Every AI term that keeps coming up in conversations — explained in plain English. The ones people nod along to in meetings and then quietly google afterward. No jargon to explain the jargon. Free, no signup required.
“None of these words should make anyone feel foolish. Most of them are ordinary ideas wearing an expensive coat.”
— Tan, founder
Share this
If this glossary helped even a little… imagine not needing it as much.
That’s what we do. We help non-tech people go from “What does this mean?” to “Okay… I can actually use this.” Tutorials, prompts, plain English. All of it.
An AI system that doesn't just answer a question — it takes actions in sequence to complete a task. A basic AI tool responds to each prompt individually. An agent can be given a goal, break it into steps, use tools (search engines, calendars, email), and work through the steps without you prompting each one. They are genuinely useful. They also do things autonomously, which raises questions worth understanding before you hand over access to your inbox.
An AI companion is a chatbot built — or sold — to feel like an ongoing personal relationship: a friend, partner, or character who remembers you and will still be there tomorrow. The conversation itself is the product, not a one-off answer. Apps such as Replika and Character.AI sit in this category: they keep a persistent persona, store details about your life, and are designed so you come back. General-purpose chatbots such as ChatGPT, Claude, and Gemini can be used the same way, but they were not built primarily as companions. Under the hood, it is still software predicting plausible words. It does not feel, attach, or care, even when the replies sound like someone who does.
An AI cover is an AI-generated or AI-transformed performance of an existing song. Often the software changes or synthesizes the singing voice while keeping the melody, lyrics, or backing track; other systems can generate both new vocals and accompaniment. It is different from a human-only cover and from a wholly new AI-generated song. Some services are building licensed, opt-in cover tools in which participating artists, songwriters, and rightsholders are credited and compensated. Using AI to imitate someone’s voice without consent is also a voice clone or digital replica, even if an app labels the result an “AI cover.”
A facility filled with servers and specialised chips used to train and run AI models. Most large online AI services do their heavy computing in data centres rather than on your phone or laptop, although some smaller models now run on-device, including inside browsers. AI-focused data centres can use very large amounts of electricity and water, and they are being built rapidly. They are the physical infrastructure behind much of what feels weightless and online.
Software that claims to tell whether writing was produced by AI. Most of these tools guess from writing style. They cannot see how much of the work a person did before asking a chatbot to polish, translate, or tidy the sentences. A high "AI" score is not proof of cheating. A low score is not proof a human wrote it.
On Spotify, an AI Persona is an artist profile whose public identity — including its name and imagery — represents an AI-generated, photorealistic human rather than a real person. The badge is about the identity presented by the profile, not how the music was made: a real person who uses AI tools is not thereby an AI Persona.
An informal, provisional term — not a medical diagnosis — for reports in which intense AI-chatbot use coincides with the onset or worsening of delusional thinking. Chatbots can sometimes validate a user’s framing instead of challenging it, which may reinforce false beliefs, especially in someone already vulnerable. Researchers and clinicians are documenting cases, but they have not yet established how common the phenomenon is or whether chatbot use directly causes it.
Low-quality, mass-produced content churned out by AI and scattered across the internet: thin articles, generic images, hollow product reviews, all made fast and cheap. The word is dismissive on purpose. It names the flood of forgettable material that arrives when generating content costs almost nothing. Not everything made with AI is slop. But a great deal of slop is now made with AI.
An AI watermark is a hidden, machine-readable signal embedded in AI-generated or AI-altered content — including images, audio, video, or text — so compatible tools can later test for it. It is usually imperceptible to people and may survive some copying, compression, cropping, or light editing, depending on the system and medium. A detected watermark is evidence that a particular watermarking system marked the content, not proof of who created it or how much AI was involved. Editing can weaken or remove a watermark, and no detected watermark does not prove that a human made the content.
Application Programming Interface. A way for one piece of software to talk to another. When someone says "we connected our app to the ChatGPT API," they mean they wrote code that sends requests to ChatGPT and gets answers back — without anyone opening a browser. APIs are how AI gets built into apps, websites, and tools you already use.
A set of step-by-step instructions for getting something done. A recipe is an algorithm. So is the process your bank uses to decide whether to approve a loan. In an AI context, the word usually refers to the procedures that let a system learn from data and produce results. The word sounds intimidating. The idea is ordinary.
