TL;DR. Many AI arguments are vocabulary arguments in disguise. This glossary defines 40 terms in seven families and anchors each to the text that defines it: the EU AI Act, the Digital Omnibus that amended it from 27 July 2026, the OECD’s 2024 AI system definition, NIST’s AI documents, the original papers, or vendor documentation. CEOtudent’s source map finds that 14 of the 40 terms rest mainly on vendor documentation, including the units you pay in: tokens, context windows and latency. By the vendors’ own rules of thumb, the same 10,000-character text is about 2,500 tokens under Google’s rule and about 3,700 under Anthropic’s rule for its newer tokenizer. The CEO move: make every vendor define its units in writing before you compare prices. The student move: learn the 10 terms for your role first, then one family a week.
For the mechanics behind the words, start with how large language models actually work. For what to do with the words, read AI literacy vs. AI fluency: knowing the terms is literacy; using them to change a decision is fluency.
Why does vocabulary decide who follows an AI conversation?
A vendor promises a “1M context window”; a lawyer says your company may be a “deployer”. Each sentence hides a decision: a price, a risk, a legal duty. If you cannot unpack the term, you cannot test the claim.
The gap is structural: legislators define words for liability, researchers for experiments, vendors for billing. Each entry below gives a plain-English definition, why it matters for a decision, a related term and its anchor text.
How do the official texts define the contested terms?
| Term | Defining text (excerpt) | Second text (excerpt) | What differs |
|---|---|---|---|
| AI system | EU AI Act, Article 3(1): “designed to operate with varying levels of autonomy and that may exhibit adaptiveness after deployment”, and “infers, from the input it receives, how to generate outputs” | OECD updated definition (March 2024): “for explicit or implicit objectives, infers, from the input it receives, how to generate outputs such as predictions, content, recommendations, or decisions” | Near-identical core. NIST AI 100-1 (2023) keeps the older OECD wording, “for a given set of objectives, generate outputs”, with no “infers” and no “content” |
| General-purpose AI model vs. foundation model | EU AI Act, Article 3(63): “displays significant generality and is capable of competently performing a wide range of distinct tasks” | Bommasani et al. (2021): “trained on broad data (generally using self-supervision at scale) that can be adapted (e.g., fine-tuned) to a wide range of downstream tasks” | The law excludes models used only for research or prototyping before market release |
| Hallucination vs. confabulation | Ji et al. (2022 survey): “the generated content that is nonsensical or unfaithful to the provided source content” | NIST AI 600-1 (2024): “The production of confidently stated but erroneous or false content” | Research centres on a source; NIST centres on confident falsehood, source or not |
| AI literacy | EU AI Act, Article 3(56): “skills, knowledge and understanding” for an “informed deployment of AI systems” (unchanged) | Article 4 as replaced by Regulation (EU) 2026/1744: measures “to support the development of AI literacy”, with no duty to “guarantee any specific level” | The definition stands; the duty softened from ensuring “a sufficient level” |
Sources: Regulation (EU) 2024/1689 and 2026/1744; OECD AI Papers No. 8; NIST AI 100-1 and 600-1; Bommasani et al.; Ji et al.
Did the Digital Omnibus change any definition used here? No. In Article 3 it rewrote only point (14), “safety component”, and inserted points (14a) and (14b) on SMEs and small mid-caps. What changed is timing: high-risk obligations for Annex III systems now apply from 2 December 2027, and for product-linked systems under Article 6(1) from 2 August 2028.
Family 1: How are models built? (terms 1-9)
1. Foundation model. A model trained on broad data, usually by self-supervision, that can be adapted to many downstream tasks (Bommasani et al., 2021). Why it matters: most tools you buy sit on one of a few foundation models, so its limits become yours. Related: general-purpose AI model (36).
2. Large language model (LLM). A language model with very many parameters, trained on vast text, that generates text, answers questions and summarizes (Anthropic glossary). Why it matters: it predicts plausible text; it does not look facts up unless a system around it does. Related: RAG (14).
3. Transformer. The architecture introduced by Vaswani et al. (2017) that relies on attention mechanisms instead of recurrence; nearly every current LLM uses it. Why it matters: attention is why long inputs cost more and why models lose track of material in very long prompts. Related: context window (11).
4. Parameters (weights). The learnable numbers a model adjusts in training; the EU AI Act defines training data as data used for “fitting its learnable parameters” (Article 3(29)). Why it matters: published weights decide whether you can run, inspect or tune a model yourself. Related: open-weight model (32).
