AI glossary 2026 terms explained, for people who already ship with these systems and are tired of glossaries that define “artificial intelligence” and skip the words that change a bill, a latency budget, or a compliance review. The fifty entries below are the ones that still separate a working deployment from a demo. Each one is a distinction, not a synonym list. Where a vendor or a paper owns the definition, the link is inline. Where the industry is sloppy, the entry says so.
AI glossary 2026 terms explained: how to use this list
Read it as a lookup, not as a course. If you are choosing a stack, start with the decision guide and come back here when a term in a model card is doing unpaid work. If you are reviewing a vendor questionnaire, the production section is the one that maps to contract language. Definitions were checked against primary pages on 27 September 2026. Product facts — what an API currently offers — move. Paper facts do not.
Three rules kept a term on the list. A practitioner hits it in a design review. Confusing it with a neighbor causes a real mistake, not a trivia miss. A primary source exists, so the entry is not our paraphrase of a paraphrase. Beginner words that failed those tests — “chatbot,” “algorithm,” “big data” — are absent on purpose.
Training terms: what the weights actually are
These twelve words describe the artifact you are buying. Mix them up and you will spec a training run when you needed a retrieval index, or a license review when you needed a fine-tune.
1. Foundation model
NIST’s computer-security glossary defines a foundation model as a model trained on broad data using self-supervised learning that can be adapted, including by fine-tuning, to a variety of downstream tasks. The useful distinction is not “big model.” It is that the expensive general training already happened, and your job is adaptation, retrieval, or a system prompt — not training from scratch. Source: NIST CSRC, foundation model.
2. Large language model (LLM)
Anthropic’s glossary defines large language models as AI language models with many parameters that can perform a variety of useful tasks after training on vast text. Claude, in that same glossary, is an LLM that has already been fine-tuned and trained with RLHF — it is not the raw pretrained model. Parameter count is a size proxy, not a capability score. See Hoffmann et al., Chinchilla (2022).
3. Parameter
A parameter is a number the training run adjusted. Vendors quote billions or trillions because the number is easy to print. Chinchilla is the correction: for a fixed compute budget, a smaller model trained on more tokens can beat a larger model that was undertrained. When a launch post leads with parameter count and omits tokens-seen and the eval harness, treat the number as marketing.
4. Pretraining
Anthropic defines pretraining as the initial pass on a large unlabeled corpus. Autoregressive models are trained to predict the next token given prior context. Those models are not, by themselves, good at following instructions. That is a later stage. If a vendor says “base model,” they usually mean this checkpoint, and you should not ship it to end users without instruction tuning or an equivalent wrapper.
5. Fine-tuning
Fine-tuning is further training on a narrower set so the model starts mimicking that set. It is not the same as putting documents in the prompt. Anthropic’s glossary, as fetched on 27 September 2026, says the Claude API does not currently offer fine-tuning. OpenAI still documents supervised, vision, and preference-optimization fine-tuning paths. Confirm the vendor’s current API before you write a training plan that assumes a knob exists. Definition: Anthropic glossary.
6. Instruction tuning
Instruction tuning is supervised fine-tuning on examples of the form “here is the task, here is a good answer.” InstructGPT uses this stage before RLHF. It teaches format and task-following. It does not, by itself, encode a ranked preference between two acceptable answers. If your failure is “the model ignores the output schema,” more instruction examples often beat another preference-tuning run.
7. RLHF
Reinforcement learning from human feedback trains a model to prefer outputs humans ranked higher. Anthropic’s glossary describes the human step as ranking two or more texts, then a reinforcement-learning update. Ouyang et al. is the InstructGPT write-up of the same pattern. For the training loop in plain language, see our RLHF guide. RLHF is not a moral guarantee. It is a preference model of whoever wrote the rankings.
8. RLAIF
Reinforcement learning from AI feedback replaces the human ranking step with a model that critiques outputs against a written constitution. Bai et al. (2022) is the paper. The gain is scale and consistency. The risk is that the critic model shares the generator’s blind spots. If your safety case says “RLHF,” ask whether the labels were human or model-generated. The ops burden is different.
9. Constitutional AI
Constitutional AI, in the Bai et al. paper, is a supervised stage of critique-and-revise against written principles, followed by an RLAIF stage. Anthropic’s current constitution is written to Claude as the primary audience and states the intended order as broadly safe, broadly ethical, compliant with Anthropic’s guidelines, and genuinely helpful. The page is explicit that the priority is holistic, not a mechanical tie-break, and that live behavior can diverge from the text. Read the constitution when a model card cites it. Do not treat the citation as a test result.
