How to Use This Glossary
Plain-language definitions of the terms you will run into on this site, in vendor marketing, and in conversations about AI at the college. Definitions come first in everyday language; where the technical meaning differs from common usage, that is noted.
This glossary is linked from the footer of every page, so you can get to it from wherever you are. If a term you need is missing, tell the ARC and it will be added.
Terms
Agent / Agentic AI
An AI system that takes actions rather than only producing text — searching, filling in forms, moving data between systems, sending things. The distinction matters because a wrong action is harder to catch and undo than a wrong paragraph.
AI (Artificial Intelligence)
An umbrella term, not a specific technology. It covers everything from spam filters to self-driving cars. On this site it usually means generative AI unless stated otherwise. Because the term is so broad, “we use AI” tells you almost nothing on its own.
Algorithm
A set of steps for accomplishing something. Ordinary and not specific to AI — a recipe is an algorithm. Worth knowing because “the algorithm” is often used to imply something more mysterious than it is.
Alt text (alternative text)
A short written description of an image, read aloud by screen readers. Good alt text describes what the image is doing on the page, not just what is pictured. See Accessibility.
Bias
In AI, patterns in the training data that show up as skew in the output — whose perspective is treated as default, which groups are described thinly or stereotypically, which viewpoints are missing. Not a matter of intent; it is a reflection of what was in the source material. See Critical Evaluation Skills.
Chatbot
A conversational interface to an AI model. The chat window is not the intelligence — it is the way you reach the model behind it.
Context window
How much text a model can consider at once — your question, the documents you pasted, and the conversation so far. When a long conversation starts “forgetting” earlier details, the context window is usually why.
Deepfake
Synthetic audio, image, or video made to look like a real person. Relevant here mainly as a literacy issue: students and employees should know that convincing fake media is now easy to produce.
Fine-tuning
Additional training of an existing model on specialized material to make it better at a particular kind of task. Distinct from prompting, which changes nothing about the model.
FERPA
The federal law protecting the privacy of student education records. Directly relevant to AI: putting identifiable student information into an AI tool means disclosing it to a third party. Treat AI tools as outside parties. See Data Privacy and Security.
Generative AI
AI that produces new content — text, images, audio, code — rather than classifying or sorting existing content. This is what most people mean by “AI” now, and it is what this site is mostly about.
Grounding
Tying a model’s answers to specific source documents you provide, rather than letting it draw on general training. Grounded tools can cite where each claim came from, which makes verification much faster and reduces fabrication.
Hallucination
When an AI produces something false while presenting it confidently — an invented citation, a fabricated statistic, a made-up quotation. The term is somewhat misleading, since it suggests a malfunction; the system is doing exactly what it was built to do, which is produce plausible text. Plausibility and truth are simply not the same thing. This is the single most important term on this page.
Inference
The act of a trained model producing an answer. You will mostly see this in discussions of cost and speed — “inference costs” means the expense of running the model, not of building it.
Knowledge cutoff
The date after which a model knows nothing, because its training data ends there. A model will usually answer questions about later events anyway rather than saying it cannot. Always relevant for policy, technology, and current events.
LLM (Large Language Model)
The kind of system behind most AI chat tools. It was trained on very large amounts of text and works by predicting what text should come next. That mechanism explains most of its behavior — including why it invents citations and why it rarely says “I don’t know.”
Local model / on-device AI
A model that runs on your own computer or on institutional hardware rather than sending your input to a vendor. Generally less capable than large hosted models, but the data does not leave the machine — which changes what is possible with sensitive material. See Emerging Approaches.
Machine learning
Systems that improve at a task by processing examples rather than by being explicitly programmed. The broader field that generative AI sits inside.
Multimodal
A model that handles more than one kind of input or output — text plus images, audio, or video. A tool you can show a diagram to, or that can transcribe a recording, is multimodal.
Neural network
The general architecture underlying modern AI, loosely inspired by brain structure. The “loosely” is doing considerable work — the comparison is a naming convention, not a claim that these systems think.
Open weights / open source
Models whose internal parameters are published, so anyone can run or modify them. Often described as “open source,” though that is not quite accurate — the training data usually is not released. Relevant to institutions considering running models themselves.
Parameters
The internal values a model adjusts during training, often quoted in billions. More parameters generally means more capability and higher cost to run, but the number alone is a poor guide to whether a model is good at your task.
Prompt
What you type. That is all it means. The surrounding mystique is mostly marketing.
Prompt engineering
The practice of writing prompts that get better results. Genuinely useful and considerably more mundane than the name implies — it is mostly being specific about audience, format, and what to avoid. See Foundational Skills.
