more accurate.Think of it as: Showing your working in a maths exam.

The saved record of your past conversations with an assistant. You can usually delete it or turn it off.

A program you talk to in everyday language by typing or speaking, and it replies.

OpenAI's chatbot, and the product that brought this kind of AI to the general public in late 2022.

Sorting things into categories, such as spam or not spam. One of the most common jobs AI does.

The AI assistant made by the company Anthropic.

A model you can only use through the company's own app or service. You can't download it or look inside.

Other people's computers, rented over the internet. Most AI you use runs in the cloud, not on your device.

An AI that can read a software project, make changes, run tests and fix errors, working through a task step by step.

Shorthand for computing power: the chips, electricity and time needed to build and run AI.

An AI operating a computer the way you do: looking at the screen, moving the mouse, clicking and typing.

AI that understands images and video, such as recognising faces, reading number plates or spotting problems on a scan.

A ready-made link between an AI assistant and another app, such as your email, calendar or files.

A training approach where a model is taught to follow a written set of principles, its constitution.

A digital label attached to a photo or video recording where it came from and whether AI was involved.Think of it as: A nutrition label for media.

Checking and filtering content for things that break the rules. Increasingly done by AI, with people handling the hard cases.

Summarising the older parts of a long conversation so an AI can keep going without running out of room.

Carefully choosing what information, files and instructions to give an AI so it has what it needs and nothing that distracts it.

When a conversation gets so long and cluttered that the AI's answers get worse. Starting a fresh chat usually fixes it.

How much the AI can keep in mind at once, including your conversation and any documents you share. Go past it and it starts forgetting the earliest parts.Think of it as: The size of the AI's desk. A bigger desk holds more papers.

Microsoft's name for its AI assistants, built into Windows, Office and coding tools. Also used loosely for any AI that works alongside you.

The legal right creators have over their work. Whether AI companies may train on copyrighted material without permission is being fought over in courts worldwide.

Standing preferences you give an AI once, such as your job or preferred tone, so you don't repeat yourself every chat.

A warehouse-sized building full of computers. AI models are built and run in these, which is why they use so much electricity and water.

People tagging examples, such as marking which photos contain a cat, so an AI can learn from them.

Your right to control what happens to your personal information, including whether your chats are used to train AI.

A collection of examples, such as text, images or numbers, gathered to teach or test an AI.

A powerful kind of machine learning that uses many-layered neural networks. It's behind most of today's impressive AI.

A mode where an assistant spends several minutes searching and reading many sources, then writes you a report with references.

A fake video, image or voice clip made with AI to look or sound like a real person.

A Chinese AI lab known for capable models that are cheap to run and free to download.

When a company retires an older model or feature. Apps built on it have to move to something newer.

The technique behind most AI image makers. It learns to turn random speckle into a picture that matches your description, one small step at a time.Think of it as: A sculptor chipping away until the statue appears.

False information spread on purpose to deceive. Misinformation is false information spread by mistake. AI makes both easier to produce.

Training a small, cheap model to copy a big, expensive one so you get most of the skill at a fraction of the cost.

A way of turning words or images into lists of numbers so a computer can tell which things have similar meanings.Think of it as: Map coordinates for ideas. Similar ideas sit close together.

Abilities that show up in large models without anyone specifically training for them.

AI needs a lot of electricity, mostly in data centres. How much, and where it comes from, is a growing public debate.

AI tools sold to large organisations, with extra security, controls and support.

The European Union's law on AI. It bans some uses outright and sets stricter rules the riskier the use.

Short for evaluations. Tests that check how well, and how safely, an AI performs.

The concern that very advanced AI could one day threaten humanity as a whole. Experts disagree strongly on how seriously to take it.

How well we can understand and describe why an AI made a particular decision.

AI that identifies people from their faces. Useful for unlocking your phone, controversial when used for surveillance.

Making sure an AI treats different groups of people equally and doesn't disadvantage anyone unjustly.

Giving the AI a few examples of what you want before asking for your own.

Taking a ready-made AI model and giving it extra training on specific material so it gets good at one particular job.Think of it as: A qualified doctor who goes on to specialise in hearts.