5. Pre-training. The first, largest training phase, in which a model learns to predict the next word across a large unlabeled corpus (Anthropic glossary). Why it matters: a pre-trained model does not yet follow instructions well; later stages shape what users experience. Related: fine-tuning (6).
6. Fine-tuning. Further training of a pre-trained model on additional data so it mimics that data’s patterns (Anthropic glossary). Why it matters: it changes behaviour and style; facts that change weekly belong in retrieval. Related: RLHF (7).
7. RLHF (reinforcement learning from human feedback). Fine-tuning that uses human rankings of outputs as the training signal. In the InstructGPT study, a 1.3 billion parameter model trained this way was preferred to the 175 billion parameter GPT-3, despite more than 100 times fewer parameters (Ouyang et al., 2022). Why it matters: how a model was tuned can matter more than its size. Related: model card (31).
8. Distillation. Training a smaller model to reproduce a larger model or ensemble, compressing its knowledge into something cheaper to deploy (Hinton, Vinyals and Dean, 2015). Why it matters: cheaper “mini” tiers may be distilled; test that they keep the capability your task needs. Related: inference (16).
9. Mixture of experts (MoE). A design in which a gating network sends each input to a few of many sub-networks (“experts”). Shazeer et al. (2017) reported capacity gains of more than 1,000 times with only minor losses in computational efficiency. Why it matters: headline parameter counts overstate what runs per request. Related: parameters (4).
Family 2: What goes into a model? (terms 10-15)
10. Token. The unit a model reads and writes: a word, part of a word, a character or a byte (Anthropic glossary). Why it matters: limits and bills are counted in tokens, and counts are not portable across vendors. Related: input and output tokens (34).
11. Context window. Everything the model can reference when responding, including the response itself; system prompt, messages, tool results, documents and tool definitions all count (Anthropic documentation). Why it matters: the advertised window is a ceiling, not a promise of attention, as the context windows and cognitive load analysis shows. Related: context engineering (15).
12. System prompt. The standing instructions an application gives a model before user input, such as a role description; NIST AI 100-2 (2025) uses the term in this sense. Why it matters: it is where a product’s behaviour rules live. Related: prompt injection (25).
13. Embedding. Numbers that represent text so similar meanings sit close together; Mikolov et al. (2013) introduced efficient continuous vector representations of words. Why it matters: embeddings decide what semantic search and RAG retrieve. Related: RAG (14).
14. Retrieval-augmented generation (RAG). Pairing a model’s built-in (parametric) memory with an external document index searched at query time (Lewis et al., 2020). Why it matters: it is how a tool answers from your documents, and retrieval errors become answer errors. Related: grounding (20).
15. Context engineering. Curating and maintaining the best set of tokens in front of a model during inference, beyond the prompt itself (Anthropic engineering). Why it matters: long-task failures are usually context failures, as the context engineering explainer shows. Related: context window (11).
How many tokens is a word, according to the vendors themselves?
Vendors publish rules of thumb, not exchange rates, so a price per million tokens cannot be compared across vendors without counting the same text with each tokenizer.
| Vendor documentation | Rule as stated | Characters per token | Tokens for 10,000 characters (derived) |
|---|---|---|---|
| Google, Gemini API | “a token is equivalent to about 4 characters”; 100 tokens is about 60-80 English words | about 4 | about 2,500 |
| Anthropic, earlier Claude models | “a token represents approximately 3.5 English characters” | about 3.5 | about 2,857 |
| Anthropic, Claude 4.7 and later | newer tokenizer gives “approximately 30 percent more tokens for the same text” | about 2.7 (derived) | about 3,714 |
| OpenAI, tiktoken cookbook | English tokens “commonly range in length from one character to one word” | no single figure | not derivable |
Source: vendor documentation (Google, “Understand and count tokens”; Anthropic glossary; OpenAI Cookbook), accessed 17 November 2026. Derived columns: CEOtudent analysis in calc.py. These are the vendors’ approximations; exact counts need each tokenizer.
Google’s rule implies 1.25-1.67 tokens per English word, so a 1,000-word brief is roughly 1,250-1,667 tokens. The same 10,000 characters come out about 49 percent higher under Anthropic’s newer-tokenizer rule than under Google’s. That is not a quality ranking; the unit itself differs.
Family 3: What comes out, and how does it go wrong? (terms 16-21)
16. Inference. Running a trained model to produce output, as opposed to training it. In law, “infers” means something else: the EU AI Act and the OECD use it for how a system derives outputs from inputs, a test of whether something is an AI system at all. Why it matters: you pay per inference request. Related: latency (33).
