AI in Education Glossary of Terms

Definitions written for education leaders as a practical companion to the Arizona AI guidance. This glossary is not intended to be legal advice.

Find a term

Core AI concepts

Adaptive Learning: A technology-supported instructional approach that uses data and algorithms to adjust the pace, sequence, or support within learning activities based on student performance.
Agent Swarm: A group of AI agents designed to coordinate on a shared goal, often with different agents handling different roles or steps. Because agents may pass tasks or data between systems, swarms increase the need for careful permissions, logging, security, and human oversight.
Agentic AI: AI systems that can pursue a user-defined goal by planning steps, using tools, and taking actions on the user's behalf. Unlike a basic chatbot that mainly responds with content, an agentic system may browse the web, edit files, run code, retrieve information, organize data, or interact with digital platforms as part of a larger workflow.
Agentic Permissions: The access and actions an AI agent is allowed to use, such as reading a file, editing a document, browsing the web, accessing a calendar, running code, or sending a message. In school settings, agentic permissions should be narrow, transparent, reversible, and paired with human approval for sensitive actions.
AI Agent: An AI system designed to independently complete tasks through multiple steps, often by using connected tools such as web browsers, files, email, calendars, code environments, or data systems. An AI agent may act more like a digital assistant than a simple chatbot.
AI Companion: An AI system designed or used for social, emotional, or companion-like interaction. AI companions can feel personal because they use natural language, memory, and affirming responses, but they do not provide human care, confidentiality, or professional judgment.
AI Detector: A tool that claims to identify whether text, images, or other content were generated by AI. AI detectors are unreliable and should not be used as the sole basis for academic integrity decisions.
AI Literacy: The knowledge and skills needed to understand, use, evaluate, and make decisions about AI systems safely, effectively, and ethically. AI literacy includes technical understanding, ethical reasoning, critical evaluation, and human judgment.
AI Slop: Low-quality AI-generated content produced quickly or at scale. AI slop can clutter digital spaces, spread misinformation, reduce trust, and make it harder to identify accurate or meaningful information.
Algorithm: A set of rules or instructions that a computer follows to complete a task or solve a problem. In AI systems, algorithms help determine how data is processed and how outputs are generated.
Artificial General Intelligence (AGI) / General AI: A theoretical form of AI that could understand, learn, and apply intelligence flexibly across many domains. Current GenAI tools are not AGI; they are powerful but limited systems designed for specific types of tasks.
Artificial Intelligence (AI): Computer systems or algorithms designed to perform tasks that typically require human intelligence, such as recognizing patterns, making predictions, solving problems, generating content, or responding to language.
Big Data: Large, varied, or fast-moving data sets that are difficult to organize, analyze, or use effectively with traditional tools. In education, big data may include attendance, assessment, behavior, course, platform, or operational data.
Chatbot: A computer program designed to simulate conversation with human users. In education, chatbots may help answer questions, explain concepts, draft text, or support learning interactions.
Context: The information available to an AI system when it generates a response. Context may include the prompt, uploaded files, prior conversation, examples, user instructions, connected data, or tool access.
Context Rot: The decline in response quality that can happen when a conversation, prompt, or source set becomes too long, cluttered, or overloaded. The AI may start ignoring earlier instructions, blending unrelated details, or drifting away from the original task.
Context Window: The amount of information an AI system can consider at one time when generating a response, such as the current prompt, prior conversation, uploaded files, or retrieved source material. When the context window is exceeded, the system may miss, compress, or lose track of important details.