RAG (Retrieval-Augmented Generation)
An approach where the system looks up relevant documents first and then answers using them. This is what makes source-grounded tools more reliable, and why they can cite where an answer came from. You do not need the acronym to use such tools, but you will see it in vendor materials.
Synthetic data
Data generated by AI rather than collected from the world. Sometimes used to protect privacy, sometimes to fill gaps. Worth knowing because synthetic data can carry over the biases of whatever generated it.
System prompt
Standing instructions given to a model before your conversation begins, setting its role and constraints. You usually cannot see it. It explains why the same underlying model behaves differently in different products.
Temperature
A setting controlling how varied the output is. Lower means more predictable and repetitive; higher means more varied and more prone to error. Some tools expose it; most do not.
Token
The chunks of text a model actually processes — roughly a short word or part of a word. Relevant because pricing and limits are measured in tokens rather than words. A rough guide: 1,000 tokens is about 750 words of English.
Training data
The material a model learned from. Almost always enormous and rarely fully disclosed. It determines what the model knows, what it is biased toward, and — in the case of image generation — whose work it absorbed. See Discipline-Specific Skills.
WEIRD data
Training data drawn primarily from Western, Educated, Industrialized, Rich, and Democratic (WEIRD) countries and cultures. Because this represents only a portion of humanity, training data skewed towards it can lead AI output to favor WEIRD concepts, ideas, and values over those cultures that aren't as well represented.
Watermarking
Techniques for marking AI-generated content so it can be identified later. An active area of work, not currently reliable. Do not assume unmarked content is human-made, or that marked content will stay marked.
These categories come from the district’s AI guiding principles, and they are more useful than the single word “AI” — particularly when deciding whether a tool is appropriate for a given task. They are arranged roughly by how much the system does on its own.
Passive AI
Produces output only in response to a direct query. Search engines, basic chatbots.
Tool-based AI
Creates content or performs a task from your prompt. Generative AI sits in this category — text, images, code, audio, or video produced from patterns in training data.
Assistive AI
Offers suggestions and recommendations from surrounding context, to support a human decision. Writing assistants, tutoring systems, recommendation engines. In clinical and professional settings this also covers purpose-built systems that support an expert’s judgment while the expert retains the decision.
Agentic AI (emerging)
Works toward a goal by planning and carrying out steps with minimal human direction. Workflow automation, adaptive learning platforms that adjust without being prompted.
Autonomous AI (emerging)
Makes independent decisions in complex environments with minimal or no human oversight. Self-driving vehicles.
Why the distinction matters: a general-purpose public tool and a purpose-built system approved for a specific use are not interchangeable, even when both are “AI.” Knowing which you are working with determines what data is appropriate to put in and how much weight the output can carry. See Choosing the Right Tool.
Bias mitigation
The ongoing work of identifying, reducing, and preventing unfair or discriminatory outcomes in AI systems — monitoring, testing, and adjusting rather than a one-time check.
Equity
Ensuring everyone has access to what they need to accomplish their goals, by actively removing barriers such as cost or technology access. Equity recognizes that fairness sometimes means providing different levels of support to different people; it does not mean identical resources for everyone.
Human-centered
An approach that prioritizes human needs, relationships, and judgment — AI designed to support and enhance the human elements of teaching, learning, and service rather than replace them. Related to HAI at Stanford University.
PII (Personally Identifiable Information)
Any information that can identify a specific person: names, Social Security numbers, email addresses, student ID numbers, photos. Protecting PII is central to using AI tools appropriately.
Transparency
Openness about when, how, and why AI is being used — telling students when AI is part of their learning experience, disclosing to colleagues when AI assisted with a work product, and being honest about both capabilities and limitations.
ARC (AI Resource Center)
The district-wide body coordinating AI guidance, resources, and professional development across Maricopa. It maintains this site. See About Us.
AITF (Artificial Intelligence Task Force)
The district task force whose work preceded and led to the establishment of the ARC. See History.
Approved AI Tools
AI applications reviewed and authorized for use in the district against criteria including data security, privacy protection, accessibility standards, and alignment with educational goals. “Approved” means vetted for a purpose, not simply available. See Choosing the Right Tool.
Domain Groups
The working groups carrying out ARC’s work in specific areas — strategy and governance, teaching and learning, and others. See Governance and Contributors.
GUIDE framework
How Maricopa’s AI guiding principles are organized, so they are easier to remember: Governed by Trust, Universal Access, Information Stewardship, Data Privacy and Security, Empowerment. See Guidelines and Safety.
Principled Innovation
Maricopa’s broader framing for adopting new technology: imagining and building new things guided by principles that create positive change.
Shared governance
Engaging internal constituencies in developing plans, policies, and procedures. In AI adoption it means diverse perspectives inform how tools are selected, implemented, and evaluated across the district.