Terminology and Reference Guide

Academic Building One

One of Virginia Tech’s graduate-studies focused buildings in the greater Washington, D.C., area. Located in Alexandria, Va., this new academic building focuses on technology, innovation, and graduate education. It is home to several research centers and institutes, including graduate programs in computer science.

Advanced Research Computing LLM Gateway

Virginia Tech’s university-managed AI platform, at llm.arc.vt.edu. Provides access to multiple large language models through a consistent interface, with institutional data protections and no third-party training on user inputs. This platform is operated by Advanced Research Computing within the Division of Information Technology.

AI Working Committee

Reporting through the IT Governance structure , this standing committee was established by the Division of Information Technology following the recommendation of the University AI Working Group. This committee provides ongoing recommendations on AI policy, investment, best practices, and risk management.

Division of Information Technology

As part of University Operations, the Division of Information Technology at Virginia Tech leads the university’s IT strategic planning and provides the enterprise applications, network infrastructure, high-performance computing resources, cloud services, collaboration platforms, software, services, and support that are pivotal to the delivery and security of AI platforms in use across Virginia Tech.

National Security Institute

A Virginia Tech research institute, part of the Hume Center for National Security and Technology, focused on national security challenges, including applications of AI and autonomous systems to defense, intelligence, and cybersecurity challenges. Located at Academic Building One in Alexandria.

Responsible and Ethical AI Framework for Virginia Tech

The comprehensive framework produced by the University AI Working Group in September 2025. Includes the seven principles, governance recommendations, implementation guidance, and a policy gap analysis identifying where existing university policies did not adequately address AI.

Sanghani Center for Artificial Intelligence and Data Analytics

Virginia Tech’s primary research center for AI and data science, housed within the College of Engineering. The center conducts foundational and applied AI research and offers graduate programs in machine learning, data analytics, and related fields.

Teaching and Learning Hub

The Center for Excellence in Teaching and Learning’s pedagogy resource guide, at teaching.vt.edu/hub . The hub provides faculty resources on course design, inclusive pedagogy, and instructional technology, including AI-related guidance .

Technology-enhanced Learning and Online Strategies

Technology-enhanced Learning and Online Strategies (TLOS), a unit of the Office of the Executive Vice President and Provost, improves student learning by fostering digital fluency, partnering with faculty to design and develop successful digital learning experiences, and creating technology-enhanced learning environments for flexible and online learning.

Virginia Tech AI Literacy Guidelines

The AI Literacy Guidelines is a five-dimensional framework based on the Digital Education Council’s AI Literacy Framework. The five dimensions, or competency areas, Understanding AI and Data, Critical Thinking and Judgment, Ethical and Responsible Use, Human-Centricity and Creativity, and Domain Expertise, are divided into three progressive development levels: Awareness, Application, and Leadership.

Virginia Tech Responsible and Ethical AI Principles

The seven principles guiding Virginia Tech’s approach to AI , that serve as one component of the broader Responsible and Ethical AI Framework. When referring to them, use the full name rather than shortening to “AI Principles,” which could refer to any institution's principles.

These principles were developed by the AI Working Group (2024–25) and approved by the Provost and Chief Operating Officer as recommendations on how to use AI across teaching, research, and operations. They are not ratified university policy. They represent the feedback gathered by the 2024–25 working group from a broad cross-section of university stakeholders and serve as guidance for responsible AI use at Virginia Tech.

Agentic AI

AI systems designed to take multi-step actions, use external tools, and pursue goals with minimal human oversight. Unlike standard language models that respond to individual prompts, agentic AI systems can browse the web, execute code, manage files, call APIs, and chain together operations. Agentic AI represents a significant shift from AI as a conversational tool to AI as an autonomous actor.

Fine-tuning

Further training an existing foundation model on a narrower, specialized dataset to improve performance on a specific task or domain. Fine-tuning adjusts the model’s parameters, as opposed to prompting, which works within the model’s existing capabilities without changing it.

Foundation Model

A large AI system trained on broad, diverse datasets that can be adapted to many downstream tasks through prompting, fine-tuning, or other techniques. GPT-4, Claude, Gemini, and Llama are examples. “Foundation” reflects the idea that one base model serves as the starting point for many applications.

Hallucination

This is when a model generates content that is factually incorrect, fabricated, or unsupported. Hallucination occurs because language models predict likely sequences of tokens but they have no mechanism for verifying whether output is true. While somewhat misleading because it implies human-like perception, the term is established in the field. No current model is hallucination-free.

Inference

The process of using a trained model to generate outputs from new inputs. For example, when a question is entered into HokieAI, inference produces the returned response. Inference is computationally expensive, though far less so than training, and accounts for much of the ongoing operational cost of AI services.

Large Language Model (LLM)

A type of foundation model designed to process and generate text. LLMs predict the most likely next token, or word or word fragment, based on patterns in training data. While responses can seem human-like, they do not understand meaning. Instead, they produce text that is statistically plausible given their training.

