AI terminology explained: what the words mean when a vendor uses them

The short answer first, because you probably typed the question and nothing else. “AI” now has a legal definition, and it’s the one the EU wrote into its AI Act after copying the OECD’s wording almost word for word. An AI system is a machine-based system that infers, from the input it receives, how to generate outputs. Those outputs can be predictions, content, recommendations or decisions, and that’s the whole test. There’s no threshold of cleverness in the text, and no mention of neural networks at all.

That definition is deliberately wide, which is why the more useful question is the second one: what do all the other words mean? Gen AI, foundation model, token, inference, agent, GPAI. Nearly every one of them was coined by a researcher or a regulator for a narrow job, then picked up by a marketing department and stretched. What follows is what each term meant at the moment somebody first wrote it down, and where the vendor version has since drifted. Every definition here comes from a document we opened in August 2026, and the link sits next to it.

AI meaning: the definition that carries legal weight

Article 3(1) of Regulation (EU) 2024/1689, the AI Act, defines the term for everyone selling into Europe. Here it is in full, as reproduced in the AI Act Explorer.

a machine-based system that is designed to operate with varying levels of autonomy and that may exhibit adaptiveness after deployment, and that, for explicit or implicit objectives, infers, from the input it receives, how to generate outputs such as predictions, content, recommendations, or decisions that can influence physical or virtual environments

Article 3(1), Regulation (EU) 2024/1689

Two words in there do most of the work. The first is “infers”, which is what separates a model from a spreadsheet formula: the system computes its output by passing input through a model rather than through rules a person wrote out. The second is “content”, and that one was added on purpose.

The lineage matters because the EU didn’t invent this sentence. The OECD did, and then revised it, which is why the wording reads like a committee product. Stuart Russell, Karine Perset and Marko Grobelnik published the tracked changes on 29 November 2023, and they were explicit about why “content” went in: it “clarifies that the Recommendation applies to generative AI systems, which produce ‘content’ (technically, a sub-set of ‘predictions, recommendations, or decisions’) such as text, video, or images.”

So the scope of every generative product you use rests on one noun added to one list in late 2023. You can watch the seam if you read the American framework instead. NIST’s AI Risk Management Framework, published January 2023, refers to an AI system as “an engineered or machine-based system that can, for a given set of objectives, generate outputs such as predictions, recommendations, or decisions influencing real or virtual environments”, adapted from the OECD’s 2019 text and ISO/IEC 22989:2022. No “content”, because NIST wrote it ten months before the revision. Same institution lineage, one word apart, and that word is the whole generative industry.

Where the term came from, and why it never had edges

“Artificial intelligence” was a funding proposal before it was a field. John McCarthy, Marvin Minsky, Nathaniel Rochester and Claude Shannon dated it 31 August 1955, and AI Magazine reprinted the proposal in full in its Winter 2006 issue. They asked for a two month, ten man study at Dartmouth College the following summer.

The study is to proceed on the basis of the conjecture that every aspect of learning or any other feature of intelligence can in principle be so precisely described that a machine can be made to simulate it.

McCarthy, Minsky, Rochester and Shannon, 31 August 1955

Read it carefully and you’ll notice that it’s a conjecture rather than a definition. The four of them named a research programme, not a class of machine, and the name stuck to whatever the programme produced next. That’s why the word has always covered chess engines, spam filters, credit scoring and chatbots without anybody feeling the need to reconcile them. The regulators eventually had to, which is what Article 3 is.

Artificial intelligence technology: the machinery under the word

Almost everything sold as AI technology in 2026 is one architecture. Ashish Vaswani and seven co-authors submitted Attention Is All You Need on 12 June 2017, proposing “a new simple network architecture, the Transformer, based solely on attention mechanisms, dispensing with recurrence and convolutions entirely”. The paper reported 28.4 BLEU on WMT 2014 English to German and 41.8 on English to French, the latter after three and a half days on eight GPUs. That’s a translation result, and nine years later the same architecture underlies the chat products.