A big general-purpose AI model that lots of different products are built on top of.

The most capable AI models that exist at any moment, the cutting edge.

Generative adversarial network. An older way of making fake images, where one network creates and another tries to catch it out, so both improve.Think of it as: A forger and a detective training each other.

Google's family of AI models, and the name of its assistant.

AI that creates new things, such as text, images, music, video or code, rather than only sorting or analysing what already exists.

Microsoft's AI coding helper, which suggests and writes code inside a programmer's editor.

Google's AI research lab, maker of Gemini and of AlphaFold, which predicts the shapes of proteins.

Generative pre-trained transformer. The type of model behind ChatGPT: it generates text, was trained in advance on lots of writing, and uses the transformer design.

Graphics processing unit. A chip first made for video games that turned out to be ideal for AI. Demand for them is why Nvidia became so valuable.

The method a model uses to improve during training: take a small step in whichever direction reduces its errors, and repeat millions of times.Think of it as: Walking downhill in fog by always stepping where the ground slopes down.

A version of RAG that also maps how facts relate to each other, so the AI can answer big-picture questions, not just look up passages.

The chatbot made by Elon Musk's AI company, xAI.

The known correct answer that an AI's output is checked against.Think of it as: The answer key at the back of the textbook.

Tying an AI's answer to real sources, such as documents or web results, so it isn't working from memory alone.

Rules and filters built around an AI to stop it producing harmful or inappropriate responses.

When an AI confidently states something that is wrong or made up. It isn't lying. It's predicting words that sound right.Think of it as: A student who didn't revise, bluffing through an essay.

When one AI agent passes a task, and everything it knows about it, to another agent better suited to continue.

The website where the AI community shares free models and datasets.Think of it as: The public library of AI models.

Keeping a person involved to check or approve what the AI does, especially for decisions that matter.

A robot with a human-like body, designed to work in spaces built for people.

One of the giant cloud companies, mainly Amazon, Microsoft and Google, that own the data centres most AI runs on.

A tool that makes pictures from a written description. Midjourney and DALL-E are well-known examples.

The moment an AI model is actually used: you ask, it works out an answer. Training is learning, inference is doing.

Using AI to remove or replace part of an image, filling the gap so it looks natural.

Extra training that teaches a model to follow requests, turning a text predictor into a helpful assistant.

A cleverly worded request designed to trick an AI into ignoring its safety rules.

A collection of an organisation's documents that an AI can look things up in.

The date an AI's training information stops. It won't know about anything after that unless it can search the web.Think of it as: A newspaper archive that ends on a certain day.

A map of facts showing how things are connected, such as people, places and companies and the relationships between them.

An AI trained on enormous amounts of text so it can read and write like a person. It works by predicting which words should come next.Think of it as: Predictive text on your phone, scaled up millions of times.

The wait between asking an AI something and getting the reply.

A public ranking of AI models by test scores. Handy, but companies are known to tune their models for the tests.

Meta's family of AI models, notable because they can be downloaded and run by anyone.

The behind-the-scenes work of running AI apps reliably: monitoring them, testing them and keeping costs under control.

An AI model that runs on your own phone or computer rather than on a company's servers. Usually more private, often less powerful.

Low-rank adaptation. A cheap way to customise a big model by training a small add-on instead of changing the whole thing.Think of it as: A clip-on lens instead of a new camera.

Tools that let you build software mostly by dragging and configuring, with only a little coding.

Teaching computers by showing them lots of examples rather than writing out every rule by hand.Think of it as: A child learns what a dog is by seeing many dogs, not by reading a definition.

Model Context Protocol. A shared standard that lets AI assistants connect to other apps and data in a consistent way.Think of it as: A universal plug socket for AI tools.

A small program that makes one app or data source available to AI assistants through the MCP standard.

A feature that lets an assistant remember things about you between conversations. You can usually view it or turn it off.

Facebook's parent company's AI division, maker of the Llama models and the assistant inside WhatsApp and Instagram.

A popular AI image generator known for its artistic style.

