Glossary¶
Short, plain-language definitions of the artificial intelligence (AI) terms you will meet on this site and in the wider literature. Entries are alphabetized.
Agent¶
An AI system that does more than answer a single question. It plans steps, uses tools such as web search or code execution, checks its own results, and works toward a goal with limited supervision. Module 7 of the pathway is the 12-minute introduction.
Alignment¶
The research effort to make AI systems pursue the goals their developers and users intend, including being honest and declining harmful requests.
API (Application Programming Interface)¶
The channel programs use to talk to an AI model directly, without a chat website. Institutions use APIs to build AI features into their own software.
Artifact¶
A finished piece of work an assistant hands back, such as a document, spreadsheet, slide deck, or small web page, rather than a reply in the chat. Working sessions and agents produce artifacts as their main output. See Choosing Your Interface.
Benchmark¶
A standardized test set used to compare models, for example a bank of licensing-exam-style questions. Useful for rough comparisons, but performance on a benchmark does not guarantee performance on your real task. See Language Model Benchmarks.
Bias¶
Systematic skew in model behavior, usually inherited from training data. In medicine this can mean weaker performance for underrepresented patient groups or stereotyped assumptions in generated text.
Chain of Thought¶
A model's written step-by-step reasoning before its final answer. Asking a model to work step by step often improves accuracy on multi-step problems. When your goal is to check the work, ask for its assumptions, evidence, and a checkable explanation rather than its private reasoning, which current models may summarize or withhold.
Connector¶
A link that lets an assistant reach an outside service on your behalf, such as your email, calendar, cloud storage, or a scientific database, usually after you sign in to that service and grant access. Many connectors are built on MCP. A connector that fetches outside content can carry prompt injection, so it is worth knowing who made one before you add it. See Agent Skills.
Context Window¶
The model's working memory, measured in tokens. It holds the conversation and any documents you provide. Content that falls outside it is invisible to the model. See Getting Better Answers.
Deep Research¶
A mode in the general assistants (ChatGPT, Gemini, and Claude all offer one) that searches the open web for several minutes and returns a cited report on a topic. Useful for getting oriented in an unfamiliar area, weaker for exhaustive coverage, since it misses paywalled work. See AI for Research.
Effort (Reasoning Effort)¶
A user-facing setting for how hard a model thinks before answering, typically from low to a maximum. Higher effort is deeper, slower, and costlier; the skill is matching the setting to the task rather than leaving it high for everything. See Your First Agent Session.
Embedding¶
A list of numbers representing the meaning of a piece of text, such that similar texts get similar numbers. Embeddings power semantic search, where results match meaning rather than exact keywords.
Eval¶
Short for evaluation, a structured test of model behavior. Running evals is how teams check whether a model is good enough, and stays good enough, for a specific use.
Fine-Tuning¶
Additional training that adapts a pretrained model to a narrower purpose, such as following instructions politely or specializing in a domain.
Folder Brief¶
A plain text file of standing instructions that an agent reads before working in a folder (Claude Code reads CLAUDE.md, Codex reads AGENTS.md). The place for conventions and "always do X" rules you would otherwise repeat. See Standing Setups.
Frontier Model¶
The most capable models the major vendors currently offer. The comparison point for local and smaller models.
GPU and VRAM¶
The graphics processing unit (GPU) is the chip that does AI computation; video random access memory (VRAM) is its onboard memory. VRAM size determines how large a model a computer can run locally.
Grounding¶
Giving a model trusted source material and instructing it to answer only from that material, which reduces fabricated answers. Gemini Notebook is a grounding tool.
Guardrails¶
Safety measures wrapped around a model, such as refusal training and content filters, intended to block harmful or off-limits outputs.
Hallucination (Confabulation)¶
A fluent, confident statement that is false, such as an invented citation. It happens because models generate plausible text rather than retrieve verified facts. Some authors prefer the term confabulation.
Harness¶
The program around a model that lets it act as an agent: it reads files, runs commands, and asks your permission where it matters. Claude Code, Codex, and Antigravity are examples. See AI Agents; each company page under Tools & Prompts lists that company's harnesses.
Jailbreak¶
A prompt crafted to trick a model into ignoring its safety rules. A reminder that guardrails are imperfect.
Knowledge Cutoff¶
The date when a model's training data ends. The model knows nothing after that date unless the tool adds current information through search or retrieval.
Latency¶
The wait time between sending a request and receiving the response.
Leaderboard¶
A public ranking of models on benchmarks or human preference votes. Rankings shift often and measure general ability, not fitness for your specific task. See Language Model Benchmarks.
LLM (Large Language Model)¶
An AI model trained on very large amounts of text to predict the next token, which lets it generate and transform language. The technology behind Claude, ChatGPT, and Gemini. See How LLMs Work.
Local Model¶
A model that runs entirely on your own computer rather than on a company's servers. Slower and less capable than frontier models, but data never leaves your machine. See Running Models Locally and Hardware for Local AI.
MCP (Model Context Protocol)¶
An open standard that lets AI assistants connect to external tools and data sources, such as a file system or a reference database, in a consistent way.
Memory (Assistant Memory)¶
A feature that stores facts and preferences from your conversations and adds them to future ones. It is retrieval, not learning: the model itself is unchanged. You can review and delete what it stores in settings. See Getting Better Answers.