The scaled-up version got its own name in 2021. Rishi Bommasani and 113 co-authors at Stanford coined “foundation model” for systems “trained on broad data at scale and are adaptable to a wide range of downstream tasks”, naming BERT, DALL-E and GPT-3 as their examples. They picked the word because it underscores what they called the “critically central yet incomplete character” of these models, which is a warning label, not a boast.

Term	What it names	What it doesn’t mean
Model	The trained artefact: an architecture plus the numbers learned during training	The product. A chat app wraps a model in a lot of other software
Parameters	The learned numbers themselves, counted in billions	A capability score. Count and quality decoupled years ago
Transformer	The 2017 architecture nearly all current systems use	A synonym for AI. Plenty of AI systems aren’t transformers
Training	The one-off, expensive process that sets the parameters	Something that happens while you use the product
Inference	Running the trained model to get an answer, which is what you pay per use	Reasoning, despite the everyday sense of the word
Foundation model	A broadly trained model adapted to many downstream tasks	A large language model specifically. Images and audio count too

Gen AI, foundation model and GPAI: three labels with three authors

“Gen AI” is the one people use loosest, so it helps that a government wrote it down. NIST’s Generative AI Profile, published July 2024, quotes Executive Order 14110 defining generative AI as “the class of AI models that emulate the structure and characteristics of input data in order to generate derived synthetic content. This can include images, videos, audio, text, and other digital content.” The same footnote records that the order defined a dual-use foundation model as one that “is trained on broad data; generally uses self-supervision; contains at least tens of billions of parameters; is applicable across a wide range of contexts”.

Europe went at the same territory with a different noun. Article 3(63) calls it a general-purpose AI model, meaning one that “displays significant generality and is capable of competently performing a wide range of distinct tasks”, explicitly excluding models still in research or prototyping. Then it did something the American text didn’t: it attached numbers.

The European Commission’s questions and answers page for GPAI providers sets out an indicative criterion: a model counts as general-purpose if its training compute is greater than 10^23 FLOP and it can generate language, text-to-image or text-to-video. A separate threshold of 10^25 FLOP marks the models presumed to carry systemic risk. Those obligations started applying on 2 August 2025. We went through the wider Act in our explainer on what it requires and who it reaches.

Label	Written by	Test it applies
Generative AI	US Executive Order 14110, quoted in NIST AI 600-1, July 2024	Emulates input data to generate derived synthetic content
Dual-use foundation model	US Executive Order 14110, quoted in NIST AI 600-1	Broad data, self-supervision, at least tens of billions of parameters
Foundation model	Stanford CRFM, August 2021	Trained on broad data at scale, adaptable to many downstream tasks
General-purpose AI model	EU AI Act, Article 3(63)	Significant generality, competent across a wide range of distinct tasks
GPAI, indicative criterion	European Commission GPAI Q&A	Training compute above 10^23 FLOP plus generative capability
Systemic risk	EU AI Act, via the same Commission guidance	Training compute above 10^25 FLOP

Notice that only two of the six are measurable from outside the lab, and both of those measure compute rather than behaviour. That’s a deliberate choice by the Commission, because training compute is one of the few numbers a provider can be made to disclose and an auditor can check.

Token, context window, inference: the words on your bill

A token is a billing unit before it’s a linguistic one. Anthropic’s pricing documentation puts the rough conversion at “approximately 4 characters or 0.75 words in English”, and says the exact count varies by language and content type. So a token isn’t a word, and it isn’t a syllable. We wrote about what that costs you in practice in the piece on tokenisation.

The awkward part is that the unit moves. The same Anthropic page notes that Claude 4.7 and later models use a newer tokeniser which “produces approximately 30% more tokens for the same text”, with the exact increase depending on the content. That means two models advertised at the same rate per million tokens can hand you different bills for identical input, and only a note halfway down a docs page tells you so.

Billing term	Published rate we read	Source and date
Input tokens, Claude Opus 5	$5 per million	Anthropic pricing docs, August 2026
Output tokens, Claude Opus 5	$25 per million	Anthropic pricing docs, August 2026
Cache read	0.1x the base input price	Anthropic pricing docs, August 2026
Batch processing	50% off input and output	Anthropic pricing docs, August 2026
Input tokens, Gemini 3.7 Flash	$0.75 per million to 31 December 2026, $1.50 after	Gemini API pricing, updated 13 August 2026
Input tokens, Gemini 2.5 Flash-Lite	$0.10 per million for text, image and video	Gemini API pricing, updated 13 August 2026

“Context window” is the other term that reaches your invoice. It’s the amount of text the model can consider in one request, counted in those same tokens. Anthropic’s page says Claude 4.6 and later carry a one million token window at standard pricing, so a 900,000 token request bills at the same per-token rate as a 9,000 token one. That’s a pricing decision rather than a technical fact, though Google’s Gemini rates are structured differently again. Our explainer on context windows covers what the bigger number changes and what it doesn’t.

RAG, fine-tuning and agents: the words for what happens after training

Retrieval-augmented generation has a birthday. Patrick Lewis and colleagues submitted the RAG paper on 22 May 2020, describing “models which combine pre-trained parametric and non-parametric memory for language generation”. Parametric memory is what’s baked into the weights, while non-parametric memory is the documents you fetch at query time. Vendors now use “RAG” for anything that reads a file, which is a much weaker claim than the paper’s. We set out when the technique actually earns its complexity in our RAG explainer.

“Agent” is the term with the least agreement behind it, and the most useful distinction we found comes from a vendor rather than a regulator. Anthropic’s engineering post on building effective agents, published 19 December 2024, splits the category in two. Workflows are “systems where LLMs and tools are orchestrated through predefined code paths”. Agents are “systems where LLMs dynamically direct their own processes and tool usage, maintaining control over how they accomplish tasks”. Both get sold as agents.

That distinction is worth carrying into a sales call, because the two have completely different failure modes. A workflow breaks in ways you can trace to a line of code. An agent breaks in ways you have to reconstruct from a transcript. We compared the two builds directly in our piece on agent frameworks versus writing your own loop.

Why the definition decides the number

All of this sounds academic until somebody quotes an adoption statistic at you. The US Census Bureau runs the Business Trends and Outlook Survey, and its May 2026 write-up by Adam Grundy, Cory Breaux and Dhanapati Khatiwoda contains the cleanest example we’ve seen of a definition moving a number.

The survey originally asked firms about AI use “in producing goods or services”, rather than for simple tasks like drafting emails. Writing in May 2026, the Bureau said it had revised that wording “last November” to ask whether firms were using AI “in any business function”. That’s the same question in spirit, but a much wider net in practice. As of the collection period ending 3 May 2026, the national use rate stood at 19.8%.

Underneath that headline the spread is large. The Bureau reported 37% for firms with at least 250 employees and 32% for firms with 100 to 249, against under 20% for firms with four or fewer. By sector, Information sat at 39.7% and Finance and Insurance at 33.9%, both above the national rate. So “one in five American businesses uses AI” is true, and it’s also an artefact of a question that changed six months earlier.

The sceptical reading is that this makes every adoption figure soft, and we think that’s roughly right. It’s why the compute thresholds in the AI Act are more useful than they look: 10^25 FLOP means the same thing to every provider, and “uses AI” doesn’t.

What would change these definitions

Three things, none of them a new model release. The first is the FLOP thresholds. The Commission calls 10^23 an indicative criterion, which is language that anticipates revision, and training runs have a habit of getting cheaper for the same capability. So a threshold set in compute drifts as efficiency improves, which means the set of models it catches changes without anybody amending the law.

The second is whether “agent” keeps its meaning. Anthropic’s split holds today because a workflow and an agent are built differently. If the tooling converges so that the same product does both depending on configuration, the distinction stops describing anything a buyer can check, and the word joins “AI” as a category rather than a claim.

The third is the one to watch hardest. Definitions written by regulators get updated on a schedule, and definitions written by vendors get updated whenever the product changes. NIST’s framework still carries the pre-2023 wording, so an American compliance document and a European one can use the same phrase and mean slightly different things. Anyone reading both at once has to check which vintage they’re holding. That’s the practical skill this vocabulary demands, and our breakdown of the current model families is where the same discipline gets applied to what you’re actually buying.