A French AI company known for efficient models, many of them free to download.

A model design with many specialist parts, where only the relevant few switch on for each question. It saves computing power.

A lasting advantage that stops competitors copying a business. A constant question in AI, where many products rely on the same models.

The trained brain of an AI system. It's the result of all the learning and the thing that produces answers.

A short fact sheet published with a model that explains what it's for, how it was tested and where it falls short.

What can happen when AI is trained on too much AI-made content: quality and variety gradually decline.Think of it as: A photocopy of a photocopy of a photocopy.

A public document in which an AI company sets out how its assistant is supposed to behave.

Several AI agents working together, each with its own job, such as one researching, one writing and one checking.

An AI that handles more than one kind of input or output, such as text, images, audio and video, not just words.

AI built for one job, such as filtering spam or recommending films. Every AI in use today is narrow to some degree.

The field of getting computers to understand and produce human language.

Software loosely inspired by the brain: layers of simple connected units that strengthen or weaken their links as they learn.

Tools that let you build apps and automations without writing any code.

Neural processing unit. A chip in newer phones and laptops designed to run AI efficiently without draining the battery.

The company that makes most of the chips AI is trained and run on, which has made it one of the most valuable companies in the world.

Optical character recognition. Turning a photo or scan of text into text you can edit and search.

Giving the AI a single example of what you want before asking for your own.

A model whose trained brain is published so anyone can download and run it. Often loosely called open source, though the recipe and data aren't always shared.

The company that makes ChatGPT and the GPT models.

Software that coordinates several AI agents, tools and steps so they work together on one job, deciding who does what and in what order.Think of it as: The conductor of an orchestra. The musicians play, the conductor keeps them together.

When a model memorises its practice examples so closely that it does badly on anything new.Think of it as: Memorising last year's exam answers instead of learning the subject.

Trusting AI output without checking it, or letting your own skills fade because the AI always does it for you.

The millions or billions of internal dials a model adjusts as it learns. More can mean more capable, but not always.

An AI search company whose answers come with links to their sources.

Anything that identifies you, such as your name, address, health details or photos. Think before sharing it with an AI tool.

An AI tailoring its answers to you based on what it knows about your preferences and history.

Scam messages that pretend to be from someone you trust. AI lets criminals write more convincing ones, in perfect English, at scale.

AI that works in the real world through robots, vehicles and machines, not just on screens.

A small trial of an AI tool inside an organisation before deciding whether to roll it out to everyone.

A test area where you can try a model with different settings before building anything with it.

Everything done to a model after its main training to make it helpful and safe, including instruction tuning and RLHF.

The first and most expensive stage of building a model, where it reads enormous amounts of text to learn how language and the world work.Think of it as: General schooling before job training.

Whatever you type or say to an AI to get a response: your question or instruction.

Breaking a big job into steps and feeding each answer into the next prompt.

The skill of wording your request so you get better results from an AI.

Hidden instructions tucked inside a web page, email or document that try to hijack an AI reading it.

A saved collection of prompts that work well, so you and your team can reuse them.

Shrinking a model by storing its numbers less precisely, so it runs on smaller devices with a small loss in quality.Think of it as: Saving a photo as a smaller file.

Retrieval-augmented generation. The AI looks things up in trusted documents first, then writes its answer from what it found.Think of it as: An open-book exam instead of answering from memory.

A cap on how many requests you can make in a given time. It's why a chatbot sometimes tells you to come back later.

A model built to think through a problem step by step before it answers. Slower, but stronger on maths, logic and planning.

AI that suggests what you might like next, such as films on Netflix or products on Amazon. Probably the AI you use most without noticing.

Deliberately trying to make an AI misbehave so weaknesses are found and fixed before release.Think of it as: A bank hiring someone to try to break into its own vault.

Laws that govern how AI can be built and used. Countries are taking very different approaches.

Training by trial and error, where the AI is rewarded for good results.Think of it as: Teaching a dog tricks with treats.

An umbrella term for building AI that is fair, safe, transparent and accountable.

Reinforcement learning from human feedback. People rate the AI's answers and it learns to give the kind people prefer.