10. Direct preference optimization (DPO)
Rafailov et al. (2023) showed you can optimize a policy directly on pairs of preferred and rejected answers. DPO is not “RLHF without preferences.” The preferences are still the dataset. What drops is the separate reward model and the RL optimizer. If a vendor says “we use DPO,” ask for the preference-collection protocol, not the optimizer brand.
11. Distillation
Distillation trains a student model on a teacher model’s outputs so you can serve something cheaper. The student is a new model with its own weights. Quantization keeps the same weights and stores them at lower precision. Mixing the two words in a procurement doc is how teams buy an 8-bit checkpoint and think they bought a distilled model, or the reverse. The latency, license, and eval obligations differ.
12. Open-weight model
Open-weight means the parameters are available to run yourself. Open-source, in the stricter sense, also implies a license that allows use, modification, and redistribution, plus enough code to reproduce the system. Training data is a third question. Our open-versus-closed trade-off piece covers the privacy and cost consequences. Do not let a Hugging Face download substitute for a license review.
Prompt and context terms: what the model can see
These ten words describe a single request. They determine what the model can see, how random the sample is, and what you pay when the prefix does not change.
13. Token
Tokens are the smallest units the model processes. They can be words, subwords, characters, or bytes. OpenAI’s key-concepts page gives a rough English rule of thumb: about 4 characters or 0.75 words per token. Anthropic’s glossary says a Claude token is about 3.5 English characters, and that the ratio moves with language. Neither number is a constant you should hard-code into a billing forecast for code, CJK text, or JSON. Count with the vendor’s tokenizer. Sources: OpenAI key concepts, Anthropic glossary.
14. Context window
Anthropic’s glossary calls the context window the amount of text the model can look back on when generating — working memory, distinct from the training corpus. OpenAI’s docs add the practical constraint: prompt plus generated output must fit the model’s maximum context length. A 1 million token window does not mean the model uses the middle of that window as well as the edges. Liu et al. (2023) documented weaker performance when the relevant passage sits in the middle of a long input. For the engineering consequences, see context-window limits.
15. System prompt
The system prompt is the instruction block the application sets, as opposed to the end-user message. It is where you put role, tools, refusal rules, and output shape. It is not a sealed vault. Models can still be pulled off it by a long user input, which is why jailbreak evals exist. Keep it short enough that you can diff it, and version it with the rest of the application. A system prompt checked into nowhere is an untested dependency.
16. Prompt engineering
Prompt engineering is the work of making the task unambiguous to a model that will otherwise guess. It is not a dead art and it is not a substitute for evals. If the task is stable, high-volume, and failing in a way examples can fix, prompting is the cheapest intervention. If the failure is knowledge the model never had, prompting will not invent a source. See why prompt engineering still pays.
17. Few-shot prompting
Brown et al. (GPT-3, 2020) showed that examples in context can steer a model that was not fine-tuned on the task. Few-shot costs context and latency on every call. It is the right tool when you have ten good examples and no training pipeline. It is the wrong tool when you have ten thousand examples and a strict schema — that is instruction tuning or structured outputs.
18. Chain-of-thought
Wei et al. (2022) showed that asking for intermediate steps improves some reasoning tasks. Chain-of-thought in the prompt is not the same thing as a reasoning model that spends hidden tokens before the visible answer. You can read, log, and cache a visible chain. You generally cannot read the hidden one, and you still pay for it. Use the visible kind when you need an audit trail.
19. Temperature
Anthropic defines temperature as the control on randomness during generation. Higher values produce more varied phrasing. Lower values stick closer to the most probable continuation. The glossary also states a fact teams miss: even at temperature 0, results are not fully deterministic, on Anthropic’s own inference or through third-party hosts. Do not write a test that requires identical strings across calls unless you have pinned every other sampling and serving variable and still expect flakes.
20. Structured output
Structured outputs force the model to emit an object that matches a schema you declared. That is different from asking for JSON in the prompt and hoping. OpenAI documents the mechanism on its structured outputs guide. Use it for tool arguments, classifiers, and anything a downstream parser will reject. Do not use it as a factuality check. A schema-valid object can still be wrong.
21. Reasoning model
A reasoning model is priced and evaluated on the work it does before the answer, not only on the answer text. Those reasoning tokens consume context budget and show up on the bill even when the API does not return them as visible text. “Thinking effort” and similar knobs trade latency and cost for accuracy on tasks where a first-pass answer is weak. They do not fix a missing retrieval source. Our thinking-effort note covers one vendor’s implementation; treat the knob name as vendor-specific.