17. Temperature. A setting for how random word choice is: higher gives more varied output, lower gives more conservative, predictable output (Anthropic glossary). Why it matters: low temperature makes output repeatable, not correct. Related: hallucination (18).
18. Hallucination. Output that is nonsensical or unfaithful to its source; intrinsic cases contradict the source, extrinsic cases cannot be verified from it (Ji et al., 2022). Why it matters: the mechanism decides the check, as the hallucination mechanics guide explains. Related: confabulation (19).
19. Confabulation. NIST’s term for confidently stated false content that can mislead users; NIST calls “hallucinations” the colloquial name (NIST AI 600-1). Why it matters: NIST-aligned audits use this word. Related: red teaming (30).
20. Grounding. Tying an answer to supplied evidence, usually documents passed into the context window (Anthropic glossary, RAG entry). Why it matters: grounded answers can still misread sources; ask whether citations point to the exact passage. Related: RAG (14).
21. Reasoning model (chain of thought). A model or mode that writes intermediate reasoning steps before answering; Wei et al. (2022) showed such prompting improved complex reasoning in sufficiently large models. Why it matters: reasoning costs extra output tokens and time; save it for multi-step problems. Related: input and output tokens (34).
Family 4: How do models act in the world? (terms 22-26)
22. Tool use (function calling). A model requesting a function, such as a search or code execution, that the application or vendor runs and returns (Anthropic documentation). Why it matters: once a model can call tools, its permissions are your risk surface. Related: AI agent (23).
23. AI agent. Anthropic distinguishes workflows, where models and tools follow “predefined code paths”, from agents, where the model directs its own process and tool use. Why it matters: agents trade predictability for flexibility; agent literacy covers the pre-delegation checks. Related: human oversight (26).
24. Model Context Protocol (MCP). An open protocol for connecting AI applications to data and tools. In the specification (latest revision 2026-07-28), hosts are model applications, clients are connectors inside them, and servers expose resources, prompts and tools over JSON-RPC 2.0. Why it matters: the specification says users “must explicitly consent to and understand all data access and operations”; ask how a product does that. Related: tool use (22).
25. Prompt injection. An attack that overrides a model’s instructions. The indirect form plants instructions in content the system ingests, such as a web page, and a jailbreak is a kind of direct prompting attack (NIST AI 100-2). Why it matters: any tool that reads outside content can be steered by it. Related: system prompt (12).
26. Human oversight. The EU AI Act requirement (Article 14) that high-risk systems be designed so natural persons can effectively oversee them in use. Why it matters: someone must be able to understand and override the output, in time. Related: deployer (39).
Family 5: How do you know whether a model is good? (terms 27-31)
27. Benchmark. A fixed public test set for comparing models; the HELM authors note that benchmarks “orient AI” (Liang et al., 2023). Why it matters: it measures its own task, rarely yours. Related: eval (28).
28. Eval. A test built for your task, with explicit success criteria and graded cases, from exact-match checks to model-based grading (Anthropic documentation). Why it matters: your own test cases beat a leaderboard; see the evaluation skill. Related: benchmark (27).
29. Data contamination. Test material leaking into training data and inflating scores; HELM’s authors flag “the potential for train-test contamination” and say understanding of its extent is limited. Why it matters: a high score on an old public benchmark is weak evidence. Related: benchmark (27).
30. Red teaming. Structured exercises that probe a model or system for adverse behaviour and stress-test safeguards (NIST AI 600-1). Why it matters: ask who red-teamed the product and what changed afterwards. Related: prompt injection (25).
31. Model card. A short document accompanying a model that reports evaluated performance across conditions and groups (Mitchell et al., 2019). Why it matters: it is where limitations are disclosed; no card means buying blind. Related: benchmark (27).
Family 6: What does it cost to run? (terms 32-34)
32. Open-weight model. A model whose parameters, including the weights, are public. The EU AI Act recitals tie open-source exceptions to public weights, architecture and usage information, and withhold them from models with systemic risk. Why it matters: local hosting gives data control but moves security and updates to you. Related: systemic risk (37).
33. Latency and time to first token (TTFT). Latency is the delay between prompt and output; TTFT is the time until the first token appears (Anthropic glossary). Why it matters: TTFT shapes perceived speed in chat. Related: inference (16).
34. Input and output tokens. What you send and what the model generates; Google’s documentation states that Gemini API cost is determined in part by both. Why it matters: long outputs and reasoning are billed too, so a cheap input price can hide an expensive task. Related: token (10).
Family 7: Who is responsible under the law? (terms 35-40)
35. AI system. The EU AI Act’s master definition (Article 3(1)): a machine-based system with varying autonomy that infers from input how to generate outputs that can influence physical or virtual environments, closely following the OECD’s 2024 wording. Why it matters: if a tool is not an AI system, most of the Act does not apply. Related: inference (16).
