AI Glossary

 

This glossary defines AI terms you are likely to encounter in CoE communications, training, vendor conversations, news coverage, and daily work.

The Communications Committee reviews and updates it quarterly. If you encounter an AI term not listed here, contact the CoE and we will add it.

A

Agenttulane, ops At Tulane

A software system that can take actions on your behalf, not just generate a response. Unlike a standard AI tool that answers one question at a time, an agent can receive a goal, break it into steps, use tools or external systems, and work through those steps with minimal hand-holding. Most agents combine a language model with the ability to search, retrieve data, write files, send messages, or trigger other software. The key distinction from a regular AI tool: an agent acts, it does not just respond.

At Tulane, in nebulaONE/TulaneAI, “agent” refers to the combination of a large language model with specific prompt instructions and system connectors that executes tasks on behalf of Tulane community users.
Agentic AIops

AI that can plan and take multi-step actions with limited human involvement. Instead of responding to one question at a time, an agentic system might research a topic, draft a summary, and email it to your team on its own.

Algorithmcore

A set of rules or instructions a computer follows to complete a task. In AI, algorithms are the processes that allow a system to learn from data and generate outputs. You do not need to understand how algorithms work to use AI tools. The key thing to know is that AI outputs are produced by pattern-matching, not by understanding or reasoning the way a person does.

Anonymizationrisk At Tulane

Removing or changing personally identifiable information in data so that no individual can be recognized from what remains.

At Tulane, anonymization is required before using Confidential (Level 3) data in most AI tools. Example: replacing employee names and TU IDs with random codes before asking an AI tool to analyze staffing patterns.
API (Application Programming Interface)tulane, core

A connection that allows two software systems to exchange information. When an AI tool pulls data from another system or sends results somewhere, it is usually using an API. You may hear this term when IT or vendors discuss integrating AI into Oracle or other enterprise platforms.

Artificial Intelligence (AI)core

Technology that enables computers to perform tasks that normally require human thinking, such as writing, analyzing data, recognizing patterns, generating images, or making recommendations. AI is a broad category. Most of the tools the COO division uses fall under a specific type called generative AI.

Audit Trail / Loggingrisk At Tulane

A recorded history of how an AI tool was used: what inputs were provided, what outputs were generated, and who was involved. Audit trails matter for compliance, accountability, and troubleshooting.

This is one of the key reasons the CoE requires enterprise tools over free tools. Enterprise platforms maintain audit trails. Free tools typically do not.

B

Bias (in AI)risk

When an AI system produces results that are skewed, incomplete, or unbalanced in ways that may not be obvious. Bias can show up as favoring certain groups of people, but it can also mean overrepresenting certain ideas, perspectives, industries, or approaches while ignoring others. Bias enters through the data the AI was trained on and through the way prompts are written. Example: an AI tool trained mostly on corporate data might consistently recommend solutions that do not fit a university’s structure or values. Checking for bias is part of responsible AI use.

BoodleBoxtulane At Tulane

An approved enterprise AI platform in use within the COO division. Like Tulane AI (nebulaONE/TulaneAI), BoodleBox is a managed environment that provides access to multiple AI models with governance controls, so staff can work with AI tools without creating individual accounts on outside platforms.

Bot / Chatbotcore

A software program that simulates conversation. Simple chatbots follow scripted rules (like a phone menu). AI-powered chatbots use language models to generate responses in real time and can handle a wider range of questions. Example: an HR chatbot that answers employee benefits questions by pulling from a knowledge base.

C

Claudetulane At Tulane

An AI language model made by Anthropic. Claude models are available through Tulane AI (nebulaONE/TulaneAI) in a governed environment. Direct subscriptions to Claude (paid or free) are not approved for Tulane business use.

Cloud Modeltulane, core

An AI model that runs on remote servers accessed over the internet, rather than on your computer or on Tulane’s own hardware. Most enterprise AI tools, including Tulane AI and Copilot, use cloud models. This is why data classification matters: when you use a cloud-based AI tool, your input travels to external servers.

Computer Visioncore

AI that can interpret and analyze visual information such as photos, videos, scanned documents, or handwritten text. You may encounter computer vision in document processing, invoice scanning, or features within Oracle and other enterprise systems that extract data from images or PDFs.

Context Windowusing

The amount of text an AI model can work with at one time, including both what you give it and what it generates back. Think of it as the model’s working memory for a single conversation. Larger context windows let you provide more background information at once, like pasting in a long document and asking questions about it.