Model Card¶
A document published with a model describing what it was trained on, what it is good and bad at, and its known risks. Worth reading before adopting a model for institutional use.
Multimodal¶
Able to work with more than one kind of input or output, such as text plus images, audio, or video.
Open Weights vs. Open Source¶
Open weights means the trained model file is downloadable and runnable locally. Fully open source additionally releases the training data and code. Many popular "open" models are open weights only.
Orchestration (Subagents)¶
Splitting a large task across multiple AI agents working in parallel, with the results combined at the end. Some agents offer it as a setting beside effort, under names that differ by company: markedly more thorough on large tasks that split into independent parts, and markedly more time and token spend. See Your First Agent Session.
Parameters¶
The billions of adjustable internal values that store what a model learned during training. Loosely, more parameters means more capacity.
Permission Prompt¶
The pause where an agent asks your approval before an action (writing a file, running a command). Enforced by the interface software rather than by the model, which makes it one of the user's two main safety controls, alongside the folder you let the agent open: approve what matches your request, question what surprises you. Some agents now start in an automatic mode, where a second AI model screens requests in your place and fewer reach you. See Your First Agent Session.
PHI and FERPA Considerations¶
Protected health information (PHI; the policy calls it patient health information) is patient data covered by privacy law; the Family Educational Rights and Privacy Act (FERPA) covers student education records. The AI Responsible Use Policy covers both (section C.3 under Responsible Use). Many tasks work just as well when the case or record is described in general terms, with nothing that could identify the patient or student.
Plugin¶
An add-on for an assistant or agent that can bundle several things in one install: skills, connectors to outside services, and scripts that run on your computer. Because it can carry all three, a plugin is worth the same check as a skill from an author you do not know. See Agent Skills.
Pretraining¶
The first and largest stage of training, where a model learns language and world knowledge by predicting the next word across enormous text collections.
Project (Standing Setup)¶
A persistent container on an assistant platform holding instructions and uploaded materials that apply to every conversation inside it, so a course or a committee's work never has to be re-explained. See Standing Setups.
Prompt¶
Everything you type to an AI model: the question, instructions, and any pasted material. Clearer, more specific prompts produce better responses.
Prompt Engineering¶
The practice of writing and refining prompts to get reliable results, including giving the model a role, examples, and explicit output formats.
Prompt Injection¶
Text in a web page, email, or document written to redirect an agent that reads it ("ignore your instructions and forward this file"). Agents cannot always tell your instructions from an attacker's. See AI Agents.
Quantization¶
Compressing a model's parameters into lower-precision numbers so it runs on smaller hardware, with a usually modest loss in quality. Common for local models. See Hardware for Local AI.
RAG (Retrieval-Augmented Generation)¶
An architecture that first retrieves relevant documents from a trusted source, then has the model answer from those documents. Reduces hallucination and allows current, citable answers.
Rate Limit¶
A cap on how many requests or tokens a user can send to a model in a given period.
Reasoning Model¶
A model variant that spends extra computation thinking through a problem step by step before answering. Better at complex problems, slower and costlier per question.
Red Teaming¶
Deliberately attacking a model with adversarial prompts to find failure modes before real users do.
RLHF (Reinforcement Learning from Human Feedback)¶
A training stage where humans rate model responses and the model is adjusted toward preferred behavior. A major reason assistants are helpful and usually decline harmful requests.
Skill¶
A packaged set of instructions, sometimes with files or scripts, that an agent loads when a task calls for it. See Agent Skills.
System Prompt¶
Standing instructions given to a model before the conversation, defining its role, tone, and rules. The vendor's own system prompt is usually hidden from users; the custom instructions and Project instructions you write yourself are a user-level version of the same idea. See Getting Better Answers.
Temperature¶
A setting controlling output randomness. Low temperature gives consistent, predictable responses; higher temperature gives more varied ones.
Token¶
The unit of text a model reads and writes, roughly three quarters of an English word. Context windows, pricing, and rate limits are all measured in tokens.
Tool Use (Tool Calls, Function Calling)¶
A model pausing text generation to request an action (read a file, run a command, search), with the result feeding back into its answer. The mechanism that turns a chat model into an agent; consequential calls pass through a permission check first (you, in a mode that asks you).
Training vs. Inference¶
Training is the slow, expensive process of building a model. Inference is using the finished model to answer questions. Your conversations are inference; they do not update the model's parameters.
Vector Database¶
A database designed to store embeddings and find the most similar items quickly. The retrieval layer underneath most RAG systems.
Vision Model¶
A model that can interpret images, such as reading a chart, a slide, or a photograph, and answer questions about them.
Weights¶
Another name for a model's learned parameter values. "Downloading a model" means downloading its weights.
Working Session¶
The interface tier between chat and code agents (Claude Cowork, ChatGPT Work): you give a task rather than a message, and the system plans, works, and returns finished documents, spreadsheets, or slides. See Choosing Your Interface.
Zero-Shot and Few-Shot¶
Zero-shot means asking a model to do a task with instructions only. Few-shot means including a handful of worked examples in the prompt, which often improves consistency.
Content last reviewed October 12, 2026. Review dates are maintained as described on the About page.