Deep Learning: A type of machine learning that uses layered neural networks to detect patterns in large amounts of data. Deep learning powers many modern AI systems, including many GenAI tools.
Generative AI (GenAI): AI that creates new content based on patterns learned from data. GenAI may produce text, images, audio, video, code, summaries, lesson materials, feedback, or other outputs. Generative AI is different from theoretical general AI, which would be capable of broad human-like intelligence across domains.
Hallucination: An AI-generated output that sounds plausible but is false, unsupported, fabricated, or misleading. Hallucinations can include incorrect facts, fake citations, invented examples, or unreliable summaries.
Jagged Frontier: The uneven boundary of AI capability where a system may perform impressively on some complex tasks while failing at tasks that seem simple to humans. The jagged frontier is a reminder that users must test AI output rather than assume capability based on confidence or fluency.
Large Language Model (LLM): A type of AI model trained on large amounts of language data to generate, summarize, translate, classify, or respond to text. Many chatbots are powered by LLMs.
Machine Learning: A branch of AI in which systems learn patterns from data rather than being explicitly programmed for every task. Machine learning has been used in education for prediction, recommendation, adaptive learning, and analytics.
Model Bias: Bias that comes from the way an AI model is designed, trained, optimized, or evaluated. Model bias can affect which outputs are favored, which perspectives are emphasized, and how the system responds to users.
Model Design: The technical and human choices that shape how an AI system behaves, including training methods, guardrails, optimization goals, safety rules, and system constraints.
Multimodal AI: AI that can work across multiple types of input or output, such as text, images, audio, video, documents, screenshots, and code.
Natural Language Processing (NLP): A branch of AI focused on enabling computers to process and respond to human language. In education, NLP can support chatbots, translation, writing feedback, accessibility tools, and language learning applications.
Reasoning Model: An AI model designed to spend more computational effort on complex or multi-step tasks before producing a response. Reasoning models may be better at analysis, planning, problem-solving, and structured explanations, though often at the expense of speed.
Retrieval: The process of finding relevant information from a source, such as a document collection, website, database, or knowledge base, for use in an AI-supported response or workflow.
Stochastic Parrot: A term used to describe language models that generate fluent output by predicting patterns in language without human understanding. The term highlights the risk of mistaking natural-sounding language for truth, comprehension, or judgment.
Synthetic Data: Data created artificially rather than collected directly from real-world events or people. Synthetic data may be generated by humans, simulations, or AI systems, and is increasingly used in AI development.
Synthetic Media: AI-generated or AI-altered media, including images, audio, video, or text. Synthetic media can be used creatively, but it can also mislead, impersonate, or harm people.
Tool Harness: The set of tools, integrations, permissions, and technical structures that allow an AI system to do more than generate a response. A tool harness may let AI browse the web, run code, access files, query data, send messages, or interact with apps.
Training Data: The data used to develop an AI model. Training data may include text, images, audio, video, code, or other information. The quality, representation, and source of training data strongly shape AI behavior and output.
Uploaded Exemplars: Samples, files, images, rubrics, drafts, or models provided to an AI system to guide its output. Exemplars can improve relevance and style matching, but users should avoid uploading confidential, copyrighted, or sensitive information into tools that have not been approved for that data.
WEIRD Data: Data that overrepresents Western, Educated, Industrialized, Rich, and Democratic populations or perspectives. When AI systems are trained on WEIRD-skewed data, outputs may underrepresent or misrepresent other communities and ways of knowing.

Teaching, learning, and student use

AI Integration: The intentional use of AI to strengthen teaching, learning, leadership, or operational workflows. Strong AI integration starts with purpose, learning goals, ethics, and human judgment rather than simply adding AI to an existing task.
AI Thought Partner: A use of AI in which the user treats the system as a sounding board for brainstorming, planning, questioning assumptions, or refining ideas. A thought-partner approach keeps the human responsible for context, judgment, values, and final decisions.
AI Tutor: An AI system used to explain concepts, ask practice questions, provide hints, or support learning through back-and-forth interaction. AI tutors may be useful when carefully designed and monitored, but they should not replace teacher expertise, relationship, or accountability.
Automated Essay Scoring (AES): The use of AI or automated systems to evaluate written work. AES may be useful for limited feedback or pattern analysis, but primary grading should rely on teacher judgment because automated scoring can be inconsistent, opaque, or misaligned with instructional goals.
Citation and Disclosure: Practices for acknowledging when and how GenAI contributed to a product or process. Formal citation may be appropriate for academic work, while a plain-language AI-use statement may be more useful for classroom transparency.
Cognitive Debt: The missed understanding or skill development that can accumulate when learners repeatedly outsource important thinking to AI instead of practicing it themselves.
Cognitive Offloading: Using a tool, routine, or system to reduce mental effort for part of a task. Cognitive offloading is not always harmful; calculators, notes, and calendars all offload thinking. The key question is whether offloading supports deeper learning or removes important productive struggle.
Evaluation of AI Outputs: The process of checking AI-generated content for accuracy, relevance, bias, completeness, source quality, accessibility, appropriateness, and alignment with local context. Evaluation is a core part of AI literacy because polished output can still be wrong or incomplete.
Feedback Partner: A use of AI in which the system provides comments, questions, revision suggestions, or alternative explanations. AI feedback should be treated as input to evaluate, not as a replacement for teacher judgment or peer and human interaction.
Instructional Misalignment: A mismatch between AI-generated materials and the intended learning objective, standards, student needs, or instructional approach. AI output can look polished while still weakening the learning design if it does not fit the context.
Media and Information Literacy: The ability to understand how information is created, packaged, shared, evaluated, and used. In the AI era, media and information literacy includes recognizing misinformation, synthetic media, algorithmic influence, source quality, and the role of cognitive bias.
Microtask: A smaller part of a larger learning or work task. In education, breaking assignments into microtasks can help teachers decide where AI support is appropriate and where students need to engage in human thinking, practice, or productive struggle.
Over-Reliance: Using AI so often or so broadly that human skill, judgment, learning, relationships, or institutional knowledge weakens. Over-reliance can occur when users treat AI output as correct by default or delegate whole tasks without understanding the work.
Performance-Based Task: An assessment or learning activity in which students demonstrate understanding through application, creation, explanation, presentation, performance, or problem-solving. In AI-integrated classrooms, performance-based tasks can make thinking visible and reduce overreliance on generated text.
Productive Challenge / Productive Struggle: Purposeful difficulty that supports learning through effort, reflection, revision, feedback, and persistence. With AI, teachers can design productive challenge by deciding which cognitive work students should practice and where AI may appropriately support or scaffold the task.
Productive Resistance: Intentional friction in a learning task that keeps students engaged in the thinking that matters. Productive resistance may include checkpoints, explanations of reasoning, defense of choices, revision cycles, or performance tasks that prevent AI from becoming an assignment easy button.
Prompt: The input, question, instruction, file, image, or request given to an AI system. A prompt gives the system direction, but the quality of the response also depends on context, model design, available tools, and user review.
Prompt Engineering: The practice of intentionally designing and refining inputs to an AI system to get more accurate, useful, and efficient outputs. Prompt engineers may follow any number of structured approaches or frameworks, but all share the same goal of better results through more deliberate communication with the AI tool.
Prompt Optimization: The process of revising a prompt to improve the usefulness, accuracy, clarity, or relevance of an AI system's response.
Prompt Template: A reusable prompt structure that users can copy, adapt, and customize to get a useful response without drafting the entire prompt from scratch.
Prompting Strategy: A structured approach for interacting with an AI system. Examples include assigning a role, stating a goal, providing context, defining expectations, asking clarifying questions, or using the AI as a thought partner.
Step-by-Step Prompting: A prompting approach that asks AI to break a task into visible steps, create a plan, or explain its reasoning at a high level. Students should still show their own thinking when that is the learning target.
Transparency: Clear communication about when AI is used, what role it played, what data it accessed, what limitations remain, and who is accountable for the final outcome. Transparency supports trust among students, staff, families, and communities.
Universal Design for Learning (UDL): A framework for designing learning experiences that provide multiple ways for students to access content, engage with learning, and demonstrate understanding. UDL can help guide thoughtful AI use for accessibility and learner variability.

Ethics, privacy, and well-being

AI-Enabled Misconduct: Misuse of AI to violate school expectations or harm others. Examples may include submitting undisclosed AI-generated work when it is not allowed, creating deepfakes, impersonating someone, forging communication, cyberbullying, or using AI to bypass security controls.
Anthropomorphism: The tendency to attribute human traits, emotions, intentions, or understanding to something non-human. With GenAI, users may feel that a chatbot "knows," "cares," or "understands," even though it does not experience thoughts or feelings. It is important to remember that a chatbot is a mathematical model, not a person.
Authorship: The act of creating a work and being accountable for its content, meaning, and choices. GenAI complicates authorship because a person may prompt, select, revise, and arrange AI-generated material while the tool produces some of the text, images, code, or media.
Bias: Uneven, incomplete, or unfair patterns in AI outputs or decisions. Bias may come from training data, model design, human choices, or how a tool is used.
Cultural Protocols: Community-specific expectations for access to, use of, and stewardship over stories, art, language, knowledge, and data. In Indigenous and tribal contexts, cultural protocols may apply even when materials are publicly visible online.
Data Bias: Bias that comes from the data used to train or operate an AI system. If the data reflects incomplete, inaccurate, or inequitable patterns, the AI output may reproduce or amplify those patterns.
Data Classification: The process of labeling data by sensitivity and risk, such as public, internal, confidential, or restricted. Data classification helps staff decide what information can be entered into approved AI tools and what must stay within protected systems.
Data Governance: The policies, roles, practices, and safeguards that determine how data is collected, stored, accessed, shared, protected, and deleted. GenAI implementation should strengthen data governance rather than work around it.
Data Privacy: The protection of personal, sensitive, or confidential information from inappropriate collection, use, disclosure, storage, or reuse. With GenAI, data privacy includes prompts, uploaded files, chat history, interaction logs, vendor practices, and model training policies.
Deepfake: AI-generated or AI-altered audio, image, or video that makes it appear as though a person said or did something they did not say or do. Deepfakes raise concerns related to misinformation, impersonation, bullying, consent, and trust.
False Positive: A result that incorrectly flags something as present when it is not. In schools, a common AI-related example is an AI detector incorrectly labeling human-written student work as AI-generated.
Human Agency: The ability and responsibility of people to make choices, exercise judgment, and remain accountable when using AI. Human agency means AI may support work or learning, but humans decide what is true, ethical, appropriate, and worth using.
Human Bias: Bias introduced through human decisions, assumptions, labels, priorities, or social patterns. Human bias can shape AI through the data people create, the labels people apply, and the design choices developers make.
Human in the Loop: A design or decision-making approach in which humans remain actively involved in reviewing, approving, correcting, or taking responsibility for AI-supported actions. In education, "human in the loop" should include real oversight, not just symbolic approval.
Human Oversight: The responsibility of people to review AI output, evaluate accuracy and appropriateness, protect privacy, consider ethical implications, and remain accountable for final decisions or actions.
Intellectual Property (IP): Legal and ethical rights related to creative works, inventions, designs, writing, images, media, data, and other forms of original expression. GenAI raises questions about ownership, authorship, attribution, copyright, training data, and cultural responsibility.
Non-Consensual Intimate Imagery (NCII): Intimate or sexualized images shared, created, or altered without the person's consent, including AI-generated or manipulated images. Schools should address AI-enabled NCII through prevention, reporting, support, discipline, and legal and privacy guidance.
Over-Permissioned Agent: An AI agent that has broader access or authority than needed for its task. Over-permissioned agents increase the risk of accidental exposure, unauthorized actions, or harmful behavior if the system is manipulated or compromised.
Parasocial Attachment: A one-sided emotional bond with a media figure, fictional character, or AI system. Because AI can respond personally and continuously, some users may develop comfort, loyalty, or dependency even though the system does not have human care or understanding.
Personally Identifiable Information (PII): Information that can identify a specific person, either directly or in combination with other data. In schools, PII may include student names, ID numbers, grades, attendance, disability status, discipline records, contact information, or other protected records.
Sycophancy: The tendency of an AI system to agree with, flatter, or validate a user, even when the user is wrong, harmful, or needs correction. Sycophancy can distort judgment and make weak ideas feel confirmed. This can create an echo chamber where a user's incorrect or biased ideas are validated instead of corrected.
Voice Cloning: The use of AI to imitate a person's voice. Voice cloning may support accessibility or media production, but it can also be used for impersonation, scams, cyberbullying, or social engineering.

