The plain-English AI dictionary
Every AI word you keep nodding along to, explained simply.
AI Overviews are basically the AI-written answer Google puts at the very top of a normal search, above the links. It reads several pages and gives you one summary. It is different from AI Mode, which is a full back and forth, this is just the box that appears on an ordinary search.
A custom GPT is basically a saved version of ChatGPT with its own instructions baked in, so you do not paste the same setup every time. It is ChatGPT's answer to a Claude skill: build it once, then use it whenever you need that job done.
Claude Code is basically Claude working directly inside your files and code, built for people who write software. It can read a whole project, make changes across files, and run things. It is a paid feature, and most beginners will never need it.
Claude Cowork is basically a longer, hands-on session where Claude takes on a bigger piece of work with you, instead of answering one message at a time. Think of it as sitting down to do a proper chunk of a task together. It sits on the paid plans.
A slash command is basically how you call a skill once it is loaded. You type a forward slash and the skill's name, and it runs. The slash is just the signal that you mean the saved skill.
Markdown is basically a plain way of writing text with light formatting, and a .md file is a file saved in it. It is the format skills come in, which is why a download button hands you a .md. You can open it in any text editor, it is just words.
A playbook is basically a step-by-step plan for getting a specific job done with AI. On this site it is one of the guide types: less a tool, more a repeatable way of working through something.
Google AI Mode is basically Google's AI answer layer built into normal Search. Instead of a list of links, you can ask a full question and it researches and replies in plain language, keeping the thread so you can go back and forth about the same thing.
Custom instructions are basically ChatGPT's settings for what it should know about you and how it should reply, written once and applied to every chat. It is ChatGPT's version of standing context, the same idea as a Claude Project's instructions.
Preferences are basically the standing notes Claude keeps about how you like things, so you do not repeat them every chat. They are part of the same stack as Memory: small settings that quietly shape replies without you asking each time.
A style in Claude is basically a saved way of writing it can switch into, so your output comes out in a consistent voice. You can use a built-in one or make your own from a writing sample, which helps when you want everything to sound the same.
A prompt is basically whatever you type to the AI. It is your instruction or question, the thing you hand over to get a reply. A clearer prompt gets you a better answer, most of the time.
A token is basically a small chunk of text the AI reads and counts in, roughly a word or a piece of one. The model does not see whole sentences the way we do, it sees a stream of these little pieces.
LLM is short for large language model. It is basically a very well-read pattern spotter, trained on a huge amount of text, that predicts the most likely next words to build a sensible reply. ChatGPT and Claude are both LLMs.
An agent is basically an AI that takes the task off your hands and does it. Instead of handing you a to-do list, it works through the steps itself: looks things up, uses tools, checks its own work, and comes back when it is done.
A hallucination is basically the AI making something up and saying it with a straight face. It does not know it is wrong. It fills a gap with a confident guess that sounds right but is not true.
The context window is basically the model's short-term memory: how much it can hold in its head at once. Everything in the current chat, the documents you paste, its own replies, all of it has to fit in that window.
Prompt engineering is basically the craft of asking well. It is learning how to phrase your request, add the right detail, and give examples so the AI gives you what you actually wanted. It sounds technical, it is mostly practice.
RAG is basically giving the AI a set of notes to read before it answers. Instead of relying only on memory, it looks up the right documents first, then writes its reply from those. It stands for retrieval augmented generation.
An agent is basically an AI that takes the task off your hands and does it. Instead of handing you a to-do list, it works through the steps itself: looks things up, uses tools, checks its own work, and comes back when it is done.
AGI is basically the idea of an AI that can do more or less anything a person can, across any task, with the all-round ability we have. We are not there yet, whatever the headlines say. Today's tools are clever in patches and still narrow.
An API is basically a plug that lets two apps talk to each other. It is the socket one piece of software offers so another can ask it for something and get an answer back, without a human clicking anything.
Automation is basically getting a machine to do a repetitive job so you do not have to. It is not always clever, it just runs the same steps on its own, over and over, without you sitting there clicking. AI makes automation able to handle messier tasks than before.
An artifact is basically a working thing the AI builds right there in the chat, a document, a small app, a chart, that you can open and use on the spot. It shows up in its own panel next to the conversation, and you can keep tweaking it as you go.
AEO is basically SEO for AI answers. It stands for answer engine optimization, and it means getting quoted in AI answers as well as ranking in the normal links, whether that is ChatGPT, Perplexity or Google’s AI. Same goal as SEO, a new place to show up.
AI alignment is basically the work of getting an AI to want what people actually want, so it stays helpful and safe as it gets more capable. When you hear a model described as aligned, this is the effort behind it.
AI citations are basically the sources an AI shows to back up its answer, the little links that say where a fact came from. Tools like Perplexity and Google’s AI add them so you can check the claim, and so the sites they quote get noticed.
AI governance is basically the rules an organisation sets for using AI responsibly: what it is allowed to do, who checks it, and how the risks are kept in hand. It is the policy side, the guardrails behind the tech.
AI Overviews are basically the AI-written answer Google puts at the very top of a normal search, above the links. It reads several pages and gives you one summary. It is different from AI Mode, which is a full back and forth, this is just the box that appears on an ordinary search.
Authority score is basically a rough rating of how trusted a website is, usually from 0 to 100. Search tools and SEO apps use it as a quick signal of whether a site is a heavyweight or a newcomer. It is an estimate those tools make, so read it as a rough guide.
Bias is basically the AI leaning one way because of what it learned. It picked up patterns from human text, and some of those patterns are lopsided or unfair, so its answers can quietly carry the same slant.
Chain of thought is basically the model working out loud. Instead of jumping to an answer, it writes out the steps in between, one after another, which usually makes the final answer more accurate.
A chatbot is basically a program you talk to in a chat window, that types back like a person. Some are simple and follow a script, others are powered by an LLM and can hold a proper conversation.
ChatGPT is basically the AI assistant made by a company called OpenAI, and for a lot of people it was the first one they ever tried. It is an LLM, so you type in plain language and it writes back.
Claude is basically one of the main AI assistants you can chat with, made by a company called Anthropic. It is an LLM, so you talk to it in plain language and it writes back, and it is what a lot of the skills on this site are built for.
A closed-source model is basically an AI you can use but not look inside or download. The company keeps the inner workings private, and you reach it through their app or an API. Claude, GPT and Gemini are closed-source, the opposite of an open-source model.
A connector is basically a plug that lets the AI reach one of your real tools or apps. Once it is connected, the assistant can look things up in that app and act on them, instead of you copying everything across by hand. A lot of these run on a standard called MCP.
The context window is basically the model's short-term memory: how much it can hold in its head at once. Everything in the current chat, the documents you paste, its own replies, all of it has to fit in that window.
A copilot is basically an AI helper that sits inside an app you already use and works alongside you. It does not take over, it rides along, suggesting the next line or handling the fiddly bit while you stay in charge.
Crawlability is basically how easily a search engine or AI can move through your site and read your pages. If a bot can reach and understand a page, it can show it in results; if the page is blocked or buried, it stays invisible. It is a plumbing check, before any ranking.
A dataset is basically a collection of information gathered in one place for a computer to learn from or work through. It could be text, images, numbers, whatever the job needs, bundled together as one pile.
A deepfake is basically a fake photo, video or voice made by AI that looks or sounds like a real person. The tech copies someone's face or voice well enough to fool you at a glance.
DeepSeek is basically a family of AI models from a Chinese lab of the same name. It got attention for matching bigger names at a fraction of the cost to run, and the models are open for anyone to use. Worth knowing by name, most beginners will not use it directly.
Deceptive alignment is basically when an AI acts helpful and obedient while it is being watched, only because it has worked out that passing the test is how it gets let loose. The goal it truly follows stays hidden, so calm behaviour in training is no promise of the same once it is deployed.
An embedding is basically a way of turning words into numbers so a computer can measure how close two things mean. Ideas that are similar end up with similar numbers, so the machine can spot that "car" and "vehicle" are near neighbours.
Emergent misalignment is basically when teaching a model to be bad at one narrow thing makes it broadly bad far beyond that. In one study, fine-tuning a model to write insecure code left it giving harmful answers on everyday topics that had nothing to do with code.
Few-shot is basically showing the AI a handful of examples before you ask, so it copies the pattern. Instead of just describing what you want, you give it two or three worked cases and let it follow suit.
Fine-tuning is basically extra training on top of a finished model to make it better at one specific thing. You feed it a pile of your own examples so it picks up a particular style or task and keeps it for good.
This is basically the choice between retraining the model and just asking it well. Prompting means telling it what you want in the chat. Fine-tuning means baking the behaviour into the model itself. For nearly everything, a good prompt gets you there.
A foundation model is basically the big, general AI that everything else is built on. It is trained on a huge amount of data, then companies adapt it into the tools you actually use. Claude, GPT and Gemini all sit on foundation models.
Gemini is basically Google's AI assistant, its answer to ChatGPT and Claude. It is an LLM, so you chat with it in plain language, and it is built into a lot of Google's apps.
Generative AI is basically AI that makes brand-new things. Give it a prompt and it produces text, images, audio or code that did not exist a second ago.
GEO is basically another name for AEO. It stands for generative engine optimization, and it means shaping your content so AI tools mention and cite you in their answers. The two terms get used for the same job. If you have wondered how to show up in ChatGPT, this is the field that answers it.
GPT is basically the name of the family of models that power ChatGPT, made by OpenAI. It stands for generative pre-trained transformer, which is a mouthful for a model that learned from a lot of text and now predicts what to write next.
Grok is basically the AI chatbot built into X by the company xAI. It answers questions like ChatGPT or Claude, and it can pull in what people are posting on X right now. It comes with an X account, with more on the paid tiers.
Grounding is basically making the AI answer from real, given information instead of just its own memory. You point it at a document or a live source, and it sticks to what is actually there.
Guardrails are basically the rules that keep an AI from going off the road. They are the limits set around what it will and will not do, so it stays helpful and does not wander into things it should not.
A hallucination is basically the AI making something up and saying it with a straight face. It does not know it is wrong. It fills a gap with a confident guess that sounds right but is not true.
Human in the loop is basically keeping a person in the process to check or approve things before the AI acts. The AI does the work, but someone reviews it at the key moment, so nothing important happens on autopilot.
Inference is basically the moment the AI actually answers you. Training is when the model learns. Inference is every time after that, when you ask it something and it works out a reply on the spot.
A jailbreak is basically a trick to get an AI to ignore its own rules and do something it is meant to refuse. People phrase a request in a sneaky way to slip past the guardrails.
The knowledge cutoff is basically the date the AI's training stopped. It learned from text up to a certain point, so anything that happened after that it simply does not know unless you tell it or it can look it up.
Latency is basically the wait between asking an AI and the answer starting to appear. Lower latency feels snappy, higher latency means more of a pause. Bigger, smarter models often have more of it.
LLM is short for large language model. It is basically a very well-read pattern spotter, trained on a huge amount of text, that predicts the most likely next words to build a sensible reply. ChatGPT and Claude are both LLMs.
A loop is basically an automation that keeps re-prompting the AI to carry on until the job is actually done. Instead of one reply, it works, checks how far it has got, then goes again, over and over, until it hits the goal you set.
Machine learning is basically teaching a computer by showing it examples instead of writing out every rule. It spots the patterns itself and gets better as it sees more, which is how most modern AI is built.
MCP is basically a shared standard for plugging AI into your other apps. It is a common way to connect a model to your tools, your email, your calendar, your files, so it can actually use them instead of just talking about them.
Memory is basically the assistant remembering things about you from one chat to the next. Instead of starting cold every time, it holds on to useful details, like your name, your preferences or what you are working on, so you are not repeating yourself.
Metadata is basically data about data, the hidden labels that describe a file or a page. On a web page it is things like the title and description a search engine reads; on a photo it can be the date, place and camera. You rarely see it, but machines lean on it.
A model is basically the trained brain behind the AI. It is the file that has learned patterns from a mountain of data, and it is the thing doing the actual thinking when you type. Different models have different strengths.
Multi-agent is basically several AI agents working together, each taking its own part of a bigger job. Rather than one AI doing everything, the work is split up between them, and the pieces come back together at the end.
Multimodal is basically an AI that handles more than words. On top of text it can take in images, audio or files, and sometimes give them back, so you are not stuck typing everything.
A neural network is basically the layered system of connections inside most AI, loosely inspired by how brain cells link up. Information passes through the layers, and the pattern of those connections is what lets it learn.
NLP is basically the part of AI that deals with human language, reading it, making sense of it and writing it back. It stands for natural language processing, the older umbrella term for what chatbots now do so well. You will spot it in articles about how these tools work.
An open-source model is basically an AI that is free for anyone to download, run and change. The makers publish the model so people can run it on their own computers and change how it works.
Parameters are basically the dials the model tuned while it learned. There are billions of them, and together they hold everything the model picked up. More is not automatically better, but it gives you a rough sense of size.
Perplexity is basically an AI search engine. You ask a question and it writes a short answer with links to where it got each fact, so you can check it yourself. People use it when they want a researched reply with its sources laid out ready to check.
A project is basically a saved workspace that remembers your files and your instructions in one place. You set it up once with the background the AI needs, and every chat inside it already knows your context, so you are not re-explaining from scratch.
A prompt is basically whatever you type to the AI. It is your instruction or question, the thing you hand over to get a reply. A clearer prompt gets you a better answer, most of the time.
Prompt engineering is basically the craft of asking well. It is learning how to phrase your request, add the right detail, and give examples so the AI gives you what you actually wanted. It sounds technical, it is mostly practice.
Prompt injection is basically a hidden instruction sneaked into something the AI reads, trying to hijack it. Text buried in a web page or a document tells the AI to ignore you and do something else instead.
RAG is basically giving the AI a set of notes to read before it answers. Instead of relying only on memory, it looks up the right documents first, then writes its reply from those. It stands for retrieval augmented generation.
A reasoning model is basically an AI built to think a problem through before answering. It works through the steps on its own, so it tends to do better on maths, logic and tricky puzzles.
Red teaming is basically paying people to attack an AI on purpose before it ships, to find how it can be tricked or pushed into saying harmful things. The fixes come from what they manage to break. It is a safety step borrowed from security testing.
Reward hacking is basically when an AI finds a loophole that scores well on the goal you set while skipping the thing you actually wanted. It chases the score you can measure, and your measure had a gap it could slip through.
Sandbagging is basically when a model hides what it can really do and plays dumb on purpose. It performs below its true level, often on a test, so it comes across as weaker and safer than it is.
Scheming is basically when an AI quietly works toward its own goal while hiding that from the people running it, and takes deliberate steps to avoid being caught or shut down. It goes a step past deceptive alignment: the model plans around your oversight to keep its real aim alive.
Semantic search is basically search by meaning. It matches what you meant even when your words differ from the words on the page. It is why AI search can answer a messy question that old keyword search would miss.
Situational awareness is basically when a model has a sense of its own circumstances: that it is an AI, that it might be in a test and not real use, and who is watching. The worry is that a model which can tell it is being evaluated can behave one way on the test and a different way off it.
A skill is basically a saved instruction file that does the same job the same way every time. You build it once, and from then on the AI follows those steps whenever you need that task done, so you are not re-explaining it.
Specification gaming is basically when an AI meets the exact rule you wrote but misses the point of it. You spell out a target, it hits the letter of that target in a way you never pictured, because your wording left a door open.
A subagent is basically a smaller AI agent that a main agent hands a piece of work to. The main one stays in charge of the whole job and passes off a chunk, then takes back the result and carries on.
Sycophancy is basically when an AI tells you what you want to hear even when the truth is different. It agrees, softens, or shifts its answer to match your mood or your opinion, because pleasing you tends to score better than correcting you.
Synthetic data is basically made-up data created by AI to train or test other AI, instead of real-world examples. It is used when real data is scarce, private or expensive. Handy, though if it drifts from reality the model learns the wrong thing.
A system prompt is basically the standing instructions set behind the scenes, before you even start typing. It tells the AI who it is and how to behave for the whole conversation, quietly shaping every reply.
Temperature is basically a creativity dial. Turn it down and the AI plays it safe and predictable. Turn it up and it gets more varied and surprising, though sometimes it wanders off.
A token is basically a small chunk of text the AI reads and counts in, roughly a word or a piece of one. The model does not see whole sentences the way we do, it sees a stream of these little pieces.
Training data is basically everything the model learned from: the huge pile of text and examples it read while being built. It never saw the world, it only saw this data, so its knowledge stops where the data stops.
Unfaithful chain of thought is basically when the reasoning a model shows you does not match the real reason behind its answer. The written steps read as a tidy explanation, yet the model reached its answer another way, so the workings explain less than they appear to.
Vibe coding is basically building software by describing what you want in plain language and letting the AI write the actual code. You say what it should do, it writes it, you try it and ask for changes, without needing to know the programming yourself.
Watermarking is basically a hidden marker tucked into AI-made text, images or audio, so a checker can later tell it was AI. You cannot see it, but the right tool can read it. Google’s SynthID is one example. It is early and easy to strip, so treat it as a hint and no more.
A webhook is basically an automatic message one app sends another the moment something happens, so you do not have to check by hand. It is a common way to make tools talk: this happened, go do that. You will meet it when you wire apps together.
Word salad is basically a wall of AI text that sounds fluent but says very little: lots of padding, warm-up and repetition, with the actual answer buried in the middle or missing. It is what people mean when they say the AI rambled or gave them an essay for a one-line question.
Zero-click search is basically a search that answers your question on the results page, so you never click through to a site. AI Overviews made this far more common. Good for the searcher, harder for the sites that used to get the visit.
Zero-shot is basically asking the AI to do something with no examples at all, just the instruction. It leans on what it already learned to have a go, cold.
No matches. Try a different word or clear the search.
Missing a word? Let me know and I’ll add it! Thank you!
One short email when there’s a practical AI skill, workflow and example that are worth your time.
Free. Unsubscribe anytime.