Model vs. Application

A model is the underlying AI system, such as GPT-4, Claude 3.5 Sonnet, Gemini, or Llama. An application is the product built on top of a model, such as ChatGPT, Copilot, or the ARC LLM Gateway. The same model can power multiple applications with different interfaces, safety filters, data handling practices, and terms of service.

When evaluating a tool, understanding the model it uses, and what the application layer adds or restricts, helps users select the best tool for their work.

Narrow AI vs. General AI

Narrow AI, also known as Weak AI, refers to systems designed for specific tasks and that are unable to understand and adapt like a human. General AI, also known as Artificial General Intelligence (AGI), is a theoretical future state for AI that aims to replicate human intelligence and cognitive ability.

While some Narrow AI applications can seem human-like, they are not able to operate beyond their specific programming. At this time, achieving General AI is hypothetical.

Parameters

A parameter is an internal variable within a machine learning model that is learned and adjusted during the training phase. A complete set of parameters represents all the knowledge the model has extracted from the training data.

Model size is often described in terms of parameter count. For example, GPT-3 has 175 billion parameters, and some models have trillions of parameters. More parameters generally allow for more complex responses, but parameter count alone does not determine quality.

Prompt Engineering

Prompt engineering is the act of crafting inputs or requests for an AI application with the goal of getting better outputs. Prompt techniques include providing examples of desired output, step-by-step reasoning, and more complex requests that define the role, constraints, and output format.

A good prompt is clear, creative, and encourages focused, thoughtful responses.

RAG (retrieval-augmented generation)

Retrieval-augmented generation (RAG) is a technique that combines a language model with an external knowledge source. The system retrieves relevant documents from a database, then provides them to the model as context.

This technique can reduce hallucination, for example when the answer exists in the external documents, and lets the model work with information that was not in its training data.

Slop

AI generated content that is low-quality and has little value. Some slop content can sound good when first read, but on analysis, lacks detail, factual information, and clarity. Slop can appear as text, images, video, or other digital content.

Tokens and Context Windows

A token is the basic unit of text a model processes. AI models break down inputs into tokens in order to process a response. Shorter words may be represented by one token while longer words can be represented by two or more tokens.

The context window is the maximum number of tokens a model can consider at once, covering input and output together. Context windows range from around 8,000 to over 1,000,000 tokens depending on the model. Content that exceeds the context window is invisible to the model.

Training Data

The datasets used to build a model during its initial training. Language models typically include books, websites, academic papers, code repositories, and other text sources.

The content in the training data directly affects what a model knows, what biases it carries, and where its knowledge has gaps. Keep this in mind when selecting a model, as AI companies may not fully disclose their training data.

AI Policy Stances

Virginia Tech recommends a three-stance taxonomy for course AI policies: Required (AI use is part of the learning activity), Limited (permitted under specific conditions, with disclosure), and Prohibited (not allowed for the assignment or course).

This approach helps make AI policies explicit and actionable rather than vague. Faculty may use different stances for different assignments within the same course.

Authentic Assessment

An assessment task grounded in personal experience, local context, real-world data, or situations AI cannot easily replicate.

Examples include analyzing a dataset the student collected, reflecting on a clinical observation, applying course concepts to a specific campus or community issue, and conducting and reporting on an original interview.

Authentic assessments are more resistant to AI-generated submissions because they require knowledge and experience not included in the AI model's training data. These assessments also tend to be more engaging for students, though they may require more design effort.

Desirable Difficulty

Desirable Difficulty is a concept from learning science (Bjork & Bjork). It suggests that tasks that are difficult in the right way produce more durable learning than tasks that feel easy.

Examples include spacing practice over time, interleaving different problem types, and retrieving information from memory. These approaches may slow initial performance but strengthen long-term retention.

AI tools often remove difficulty from academic tasks. Faculty are encouraged to consider whether the difficulty being removed by AI is desirable (serving learning) or undesirable (simply a barrier) when designing assessments.

Illusion of Fluency

AI output that reads smoothly, uses appropriate vocabulary, and follows disciplinary conventions may present an illusion of fluency. This fluency can create a false sense of understanding in both the system and the reader.

A student who accepts AI-generated text without critical engagement may feel they understand the material because the text sounds authoritative, without having done the cognitive work that produces actual understanding.

Illusion of fluency is one of the most significant pedagogical risks associated with AI writing tools.

Process-based Assessment

Process-based assessment evaluates intermediate work products, not just final deliverables.

Examples include annotated drafts, revision histories, reflective essays, in-class writing samples, and oral defenses. This type of assessment is more resistant to AI misuse because it makes thinking visible throughout the learning process.

In addition, this approach provides better formative feedback. The trade-off is increased grading workload, which can often be addressed through thoughtful course design.

Productive Friction

Productive friction is the cognitive effort required for genuine learning. Writing a first draft is difficult because organizing ideas into prose forces the writer to clarify thinking, identify gaps, and make choices about emphasis.

This friction is where much of the learning occurs. AI can eliminate productive friction (the struggle that builds understanding) alongside unproductive friction (formatting, boilerplate, or repetitive tasks).

Faculty benefit from identifying which sources of friction in their assignments are worth preserving to support meaningful learning.