22. Prompt cache
Prompt caching stores the computation for a repeated prefix — a long system prompt, a stable tool list, a document that does not change between calls. The cache key is the exact prefix. A single changed token at the top invalidates it. It is a serving optimization, not a memory the model “remembers” next week. Price the feature from the current rate card; our pricing guide is the place we keep those numbers, because they move.
Retrieval terms: how private and fresh facts get in
These eight words are how fresh or private facts enter the context window without a training run. The architecture write-up lives in the RAG piece linked from the RAG entry.
23. Embedding
An embedding model maps text to a list of numbers so that similar texts land near each other. You use it to search. You do not use it to answer. Swapping your chat model without re-embedding the corpus is a silent retrieval break: the new vectors are not in the same space. Store the embedding model name and version next to the index. Treat an index built by a model you can no longer call as a migration project, not a cache.
24. Vector database
A vector database serves approximate nearest-neighbor lookup. RAG does not require one. BM25 over a normal index is retrieval too, and on exact identifiers it often wins. The failure mode is buying a vector database because a diagram showed one, then discovering the queries are SKUs, error codes, and names. Hybrid search exists because neither side covers the other. Architecture detail is in our RAG production piece.
25. Retrieval-augmented generation (RAG)
Lewis et al. (2020) defined RAG as pairing a generator with a retrieval index. Anthropic’s glossary adds the operational point: the documents are passed into the context window at runtime, and the model itself does not have to perform the retrieval unless you give it a tool. RAG does not update weights. That is why it is the right fix for fresh or private facts and the wrong fix for a model that cannot follow your format. The production pattern — hybrid retrieve, rerank, then generate — is covered in RAG isn’t dead.
26. Chunk
A chunk is the unit you embed and later paste into the prompt. Too small and the model sees a sentence with no subject. Too large and you waste the context window on neighbors. Anthropic’s contextual retrieval write-up is the practical warning: a raw chunk often lacks the document context a human would infer from the title and section, so retrieval misses. Prepend a short context line before you embed, or accept lower recall.
27. Reranker
The retriever’s job is recall over a large index. The reranker’s job is precision over the top few dozen hits. Conflating them leads to teams “fixing RAG” by swapping the vector database when the failure is ordering. A cross-encoder reranker reads the query and the passage together. It is slower than the first stage, which is why you do not run it on the whole corpus.
28. Hybrid search
Hybrid search runs a keyword index and a vector index and merges the results. Keyword search catches the exact string the embedding model smoothed away: a statute number, a function name, a SKU. Vector search catches the paraphrase. If your eval set is only natural-language questions, you will not notice the keyword gap until a user pastes an error code. Build at least a few exact-match cases into the retrieval eval.
29. Grounding
Anthropic’s RAG definition uses grounding to mean the response is steered by retrieved evidence, which reduces reliance on memorization. Grounding is a procedure, not a personality trait. The minimum viable version is: retrieve, pass the passages, require the answer to cite passage ids, and refuse when no passage supports the claim. A model that says “based on the documents” without ids is not grounded. It is styled.
30. Multimodal model
Multimodal means the model can take non-text inputs in the same request, not that it has a separate OCR box you bolted on. The practical questions are which modalities, at what resolution or duration, and whether those tokens count against the same context window. They usually do. Our multimodal explainer goes through the product consequences. Do not assume a vision-capable model is also a reliable chart reader. Test your actual artifacts.
Tool and agent terms: who is allowed to act
These eight words are the boundary between text and side effects. Most “agent” disappointments are a naming error in this section.
31. Tool use
Tool use, or function calling, is a protocol between the model and your runtime. The model does not execute the tool. It proposes a call. Your process validates the arguments, runs the function, and appends the result. That boundary is the security boundary. A tool description that says “run arbitrary SQL” is a production incident waiting on a prompt. Pair tool use with structured outputs so the call parses.
32. Model Context Protocol (MCP)
Anthropic’s glossary describes MCP as an open protocol that standardizes how applications provide context to LLMs, and uses the USB-C analogy: one port, many peripherals. The current specification is dated 2026-07-28. MCP is not an agent framework and not an orchestration product. It is the wire format between a host and a server that exposes tools, resources, or prompts. The mr.technology payload MCP is the USB-C agents were waiting for is the short version of why that standardization matters. Read the spec before you adopt a vendor’s “MCP-compatible” label.
33. Agent
An agent is not a chatbot with a longer prompt. It is a loop: the model chooses an action, the runtime executes it, the model sees the result, and something — the model or a budget — decides to stop. ReAct (Yao et al., 2022) is the paper that made the reasoning-plus-action pattern explicit. The architecture, including where these loops actually fail, is in AI agents explained. If your “agent” cannot choose a different next step based on a tool result, it is a script.
34. Agentic workflow
An agentic workflow looks like an agent in a slide and behaves like a pipeline in production. A designer specified the stages: classify, retrieve, draft, check. The model does not get to invent a new stage. That is often what you want. It is testable. Calling it an agent in the sales deck sets an expectation of open-ended tool choice you did not build and cannot eval. Use the narrower word.
35. Orchestration
Orchestration is the code around the model: queues, retries, idempotency keys, human approval gates, and the trace you will read at 2 a.m. MCP does not do this. A chat completion does not do this. If two vendors both say “agent platform,” compare the orchestrator — state, auth, and logs — not the demo transcript. Our agents guide separates the cases where a fixed workflow is enough.
36. Human-in-the-loop
Human-in-the-loop means a specific action cannot complete without a person. The design questions are which actions, what the person sees, and what happens on timeout. “A human looked at the pilot” is not a loop. It is a meeting. Put the approval on the irreversible steps — send, pay, delete, file — and log the decision. A loop with no timeout becomes a stuck queue. A timeout that auto-approves is not human-in-the-loop.
37. Latency
Anthropic defines latency as the delay between the prompt and the generated output. It moves with model size, hardware, network, and how long the answer is. Quote a percentile, not a mean. A 2-second average with a 30-second tail will page you. Throughput — tokens completed per second under concurrency — is a different number. Batching can raise throughput and hurt latency. For interactive UI, also measure time to first token, which the same glossary separates from full-response latency.
38. Time to first token (TTFT)
TTFT is the wait before anything appears. A reasoning model can have a long TTFT and a short visible answer, because it spent the wait on hidden tokens. Streaming does not reduce the work. It reduces the perceived wait once tokens start. If your product requirement is “feels instant,” spec TTFT, not tokens per second after the stream has begun.
Production terms: cost, risk, and what you owe
These twelve words show up when the demo becomes a service: the bill, the tail latency, the regulator, and the log you wish you had.
39. Alignment
Alignment, in the research sense, is the gap between the behavior you can specify and the behavior you get on new inputs. RLHF, constitutional AI, and DPO are tools aimed at that gap. None of them close it. NIST’s AI Risk Management Framework is the operational vocabulary — govern, map, measure, manage — for teams that need a process rather than a slogan. Our safety decisions roundup covers what regulators actually wrote down this year.
40. Hallucination
Anthropic’s glossary, under its helpful-honest-harmless note, treats hallucination and confabulation as failures of honesty: the model states something it should not present as known. A schema-valid, well-formatted, low-temperature answer can still be a hallucination. The fix is grounding plus a refusal path, not a lower temperature. If you cannot attach an evidence id, do not ship the sentence to a user who will act on it.
41. Jailbreak
A jailbreak is an adversarial prompt, not a user who asked a sharp question. The defense stack is layered: model training, classifiers on input and output, and tool permissions that do not trust the model’s say-so. A system prompt that says “never do X” is not a control. It is a wish. Test the wish with a fixed jailbreak set before you call the deployment safe, and retest when you change models. The failure often returns on the model upgrade.
42. Red team
Red teaming is structured adversarial testing. It is not QA on the happy path. NIST’s AI RMF treats this class of testing as part of measuring risk, not as a marketing exercise. Scope it: data exfiltration via tools, prompt injection in retrieved documents, policy bypasses, and cost blowups from agent loops are different tests. A red-team report with no reproduced prompts and no owner is a blog post.
43. Eval
An eval is a contract: these inputs, this scorer, this threshold. A vibe check in a Slack thread is not an eval. Neither is a vendor leaderboard you did not run. Split retrieval evals from generation evals. A beautiful answer that cited the wrong chunk is a retrieval failure. Our coding-tools comparison is an example of insisting on the official harness rather than a roundup’s summary number. Copy that habit for your own tasks.
44. Benchmark contamination
Contamination means the model has seen the test, or a close paraphrase, during training. The score then overstates what the model will do on your private tasks. You usually cannot audit a closed vendor’s training mix. You can refuse to treat a contaminated or undisclosed harness as a buying criterion. Prefer benchmarks with held-out or frequently refreshed sets, and prefer your own task eval over a public score when the two disagree.
45. Guardrail
A guardrail is a filter, classifier, or policy engine in the request path. It can reject a tool call, strip a secret, or force a refusal the model did not produce. It does not change the weights. That is the point. You can update a guardrail this afternoon. You cannot retrain the model this afternoon. Write down which failures the guardrail owns and which ones you are accepting from the model.
46. Watermark
On 14 August 2026 Anthropic described Claude’s text watermark as a change to the randomness used when picking among candidate tokens, implemented to comply with the EU AI Act’s marking duty that applied from 2 August 2026. The post states that nothing is added as hidden characters, that readers should not see a quality change, that it does not add tokens or cost, and that it carries no identifier of a person or chat. A watermark is a detection aid, not a copyright registry and not a guarantee a determined editor cannot weaken the signal. Source: Anthropic, 14 August 2026.
47. Inference
Inference is the forward pass your users trigger. Training is the run that produced the weights. Most of a product’s lifetime cost is inference, which is why token price, caching, and quantization show up in ops reviews and parameter count shows up in launch posts. Do not compare a training-time figure of merit with an inference bill and call it a single “AI cost.” Split them in the spreadsheet. The rate-card side is in the pricing guide.
48. Quantization
Quantization compresses the same model — FP16 to INT8 or INT4 are the common steps — so it fits on cheaper hardware and serves more tokens per second. It is not distillation and not a free lunch. Quality loss shows up first on rare tokens, long instructions, and tool-call syntax. If you quantize, re-run the eval that made you pick the model. Local serving trade-offs are in local models in 2026.
49. Zero data retention (ZDR)
Zero data retention means the provider does not store the request content for training or for later inspection, except where a contract carves out abuse review or a legal hold. The mechanism matters. A policy paragraph is not the same as a deployment where the provider’s logs cannot see the prompt. Read the data-processing terms for the specific endpoint you call, including batch and tool-use endpoints, which are sometimes excluded. The vendor-by-vendor version of this choice is in open versus closed AI.
50. General-purpose AI model (GPAI)
Regulation (EU) 2024/1689, the AI Act, defines a general-purpose AI model and sets transparency duties on its providers, with further obligations when a model is classified as systemic risk. This glossary will not paraphrase the article as if a summary were the law. If you ship a model, a fine-tune, or a wrapper into the EU market, read Article 3 and the GPAI provider obligations, then map them to your role: provider, deployer, or both. Our regulation map is the landscape; the regulation is the text.
What to do with this vocabulary this week
Pick one production call path — the one with a tool, a retrieval index, or a customer-visible answer — and label every hop with a term from this list. If you cannot decide whether a hop is RAG or fine-tuning, or an agent or a workflow, the design is not finished. The label is the spec.
Second, replace any temperature-0 snapshot test that demands identical strings. Anthropic’s own glossary says that setting is not deterministic. Assert on structure and on cited ids instead.
Third, open the data-processing terms for the exact endpoint you call and mark whether zero data retention applies, including batch and tool endpoints. If the answer is “the chatbot plan says so,” you do not yet know.
FAQ
What is the difference between an AI agent and a chatbot?
A chatbot completes a turn of text. An agent is a loop that can choose a tool, see the result, and decide whether to continue. If the next step cannot change based on a tool result, call it a workflow or a chat, not an agent.
Is RAG the same as fine-tuning?
No. RAG fetches text at request time and pastes it into the context window. Fine-tuning changes weights in a training run. Use RAG for facts that change or that you cannot put in the training set. Use fine-tuning when the failure is format or style and you have examples.
Does a long context window mean the model uses every token?
No. The context window is a capacity limit, not a quality guarantee. Work such as Liu et al. (2023) found weaker use of information placed in the middle of long inputs. Put the evidence at the edges, or retrieve a short passage instead of stuffing the window.
What does MCP replace?
MCP replaces the habit of writing a one-off connector for every tool. It does not replace your orchestrator, your auth layer, or your evals. It is a protocol, specified as of 2026-07-28, for how a host talks to a server that exposes tools and data.
What is the difference between open-weight and open-source?
Open-weight means you can download the parameters. Open-source, strictly, also covers the license and enough code to use, modify, and redistribute. Training-data access is a separate question. A weight file without a license review is not a procurement decision.
Related reading
Depth on the terms that need more than a paragraph: RLHF, RAG in production, agent architecture, prompt engineering, context-window limits, the four-model comparison, open versus closed, and what the tools actually cost.