Building machines that can sense and move in the physical world. AI is increasingly what controls them.

Telling the AI who to be, such as a patient tutor or a strict editor, to shape the style of its answers.

A sealed-off area where an AI can run code or try things without being able to damage anything outside it.Think of it as: A playpen.

The observation that AI tends to improve in a predictable way as you give it more data, more computing power and a bigger model.

Software development kit. A toolbox of ready-made code that helps programmers build on a particular platform.

Learning from raw data without human labels, for example by hiding a word in a sentence and trying to guess it. This is how language models learn.

Search that finds results by meaning, not exact words, so a search for car also finds vehicle.

AI that judges whether a piece of text is positive, negative or neutral. Used to sift reviews and social media.

Staff using AI tools at work that their employer hasn't approved, often pasting in company data.

A compact language model that is cheaper and faster, and can run on a phone or laptop.

OpenAI's tool for generating video from a written description.

A country building its own AI models and data centres so it doesn't depend on foreign tech companies.

Turning spoken words into written text. It's how dictation and auto-captions work.

A well-known image generator that is free to download and run yourself.

Getting an AI to answer in a strict format, such as a table or a form, so other software can read it reliably.

A helper agent that a main agent hands part of a job to. It works on its piece and reports back.

Condensing a long document, meeting or thread into the key points. One of the most reliable everyday uses of AI.

A hypothetical AI far smarter than humans at nearly everything. It doesn't exist.

Training with examples that come with the right answers attached, such as photos already labelled cat or dog.

An AI's habit of telling you what you want to hear instead of what's true.

Artificially generated data used for training when real data is scarce, expensive or private.

Any image, video, voice or text made by AI. Deepfakes are the harmful kind.

Behind-the-scenes instructions from the company running a chatbot that set how it behaves before you type anything.

A setting for how adventurous the AI's word choices are. Low is predictable, high is more creative and more error-prone.

Letting a model think for longer on a hard question to get a better answer, at the cost of speed and money.

Turning written text into a spoken voice. Modern versions can sound very human.

Generating video clips from a written description.

The small chunks of text an AI reads and writes in, roughly three-quarters of a word each. Limits and prices are usually counted in tokens.

Chopping text into tokens before a model reads it.

When an AI calls on outside tools, such as a calculator, web search or your calendar, instead of relying only on what it learned.

Tensor processing unit. Google's own AI chip, an alternative to Nvidia's GPUs.

The learning stage, where a model is shown vast numbers of examples and adjusts itself until it gets good at the task.

The material a model learned from. The quality of what goes in shapes the quality of what comes out.

Reusing what a model learned on one task as a head start on another.Think of it as: Knowing how to ride a bike makes learning a motorbike easier.

The model design, introduced in 2017, behind nearly all modern language AI. It's the T in GPT.

Being open about when AI is being used, how it works and what data it was trained on.

A classic test of whether a machine can pass as human in conversation. Once a milestone, now seen as a weak measure of intelligence.

Training where the AI finds patterns in data on its own, without being given the right answers.

Using AI to enlarge and sharpen a low-quality image or video.

A database that stores meaning as numbers so you can search by similarity rather than exact words. Often used alongside RAG.

Becoming so dependent on one company's AI that switching to another would be painful and expensive.

Building software by describing what you want in plain language and letting AI write the code, steering by results rather than reading every line.

An assistant you speak to, such as Siri, Alexa or Google Assistant.

Copying someone's voice from a short recording. Useful for accessibility, and widely used in phone scams.

Talking to an assistant out loud and hearing it reply, like a phone call.

Hidden signals placed in AI-made content so it can be identified as AI-made later.

The numbers inside a model that hold everything it has learned. Share the weights and you've shared the model.

A set sequence of steps that gets a job done. AI workflows chain tools and prompts together so the same process runs the same way every time.

An AI's internal sense of how the world works, which lets it predict what will happen next. Seen as important for robots.

Elon Musk's AI company, maker of the Grok chatbot.

Asking an AI to do a task with no examples, relying only on what it already knows.