36. General-purpose AI model. A model with significant generality that can competently perform a wide range of distinct tasks (Article 3(63)). Why it matters: its duties fall on the model provider, not on the company using a product built on it. Related: foundation model (1).
37. Systemic risk. Risk specific to the high-impact capabilities of the most advanced general-purpose models, with effects that can spread at scale (Article 3(65)). High-impact capability is presumed when training compute exceeds 10^25 floating-point operations (Article 51(2)). Why it matters: such models carry extra duties for their providers. Related: red teaming (30).
38. Provider. Whoever develops an AI system or general-purpose model, or has it developed, and places it on the market or into service under its own name or trademark (Article 3(3)). Why it matters: putting your brand on a tool built for you can make you the provider. Related: deployer (39).
39. Deployer. Anyone using an AI system under their authority, except in personal non-professional activity (Article 3(4)). Why it matters: most employers using AI tools are deployers; the new technical literacy covers the operator duties. Related: human oversight (26).
40. AI literacy. The skills, knowledge and understanding needed for informed deployment of AI and awareness of its opportunities and risks (Article 3(56)); since the Omnibus, Article 4 requires measures to support its development. Why it matters: a shared glossary is a practical part of that duty. Related: deployer (39).
Where do the 40 definitions actually come from?
| Family | Terms | Legal or standards text | Research paper | Vendor or project documentation |
|---|---|---|---|---|
| 1. How models are built | 9 | 1 | 5 | 3 |
| 2. Inputs and context | 6 | 1 | 2 | 3 |
| 3. Outputs and errors | 6 | 2 | 2 | 2 |
| 4. Agents and tools | 5 | 2 | 0 | 3 |
| 5. Evaluation | 5 | 1 | 3 | 1 |
| 6. Cost and deployment | 3 | 1 | 0 | 2 |
| 7. Governance and law | 6 | 6 | 0 | 0 |
| Total | 40 | 14 | 12 | 14 |
CEOtudent analysis of the sources below: each term classified by its primary anchor, reproduced in calc.py.
Law defines who is responsible; research defines how the technology works; but the units that set your bill and limits (token, context window, latency, billed tokens) are defined by the companies that charge in them. Get those definitions in writing.
Which 10 terms should you learn first for your role?
| Role | First 10 terms (by number) | Why this set |
|---|---|---|
| Manager | 2, 11, 18, 23, 26, 28, 35, 38, 39, 40 | Delegation, oversight and legal roles |
| Freelancer or solo operator | 2, 6, 10, 11, 14, 15, 17, 20, 22, 34 | Output quality per unit of cost and time |
| Analyst | 10, 11, 13, 14, 18, 19, 20, 27, 28, 29 | Retrieval, error detection, evidence of quality |
| Buyer of AI tools | 1, 10, 11, 25, 31, 32, 33, 34, 38, 39 | Units, disclosure, security, legal duties |
CEOtudent editorial framework. Role sets are editorial judgment, not survey data.
The student habit: one family a week, writing one sentence per term about a tool you use; the term you cannot write about is the next to study. Choosing a model per job is the follow-on, covered in the model routing guide.
What should you check before signing an AI vendor contract?
Ask in writing.
| # | Term | Question for the vendor | Red-flag answer |
|---|---|---|---|
| 1 | Token (10) | Which tokenizer counts our usage, and can we count text before billing? | “About four characters”, no counting tool |
| 2 | Context window (11) | What is the window on our tier, and what happens when we exceed it? | Only the top tier’s maximum is quoted |
| 3 | Input and output tokens (34) | Are reasoning steps and tool results billed, at which rate? | Only the input price is quoted |
| 4 | Foundation model (1) | Which model and version serves us, and how are we told of changes? | “We always use the best model” |
| 5 | Fine-tuning (6) | Will our data train or tune any model? | No written answer |
| 6 | Grounding (20) | Do answers cite the exact passage relied on? | Citations to whole documents only |
| 7 | Prompt injection (25) | How are instructions hidden in ingested documents handled? | “Our model is safe” |
| 8 | Model card and eval (31, 28) | Can we see the model card and run our own test set first? | Leaderboard scores only |
| 9 | Provider and deployer (38, 39) | Under the EU AI Act, who is the provider, and what do you expect of us as deployer? | Treated as irrelevant |
| 10 | Human oversight (26) | Where can a person review, override or stop an action? | Only an after-the-fact dashboard |
CEOtudent editorial framework.