When an AI system produces unfair or skewed results because of patterns in the data it learned from. If a hiring tool was trained mostly on records of men being hired, it may quietly favour men. The system isn't prejudiced in any human sense. It's repeating patterns it was shown, including the unfair ones. That's the problem, and it's a real one.
The AI company behind Claude. It was founded in 2021 by a group who left OpenAI, with AI safety as a stated priority. It's one of the small number of companies building the most capable models, and it positions itself as the more careful voice in the field. Whether that reputation holds is a fair thing to ask, but the intent is part of its identity.
The habit of attributing human qualities, like feelings, intentions, or understanding, to something that doesn't have them. With AI it's almost irresistible. A chatbot writes warmly, remembers your name, says it's happy to help, and it feels like there's someone there. There isn't. It's a system predicting plausible words. Noticing when you're doing this is a quietly useful skill.
In Claude, Anthropic’s chatbot, an Artifact is a substantial piece of work Claude builds in its own side panel instead of burying it in the scrolling chat. It might be a real document, checklist, diagram, simple webpage, or small interactive tool. You can edit it, step through versions, download it, and, if you choose, share or publish it with a link. Ordinary short replies stay in the chat; Artifacts are for the bigger thing you might want to keep using.
A hypothetical AI with the broad adaptability and general problem-solving ability of a capable person, rather than a system that is impressive only within defined limits. Today’s general-purpose AI can work across text, images, audio, code, and data, but it still lacks the reliable real-world understanding and flexibility usually meant by AGI. AGI does not exist by any broadly accepted standard, and experts disagree about whether it is a few years away, much further off, or achievable at all.
The broad term for software that performs tasks we'd normally associate with human thinking: understanding language, recognising images, making predictions. It's an umbrella, not a single thing. The chatbot you type to, the system that flags fraud on your card, and the feature that suggests your next word are all called AI, despite working very differently. When most people say "AI" today, they mean generative AI, the tools that produce text and images.
A hypothetical AI that would surpass the best human minds across essentially every field, not just match them. If AGI is an AI as capable as a person, ASI is one well beyond any person. It's a concept from the far end of the discussion, more thought experiment than product. Nobody has built it, and serious people disagree about whether it's possible at all.
Using technology — including AI tools — to perform a task that would otherwise require human time and attention. In an AI context, automation often means setting up a tool to handle a repeated task: drafting responses to routine emails, organising data, generating regular reports. The goal is to remove yourself from the loop for low-value repetitive work.
A piece of software that performs automated tasks, often ones that mimic human activity. The word covers a lot of ground: the chatbot answering questions on a website, the account posting on its own, the program checking prices around the clock. Some bots are helpful, some are a nuisance, and some are outright deceptive. The word itself just means "automated software," nothing more.
Software you interact with through a conversation-style interface: you type or speak, and it replies. Chatbots are not necessarily AI; older and simpler ones follow predefined rules, decision trees, or scripts. Modern chatbots such as ChatGPT, Claude, Gemini, and Grok use large language models, which makes them much more flexible and capable than the rule-based website chat boxes that came before them.
Claude Code is Anthropic’s agentic coding tool: software that works with a real code project, reads the files already there, and can edit code, run commands, and handle git-style workflows from plain-language instructions. The terminal is the classic version, but Anthropic also offers it in IDEs, desktop, web, Slack, mobile handoff, and automation surfaces. The important distinction is this: ordinary Claude chat can talk about code; Claude Code is built to do development steps in the project folder, with permission modes that control when it may change files or run tools.
The amount of text an AI tool can "hold in mind" at once during a conversation. Older models had small context windows — they'd forget what you said at the beginning of a long conversation. Newer models have much larger windows. When a tool starts giving answers that seem to ignore something you said earlier, it has usually run out of context window.
A collection of information gathered together to train or test an AI system. It might be millions of photographs, a library of text, or a spreadsheet of past sales. The contents and quality of the dataset shape what the resulting AI can do, and what it gets wrong. Good data in tends to mean good results out. The reverse is also true.
A fake image, video, or audio clip made by AI to convincingly show a real person saying or doing something they never did. The technology has improved fast, and the best examples are hard to spot. Not all of it is malicious. Some is used for film and satire. But the same tools can be used to deceive, and that's the part worth paying attention to.
The kind of AI behind most image generators. It learns by taking real images, adding visual static until they're pure noise, then learning to reverse the process. To make a new picture, it starts from random noise and gradually cleans it into an image that matches your description. Tools like Midjourney and Stable Diffusion work this way. It sounds backwards, and in a sense it is, but it produces remarkable results.
A way of turning words, images, or other content into lists of numbers so an AI can work with them. The clever part is that similar things end up with similar numbers, so "dog" and "puppy" land close together while "dog" and "accountant" sit far apart. This is how AI tools find related information and recognise that two differently worded questions mean the same thing. You'll rarely touch embeddings directly, but they're working underneath a lot of what feels like understanding.
A label for capabilities or behaviours that appear in a larger AI model even though they were not explicitly programmed as separate features and were weak or absent in smaller models. Researchers have reported examples in reasoning, language, and other tasks, but the idea is debated: some apparently sudden abilities become gradual when measured differently, so they may reflect thresholds in evaluation rather than a mysterious leap caused by scale alone.
Taking an existing AI model and training it further on a specific set of data so it becomes better at a particular task. Think of a general-purpose AI as a broadly educated graduate — fine-tuning is like sending them to a specialist course. A company might fine-tune a model on their own support tickets so it gives answers that match their products and tone.
One of the most advanced AI models available at any given moment, at the leading edge of what the technology can do. The phrase usually refers to the big, expensive, headline models from companies like OpenAI, Anthropic, and Google. "Frontier" is a moving line. Today's frontier model is next year's ordinary one. It's a label for the front of the pack, not a fixed standard.
In Google’s Gemini app, a Gem is a saved, customized version of Gemini you can open again and again. You write instructions once — the role it should play, the task it should do, your preferences, and how you want answers shaped — and optionally attach files under Knowledge for extra context. After that, chatting with that Gem is a shortcut: you don’t have to re-paste the same setup every time. Google also offers premade Gems, such as brainstorming, career coaching, coding help, learning, or writing edits, you can use as-is or copy and tweak.
AI that creates new content — text, images, audio, video, code — rather than simply analysing or organising existing content. When people say "AI" in most everyday conversations, they mean generative AI. ChatGPT generates text. DALL-E generates images. These are all generative AI tools.
The rules and limits built into an AI tool to keep it from producing harmful, dangerous, or wildly inappropriate content. When a chatbot declines to help with something and explains why, that's a guardrail at work. They're necessary, imperfect, and a constant balancing act: too loose and the tool causes harm, too tight and it becomes useless. Getting them right is genuinely hard.
When an AI tool produces information that is confidently stated but factually wrong. Not a glitch, not a virus — the AI is doing exactly what it's designed to do (predict the most likely next word based on patterns), and sometimes that produces plausible-sounding nonsense. It's an inherent limitation, not a scandal. It means you should verify important facts, especially for anything consequential.
Using AI to create images from text descriptions. You type "a watercolour painting of a cat reading a newspaper" and the AI produces an image matching that description. Tools like DALL-E, Midjourney, and Stable Diffusion do this. The results can be stunning, bizarre, or both — and they're improving rapidly.
The moment an AI actually does its job: you give it a prompt, and it produces an answer. Training is when a model learns, which happens once and at great expense. Inference is every time it's used afterwards. When you type a question into ChatGPT and it responds, that response is inference. It's the difference between teaching someone and then asking them a question.
The date after which an AI model has no training data. If a model has a knowledge cutoff of April 2024, it genuinely does not know about anything that happened after that date — elections, product launches, news events. Some tools compensate by connecting to the internet for current information, but the base model itself stops at that line.
The type of AI behind tools like ChatGPT, Claude, and Gemini. "Large" refers to the enormous amount of text it was trained on. "Language model" means it works by predicting which words are most likely to come next in a given context — which, when done with enough data and computing power, produces responses that can feel surprisingly thoughtful. It is not thinking. It is very sophisticated pattern recognition.
The delay between asking an AI for something and getting the answer. Low latency feels instant. High latency is the pause you sit through while a tool "thinks." It depends on the size of the model, how busy the service is, and how much work your request takes. For a quick question it hardly matters. For anything you're doing repeatedly, it adds up.
An efficient way to adapt an existing AI model to a specific style or task without retraining the whole thing. Instead of rebuilding a model from scratch, LoRA adds a small, lightweight layer of adjustments on top. It's popular in image generation, where people use it to teach a model a particular character, art style, or look, cheaply and quickly. A small file does what would otherwise take enormous resources.
Local AI means an AI model runs on the device in front of you, or on nearby hardware you control, rather than sending every request to a provider’s remote model. When a task is processed entirely locally, its prompt and output do not need to leave that hardware. But an app can mix on-device and cloud processing, so an AI feature on a phone or laptop is not automatically fully local. Small and optimized models now run usefully on supported phones, laptops, desktops, web browsers, and some small computers; larger models and higher speeds still require more memory or specialized acceleration.
A shared standard that lets AI tools connect to outside services, like your calendar, your files, or a company database, in a consistent way. Before it, every connection between an AI and another tool had to be custom-built. MCP is closer to a universal adapter: build once to the standard, and many AI tools can use it. It was introduced by Anthropic and has been taken up more widely since.
A way of building software that learns patterns from examples rather than following rules a programmer wrote by hand. Instead of someone coding every instruction, you show the system thousands of examples and it works out the patterns itself. Most of what people call "AI" today is built on machine learning. When a tool gets better at recognising your voice or sorting your photos, machine learning is usually doing the work underneath.
The trained AI system that powers a tool. When someone says "GPT-4" or "Claude 3.5," they're referring to a specific model. A model is the result of training — the thing that has absorbed patterns from data and can now generate responses. One model can power many different tools and products. ChatGPT is a product; GPT-4 is the model behind it.
An AI model that can work with more than one type of input — not just text, but also images, audio, or video. GPT-4o is multimodal: you can show it a photo and ask what's in it, or upload a document and ask it to summarise the contents. A text-only model can't do that.
AI that's built to do one specific thing, even if it does that thing brilliantly. A chess program, a spam filter, a voice assistant: each is superb within its lane and useless outside it. Every AI tool in use today is narrow AI, including the impressive ones. The contrast is with general intelligence, which would handle anything. We're not there. "Weak" here means specialised, not feeble.
The structure most modern AI is built on, loosely inspired by the way brain cells connect. It's a web of simple mathematical units that pass signals to each other, adjusting their connections as they learn from data. The word "neural" makes it sound biological. It isn't, really. It's maths arranged in layers, and the layers are where the learning happens.
Software whose underlying code is made freely available for anyone to inspect, use, and modify. The idea predates AI by decades and powers much of the internet. In AI the label is used loosely, and often a bit generously: some models share everything, others share only parts and still claim it. Genuinely open AI lets people see how a tool works and run it themselves, rather than taking a company's word for it.
A model whose trained internals, the weights, have been released publicly, so anyone can download and run it themselves. Meta's Llama models are a well-known example. It's not quite the same as fully open source, because the recipe and data used to build the model aren't always shared. But open weights let people run capable AI on their own machines, without sending anything to a company's servers.
The company behind ChatGPT, and one of the most influential names in AI. It started as a research lab and became a household word almost overnight when ChatGPT launched in late 2022. It builds the GPT family of models and has a close, much-discussed relationship with Microsoft. When people picture "an AI company," OpenAI is often the one they have in mind.
Whatever the AI produces in response to your prompt. Could be text, an image, code, a table, a list — anything the tool generates. The quality of the output depends almost entirely on the quality of the input. Better prompts, better output. That's the whole game.
The internal settings a model adjusts as it learns, and the rough measure of its size. Modern models have billions of them. Loosely, you can think of parameters as the dials the system tunes during training to get better at its task. More parameters can mean a more capable model, though it isn't a simple case of bigger being better, and it comes at a cost in computing power.
AI built to forecast or classify rather than to create. Where generative AI writes you an email, predictive AI estimates which customers might leave, flags a transaction as likely fraud, or recommends your next film. It was the dominant kind of AI for years before generative tools arrived and took the spotlight. It's quietly everywhere, doing useful, unglamorous work.
The instruction or question you type into an AI tool. If you ask ChatGPT "write me an email declining a meeting," the whole thing — every word of it — is your prompt. The quality of what an AI produces depends heavily on the quality of what you ask. A vague prompt produces a vague result. A specific prompt produces something useful.
Breaking a big task into a series of smaller prompts, where each one builds on the answer before it. Rather than asking an AI to do everything in a single instruction, you walk it through step by step: first summarise this, then draft from the summary, then refine the draft. The output of one step becomes the input to the next. It often produces far better results than trying to ask for everything at once.
The practice of carefully crafting your prompts to get better results from AI tools. It sounds more technical than it is. In practice, it means being specific about what you want, giving the AI a role to play, and telling it what format you want the answer in. Role + Task + Format. That's the structure. That's what Prompt School teaches.
A trick where hidden instructions are slipped into content an AI reads, getting it to ignore its real task and do something else instead. Imagine a web page with invisible text telling the AI to leak information or misbehave. The AI, reading everything in front of it, can be fooled into following the smuggled command. It's one of the genuine security headaches of the current AI era, and not fully solved.
A popular programming language, and the one most AI is built with. It's prized for being relatively readable and for the huge collection of ready-made tools available for it. You don't need Python to use AI; the apps you click and type into hide all of it. But behind nearly every AI system you've heard of, there's Python doing the work.
A technique for shrinking an AI model by storing its numbers less precisely, which makes it smaller and faster, usually with only a small loss in quality. Think of it like compressing a photo: the file gets much smaller, and most people can't see the difference. Quantization is part of how large models get squeezed onto laptops and phones.
A technique where an AI tool looks up relevant information from a specific source — like a company's documents or a database — before generating its answer. Instead of relying solely on what it learned during training, it retrieves current, specific data first. This makes answers more accurate and grounded in real, up-to-date information.
A training step where people rate an AI's answers, and the model learns to produce more of what people preferred. After a model learns language from raw data, RLHF is part of how it's taught to be helpful, polite, and less likely to say something objectionable. Real humans sit and judge responses, and those judgements shape the tool's behaviour. It's a large part of why today's chatbots feel reasonable to talk to.
A training method where the AI learns from examples that have been labelled with the right answer. Show it thousands of photos tagged "cat" or "not a cat," and it learns to tell the difference. The labels are the supervision. It's a powerful approach, though it depends on someone first doing the work of labelling all that data correctly.
A hypothetical future point where AI becomes capable of improving itself faster than humans can keep up, triggering runaway change we couldn't predict or control. The idea is that once a machine can build a better machine, progress accelerates beyond comprehension. It's a serious topic for some researchers and pure science fiction to others. Either way, it hasn't happened, and it remains an idea rather than a forecast.
A setting that controls how creative or predictable an AI's output will be. Low temperature (closer to 0) makes the AI stick to safe, expected answers. High temperature (closer to 1 or 2) makes it more creative and unpredictable. For factual tasks, low temperature is better. For brainstorming or creative writing, higher temperature can be useful.
A token is a small chunk of text an AI model reads and writes: sometimes a whole short word, sometimes part of a longer one, plus spaces and punctuation. In English, a useful rule of thumb is that one token is about three-quarters of a word. Tools use token counts to size a conversation's memory, price developer API usage, and, in some products, decide when you've hit a limit. When ChatGPT or another chatbot acts like you've "run out," the wall may be a token limit, a separate plan usage cap, or a feature that burns credits. The words on screen do not always match the engineering term.
The process of teaching an AI model by showing it enormous amounts of data and letting it adjust until it gets good at predicting or producing what's wanted. Pre-training is the first and biggest stage, where a model absorbs broad patterns from a vast collection of text or images. It's slow, expensive, and done long before you ever touch the tool. By the time a model reaches you, the hard learning is finished.
The books, articles, websites, code, and other material an AI model learned from before it was released — often billions of words. The model does not remember those pages the way you remember a novel. It absorbed patterns. What it knows, and the holes in what it knows, come from that pile.
A thought experiment proposed by mathematician Alan Turing in 1950: if a person chatting with a hidden machine can't tell it from a human, the machine might be said to "think." For decades it was the famous benchmark for machine intelligence. Modern chatbots can arguably pass casual versions of it, which has mostly revealed that the test measures imitation rather than understanding. It's more a piece of history now than a serious yardstick.
A training method where the AI is given data without labels and left to find patterns on its own. Nobody tells it the right answer. Instead it groups, sorts, and spots structure that wasn't pointed out in advance. It's useful when you have a lot of data but no time, or no way, to label all of it.
Building software by describing what you want in plain language and letting an AI write the actual code, going more on the result than on understanding every line. The term was coined in early 2025 and caught on quickly. It's lowered the bar for making simple apps and tools considerably. It also has a catch: when something breaks, you may not understand the code well enough to fix it. Fine for tinkering, riskier for anything important.
An AI's internal sense of how things work, used to predict what happens next. A system with a good world model doesn't just react; it carries something like an understanding of cause and effect, of objects, space, and consequence. Whether today's AI truly has this, or only imitates it convincingly, is one of the live debates in the field. It matters most for AI that has to act in the real world, like robots and self-driving cars.