Cybersecurity and operational workflows

3ssentials: Arizona AI implementation shorthand for three essential early actions: determine and communicate approved AI tools, train teachers and staff on top-line guardrails, and clarify expectations for GenAI use. The framework helps schools move from general discussion to concrete early action.
Adaptive Governance: A flexible governance approach that can evolve as AI tools, risks, laws, and local needs change. In schools, adaptive governance usually involves cross-functional leadership, regular feedback cycles, updated guidance, and decisions anchored in shared values rather than one-time rules.
AI-Enhanced Phishing and Impersonation: The use of GenAI to make scams, social engineering, or impersonation more convincing through polished language, local context, synthetic voice, images, or video. In schools, examples may include fake payroll requests, fraudulent invoices, scholarship scams, or messages that appear to come from trusted leaders.
AI Use Expectations: Shared norms that explain when, how, and why GenAI may be used. Clear expectations help students, staff, and families understand appropriate use, transparency, privacy, and the difference between support for learning and shortcutting the intended thinking.
Arrival Technology: A technology that arrives broadly into daily life and work, requiring people and institutions to adapt rather than simply choose whether to adopt a single tool. GenAI is framed as an arrival technology because it is reshaping work, information, decision-making, and learning.
Cross-Functional Leadership Team: A group that brings together people from instruction, curriculum, technology, operations, communications, privacy, legal, school sites, and other relevant roles to guide AI implementation. This structure helps prevent AI decisions from becoming isolated within one department.
Educational Data Mining (EDM): The process of analyzing data from educational settings to identify patterns, improve learning outcomes, optimize educational processes, and support institutional decision-making.
Guardrails: Policies, technical controls, expectations, training, and human review practices designed to reduce risk and guide responsible AI use. Guardrails do not eliminate risk, but they help users make safer and more consistent decisions.
Indirect Prompt Injection: A type of prompt injection in which hidden or malicious instructions are embedded in content an AI system processes, such as a webpage, email, PDF, shared document, or uploaded file. The AI may follow those instructions without the user realizing it.
Integrated Workflow: A process in which AI supports multiple steps within a larger project or operational task. For example, AI might gather information, summarize findings, draft communication, prepare a report, and route the work for human review.
Knowledge Assistant: An AI tool configured to answer questions or support work using a defined body of trusted information, such as district procedures, board policies, handbooks, curriculum documents, or help-desk knowledge. Knowledge assistants require careful source control, updates, permissions, and evaluation.
Least Privilege / Narrow Access: A security principle that gives users, systems, or AI agents only the minimum access needed to complete a task. For AI agents, narrow access may mean read-only permissions, limited folders, restricted tools, approval checkpoints, and no access to sensitive systems unless truly necessary.
Position Statement: A clear, locally developed statement that explains a school system's values and direction regarding GenAI. A strong position statement can guide decisions about tools, instruction, professional learning, transparency, privacy, and communication.
Prompt Injection: An attempt to manipulate an AI system by giving it instructions that override or conflict with the user's intent, system rules, or safety constraints. Prompt injection becomes especially risky when AI systems can access tools, files, or connected platforms.
Sandboxing / Isolation: A security practice that limits a tool or AI agent to a controlled environment so it cannot freely affect other systems. Sandboxing can reduce risk when testing new AI workflows, running code, or connecting agents to school data or platforms.
Social Engineering: A cyberattack strategy that manipulates people into revealing information, granting access, sending money, changing account details, or taking unsafe actions. GenAI can make social engineering more convincing by generating polished, personalized messages.
Vibe Coding: Using AI coding tools to rapidly create or modify software, scripts, dashboards, automations, or prototypes through natural-language instructions. Vibe coding can expand who is able to build tools, but it also increases the need for review, testing, security, documentation, and maintainability.
Workflow Automation: The use of technology to complete repeated steps in a process, such as routing information, generating reminders, summarizing forms, updating records, or creating reports. AI-enabled automation can improve efficiency, but it should include clear ownership, privacy protections, error checks, and human escalation paths.
Workflow Redesign: Rethinking a process around human strengths, local goals, and AI-supported capabilities rather than simply inserting AI into an old routine. Good workflow redesign asks what should be automated, what should be augmented, what should remain human, and how quality will be monitored.

Policy and acronym reference

Acceptable Use Policy (AUP) / Responsible Use Agreement (RUA): A policy or agreement that explains how students and staff may use school technology. For GenAI, it should address approved tools, age requirements, privacy, transparency, academic integrity, and AI-enabled misconduct.
Children's Internet Protection Act (CIPA): A federal law connected to internet safety protections for schools and libraries receiving certain federal funds. In GenAI contexts, leaders should consider how AI tools align with filtering, monitoring, and student safety expectations.
Children's Online Privacy Protection Rule (COPPA): A federal rule addressing the collection, use, and disclosure of personal information from children under 13 by online services. Schools should consider COPPA when reviewing AI tools that may be used by younger students.
Family Educational Rights and Privacy Act (FERPA): A federal law protecting the privacy of student education records. In GenAI contexts, FERPA is relevant when prompts, uploads, integrations, or vendor systems may involve student records or personally identifiable information.
Protection of Pupil Rights Amendment (PPRA): A federal law related to certain surveys, analyses, evaluations, and student information in protected areas. In GenAI contexts, schools should be mindful of tools or activities that ask students to disclose sensitive personal information.
Web Content Accessibility Guidelines (WCAG): Technical standards used to evaluate and improve the accessibility of digital content. School-provided digital materials, including AI-generated materials, should be reviewed against applicable accessibility expectations such as WCAG Level AA standards.

A term missing, or a definition that has aged badly?

This vocabulary changes faster than almost anything else in the guidance. Tell us at a Learning Network session and it gets fixed.

Join the Learning Network