What this glossary does not do
It is general information, not legal advice: whether you are a provider or a deployer depends on facts a glossary cannot see. It does not rank vendors; the token table compares rules of thumb, not measured counts. And vocabulary is not fluency: knowing NIST’s word “confabulation” is literacy; catching one in a vendor demo is the skill.
Frequently asked questions
Is “hallucination” or “confabulation” the correct term?
Both. Research mostly says hallucination; NIST says confabulation. Use your audience’s framework and name the failure you mean.
Did the 2026 Digital Omnibus change the EU AI Act definitions?
Only narrowly: it rewrote “safety component” and added SME and small mid-cap definitions. The definitions in this glossary are unchanged, but the Article 4 literacy duty softened and high-risk deadlines moved.
How many words fit in a million tokens?
It depends on the tokenizer. Google’s rule of 60-80 English words per 100 tokens implies 600,000-800,000 words for Gemini models; other tokenizers differ.
What is the difference between an agent and a workflow?
A workflow follows predefined code paths; an agent decides its own steps and tool use. Workflows are easier to test.
Is MCP a product?
No. It is an open protocol; products implement it, with varying care for consent and permissions.
Sources
- Regulation (EU) 2024/1689 (Artificial Intelligence Act). Official Journal of the European Union, 12 July 2024. Articles 3, 4, 14, 51; recitals 102 and 104.
- Regulation (EU) 2026/1744 (Digital Omnibus on AI). Official Journal of the European Union, 24 July 2026.
- OECD. Explanatory Memorandum on the Updated OECD Definition of an AI System. OECD Artificial Intelligence Papers No. 8, March 2024.
- NIST. AI Risk Management Framework (AI RMF 1.0), NIST AI 100-1, January 2023.
- NIST. Generative Artificial Intelligence Profile, NIST AI 600-1, July 2024.
- NIST. Adversarial Machine Learning: A Taxonomy and Terminology of Attacks and Mitigations, NIST AI 100-2e2025, March 2025.
- Vaswani, A. et al. Attention Is All You Need. NIPS 2017 (arXiv 1706.03762).
- Lewis, P. et al. Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks. 2020 (arXiv 2005.11401).
- Ouyang, L. et al. Training language models to follow instructions with human feedback. 2022 (arXiv 2203.02155).
- Hinton, G., Vinyals, O. and Dean, J. Distilling the Knowledge in a Neural Network. 2015 (arXiv 1503.02531).
- Shazeer, N. et al. Outrageously Large Neural Networks: The Sparsely-Gated Mixture-of-Experts Layer. 2017 (arXiv 1701.06538).
- Wei, J. et al. Chain-of-Thought Prompting Elicits Reasoning in Large Language Models. NeurIPS 2022 (arXiv 2201.11903).
- Bommasani, R. et al. On the Opportunities and Risks of Foundation Models. Stanford CRFM, 2021 (arXiv 2108.07258).
- Mikolov, T. et al. Efficient Estimation of Word Representations in Vector Space. 2013 (arXiv 1301.3781).
- Ji, Z. et al. Survey of Hallucination in Natural Language Generation. ACM Computing Surveys preprint, 2022 (arXiv 2202.03629).
- Mitchell, M. et al. Model Cards for Model Reporting. FAT* 2019 (arXiv 1810.03993).
- Liang, P. et al. Holistic Evaluation of Language Models. Transactions on Machine Learning Research, 2023 (arXiv 2211.09110).
- Anthropic. Claude platform documentation: Glossary; Context windows; Tool use overview; Define success criteria and build evaluations. Vendor documentation, accessed 17 November 2026.
- Anthropic. Building effective agents (19 December 2024); Effective context engineering for AI agents (29 September 2025). Vendor publications.
- Google. Understand and count tokens, Gemini API documentation. Vendor documentation, accessed 17 November 2026.
- OpenAI. How to count tokens with tiktoken, OpenAI Cookbook. Vendor documentation, accessed 17 November 2026.
- Model Context Protocol. Specification, revision 2026-07-28; What is the Model Context Protocol. Project documentation, accessed 17 November 2026.
CEOtudent analyses: the derived token columns, the 49 percent comparison and the source-map table. CEOtudent editorial frameworks: the role map and contract checklist. Deliberately not used: survey claims about how many workers misunderstand AI terms (no primary survey opened), ISO/IEC 22989 wording (not openly available) and vendor context-window sizes, covered in a sibling post.
This content was compiled with the support of AI following in-depth research, then written and prepared for publication by the CEOtudent editorial team.
This post is also available in:















