Written and maintained by CASRAI Editorial Board
Last updated
People searching for a plain-English explanation of AI vocabulary rarely type the same phrase twice. Some look for “AI terminology,” some for “AI terms,” some for “AI acronyms,” and some for “AI glossary” or “artificial intelligence glossary.” They almost always mean the same thing: I keep running into words and abbreviations in AI news, vendor pitches, and policy documents, and I need to know what they actually mean before I can do my job. This guide is built for that person — a compliance officer, governance lead, procurement reviewer, or manager who is not trying to become a machine-learning engineer, just trying to stop nodding along in meetings.
It covers two layers. First, the everyday AI and machine-learning vocabulary you’ll meet in almost any AI conversation — model, training, inference, hallucination, fine-tuning, and the rest. Second, a decoded list of the acronyms that show up specifically in AI governance and compliance work — RSP, FSF, NIST AI RMF, GPAI, CAISI, AISI, and others — each one linked to a full explainer if you need more than a one-line definition.
Common AI and machine-learning terms, in plain English
These are the words that get used loosely in almost every AI conversation, often without anyone stopping to define them. Here’s what each one actually means.
| Term | What it means |
|---|---|
| Artificial intelligence (AI) | Software that performs tasks — recognizing patterns, generating text or images, making predictions — that would otherwise require human judgment. It’s an umbrella term, not one specific technology. |
| Machine learning (ML) | A subset of AI in which a system improves at a task by learning patterns from data, rather than following rules a person wrote by hand. Almost all AI systems discussed in governance and compliance contexts today are built on machine learning. |
| Model | The trained mathematical system that takes an input (text, an image, a number) and produces an output (a prediction, a classification, generated text). “The model” is what people usually mean when they say “the AI.” |
| Training | The process of exposing a model to large amounts of data so it learns the patterns needed to perform its task. Training is where a model’s internal parameters are set; it typically happens once, before the model is deployed. |
| Training data | The dataset a model learns from during training. Where that data came from — and whether the developer had the rights to use it — is one of the most common compliance questions raised about a given model. |
| Fine-tuning | Taking an already-trained model and training it further, usually on a smaller, more specific dataset, to adapt it to a particular task or domain (for example, fine-tuning a general model on an organization’s own documents). |
| Inference | The step where a trained model is actually used — you give it an input and it produces an output. If “training” is teaching the model, “inference” is the model answering a question. This is the stage most day-to-day AI use happens at. |
| Parameter | An internal value a model learns during training and uses to produce its outputs. “A 70-billion-parameter model” is a rough proxy for size and, loosely, capability — though parameter count alone tells you little about safety or reliability. |
| Prompt | The input — usually text — a person or system gives a generative AI model to get a response. “Prompt engineering” is the practice of writing prompts deliberately to get more reliable or accurate output. |
| Generative AI | AI systems designed to produce new content — text, images, audio, code — rather than just classify or score existing content. This is the category most current AI governance policy is written to address. |
| Large language model (LLM) | A type of generative AI model trained on very large amounts of text, used to understand and generate human language. ChatGPT, Claude, and Gemini are all built on LLMs. |
| Foundation model | A large model trained on broad data that can be adapted (often by fine-tuning) to many different downstream tasks, rather than being built for one narrow purpose. Most current LLMs are foundation models. |
| Multimodal model | A model that can take in or produce more than one type of content — for example, text and images together — rather than being limited to a single format. |
| AI agent | An AI system set up to take actions toward a goal with limited human intervention — not just answering a question, but calling tools, browsing, or executing multi-step tasks on its own. This is a fast-moving area of both product development and governance concern. |
| Context window | The amount of text a model can “see” and take into account at once when generating a response, usually measured in tokens. |
| Token | A chunk of text (often a word or part of a word) that a language model processes as a unit. Model pricing, context-window size, and processing limits are usually expressed in tokens rather than words or characters. |
| Hallucination | When a generative AI model produces output that is factually wrong or entirely fabricated, but stated confidently enough to sound plausible. This is one of the most consequential terms for a compliance audience, since it’s the direct source of real-world liability and misinformation risk from AI outputs. |
| Bias (algorithmic bias) | Systematic, unfair skew in a model’s outputs, usually traceable to skew in its training data or design choices. This is a distinct concept from hallucination: a biased model can be factually consistent and still produce unfair outcomes. |
| Alignment | The general effort to make an AI system’s behavior match human intentions and values, rather than optimizing for its stated objective in unintended or harmful ways. In frontier AI safety specifically, this term has a more technical, load-bearing meaning — see the dedicated explainer linked below. |
| Guardrails | Technical or policy controls placed around a model to constrain what it will do or say — content filters, usage restrictions, or automated checks on outputs before they reach a user. |
| Red teaming | Deliberately trying to make a model misbehave, produce harmful output, or bypass its guardrails, in order to find and fix weaknesses before real users encounter them. |
| Benchmark / evaluation | A standardized test used to measure a model’s performance or behavior on a specific task, so that different models — or different versions of the same model — can be compared on consistent terms. |
| Retrieval-augmented generation (RAG) | A technique where a model retrieves relevant information from an external source (a document set, a database) at the time of a query, then uses that retrieved content to generate its answer — rather than relying only on what it learned during training. |
| Explainability / interpretability | The degree to which a model’s decision-making process can be understood or explained, rather than treated as an opaque “black box.” This matters directly for audit and accountability requirements in regulated sectors. |
| Synthetic data | Artificially generated data, produced by a model or algorithm rather than collected from real-world events, sometimes used to train or test other models when real data is scarce, sensitive, or expensive to obtain. |
| Deepfake | Synthetic audio, video, or images generated to convincingly depict a real person doing or saying something they did not, produced using generative AI techniques. Increasingly the subject of specific disclosure and labeling laws. |
| AGI (artificial general intelligence) | A hypothetical future AI system capable of performing at or above human level across essentially any intellectual task, rather than being limited to the narrow tasks today’s models are built for. No current system meets any widely agreed definition of AGI, but the term shows up constantly in frontier-lab governance commitments and policy debate. |
AI acronyms decoded
The general terms above get you through most AI conversations. The acronyms below are the ones that show up specifically once AI governance, compliance, and regulation enter the conversation — and they multiply fast, because every regulator, standards body, and lab tends to name its own framework. Each entry below links to a full explainer if you need the detail behind the one-line definition.
| Acronym | What it stands for |
|---|---|
| RSP | Responsible Scaling Policy — a frontier AI developer’s own published commitment to test its models against defined risk thresholds and apply stronger safeguards before crossing one. The term originates with Anthropic. |
| FSF | Frontier Safety Framework — Google DeepMind’s version of a responsible-scaling-style policy. See how it and other labs’ frameworks compare in Frontier AI Labs: Who They Are and What Safety Frameworks They Publish. |
| NIST AI RMF | The National Institute of Standards and Technology’s AI Risk Management Framework — a voluntary U.S. framework built around four functions (Govern, Map, Measure, Manage), with a dedicated Generative AI Profile addendum. |
| GPAI | General-purpose AI — the EU AI Act’s term for broadly capable models like large language models, as opposed to narrow, task-specific AI systems. See the GPAI Code of Practice explainer. |
| CAISI | The U.S. Center for AI Standards and Innovation — the renamed successor (June 2025) to the U.S. AI Safety Institute, focused on voluntary standards and evaluations rather than binding rules. |
| AISI | AI Safety Institute (or, in the UK since February 2025, AI Security Institute) — a national government body that evaluates advanced AI systems. Several countries’ institutes coordinate through what is now called NAAIMES. |
| ISO/IEC 42001 | The international standard for an AI management system — the AI-specific equivalent of an ISO 9001 quality management system. See the certification guide for scope and timeline. |
| EU AI Act | The European Union’s binding AI regulation, which sets risk-tiered obligations for AI systems sold or used in the EU. See AI Regulations Around the World for how it compares to other jurisdictions’ approaches. |
| NLP | Natural language processing — the broader field of AI concerned with understanding and generating human language, of which today’s LLMs are one (very successful) approach. |
| API | Application programming interface — the standard way most organizations actually access a commercial AI model: sending it requests programmatically rather than through a chat interface. |
Why AI governance uses so much of this vocabulary at once
If you’re reading this guide because of a policy review, a vendor questionnaire, or a board briefing rather than general curiosity, the reason the vocabulary above feels dense all at once is structural: AI governance sits at the intersection of the technical terms in the first table (what the system actually does) and the regulatory acronyms in the second (who says it has to do that safely). A risk assessment, an AI usage policy, or an incident-reporting process will typically use both kinds of term in the same paragraph — describing what a model does technically, and which framework or law governs how that has to be handled. If you’re building out that kind of program from scratch, the AI Governance Framework Template and AI Risk Assessment Framework and Risk Register guides put this vocabulary to work end to end, and AI Regulations Around the World maps which of the acronyms above actually carries legal force in which jurisdiction.
Looking for frontier-AI-safety vocabulary specifically?
The two glossaries above are deliberately general — they cover the AI and AI-governance vocabulary you’ll meet across almost any organization’s AI work. If what actually brought you here is more specific — capability thresholds, safeguards, incident reporting, or who inside an organization is accountable for a safety commitment — that is a narrower, more specialized vocabulary than this page covers. CASRAI’s own NIKOLAI project is an independent dictionary of 64 frontier-AI-safety terms organized across 10 tracks, built specifically for that layer: thresholds and checkpoints, mitigations and security, incident reporting, and accountability roles among them. It is CASRAI’s own reference work, not an industry standard and not endorsed by any lab or regulator — every crosswalk between a NIKOLAI term and how a specific organization actually uses it is a “shadow mapping” unless that organization has filed its own Mapping Declaration to confirm it. See What Is NIKOLAI? for the full picture, or go straight to a specific track, such as Incidents or Commitments and Governance.
Frequently asked questions
What is the difference between AI terminology, AI terms, and AI jargon?
In practice, none. “AI terminology,” “AI terms,” and “AI jargon” are all ways of describing the same thing — the specialized vocabulary used to discuss artificial intelligence systems, from basic concepts like “model” and “training” to more technical or field-specific terms. People searching for any of these phrases are almost always looking for the same kind of plain-English glossary.
What are the most common AI acronyms?
The acronyms people encounter most often split into two groups: everyday technical acronyms like AI, ML, LLM, NLP, and AGI, and AI-governance acronyms like NIST AI RMF, GPAI, CAISI, AISI, and ISO/IEC 42001. The second group is where most confusion comes from, because each regulator or standards body names its own framework differently. Both groups are decoded in the tables above.
Is there a single official AI glossary?
No single glossary is universally authoritative. NIST maintains its own AI terminology as part of the AI Risk Management Framework, the OECD and ISO each publish their own AI-related vocabulary standards, and individual labs and vendors sometimes use terms slightly differently from one another. This guide draws on those sources to give consistent, plain-English definitions rather than pointing to one single “official” list.
What is an AI acronym that specifically relates to AI safety, not just AI generally?
Acronyms like RSP (Responsible Scaling Policy) and FSF (Frontier Safety Framework) are specific to frontier AI safety — the practice of major AI labs testing their most capable models against defined risk thresholds before release. That vocabulary goes considerably deeper than general AI terminology; CASRAI’s NIKOLAI project, discussed above, is built specifically to catalog it.
Do I need to know all of these terms to write an AI policy?
No. Most organizational AI policies lean on a small working subset — typically model, training data, hallucination, bias, and whichever governance acronyms apply to your jurisdiction (for example GPAI and the EU AI Act, or NIST AI RMF in the U.S.). The full glossary above is a reference to return to as new terms come up, not a list to memorize up front.