Copilot (Microsoft Copilot)tulane At Tulane

An AI assistant built into Microsoft’s productivity tools (Word, Excel, Outlook, Teams). Covered under Tulane’s existing Microsoft agreement. Appropriate for Public and Internal data (Levels 1 and 2) only.

D

Data Classificationrisk At Tulane

A system for categorizing data by how sensitive it is. At Tulane, four levels determine which AI tools you can use: Level 1 (Public), Level 2 (Internal), Level 3 (Confidential), and Level 4 (Restricted). The Quick Reference Card has the full breakdown. The most important rule: classify your data first, then choose the tool that matches.

Data Privacyrisk At Tulane

Protecting personal information that AI systems access or process. This includes complying with laws and policies on how data is collected, stored, used, and shared.

At Tulane, data privacy requirements vary by data classification level and by the type of data involved (employee records, student records, financial data, etc.). The AI Usage Guidelines and the data classification framework are the primary references.
Data Steward / Data Ownerops

A person or role responsible for the quality, protection, and proper use of a specific category of data. For example, HR is the data steward for employee records, and the Registrar is the data steward for student enrollment data. When an AI tool needs access to a particular type of data, the data steward for that category should be involved in the review.

Deepfakerisk At Tulane

AI-generated video, audio, or images that realistically depict a real person saying or doing something they never actually said or did. Creating deepfakes is prohibited under the AI Usage Guidelines.

E

Enterprise AI Tooltulane At Tulane

An AI tool provided under a paid institutional license with data protection, access controls, and audit trails. Tulane’s enterprise tools are Tulane AI (nebulaONE/TulaneAI), BoodleBox, and Microsoft Copilot. Enterprise tools are distinct from free or personal-subscription tools, which lack these protections and are not approved for Tulane business use.

F

Fine-Tuningskills

The process of taking a general-purpose AI model and training it further on a specific set of data so it performs better for a particular task or domain. Example: a general language model might be fine-tuned on university procurement documents so it understands Tulane’s contract terminology and formatting. Fine-tuning is a technical process handled by IT or vendors, not something individual staff need to do.

Foundation Modelcore

A large AI model trained on broad data that serves as the starting point for many different applications. GPT-4, Claude, Gemini, and Llama are foundation models. Think of them as general-purpose engines. You interact with foundation models every time you use Tulane AI or Copilot, even though you do not see the model itself.

G

Generative AI (GenAI)core

AI that creates new content, such as text, images, code, audio, or video, based on patterns learned from its training data. When you type a question into Tulane AI and it writes a response, that is generative AI at work. Most of the AI tools the COO division uses are generative AI tools.

GPT (Generative Pre-trained Transformer)tulane, core

A family of AI language models developed by OpenAI. GPT is the technology behind ChatGPT. The term has become common enough that people sometimes use “GPT” loosely to mean any large language model, though it technically refers to OpenAI’s models specifically. GPT models are available through Tulane AI (nebulaONE/TulaneAI).

Groundingusing

Connecting an AI model to verified, specific data sources so its responses draw on real information rather than only its general training. A grounded AI system might pull from your department’s documents or Tulane’s knowledge bases when answering questions. Grounding reduces hallucinations and makes outputs more relevant. You may hear this term when IT or vendors discuss how Tulane AI will connect to internal systems.

H

Hallucinationrisk

When an AI tool generates information that sounds confident and plausible but is factually wrong. AI tools do not “know” things the way people do. They predict text based on patterns, which means they can produce convincing nonsense. Always verify AI outputs before using them in decisions, reports, or communications. This is one of the core reasons for the Human at the Helm principle.

Human at the Helmrisk At Tulane

The CoE’s first guiding principle. Every AI output requires human review before it is used, shared, or acted upon. AI augments human judgment. It does not replace it. In practice, this means you treat AI-generated text, data, recommendations, and analysis as a starting point, not a final answer. You are always the decision-maker.

Human-in-the-Loop (HITL)risk At Tulane

A design approach where humans review, approve, or override AI outputs before they take effect. This is especially important for decisions that affect people, such as hiring recommendations, financial approvals, or student services.

Human at the Helm is Tulane’s guiding principle. Human-in-the-loop is the technical design pattern that puts that principle into practice.

I

Image Generationusing At Tulane

Using AI to create images from text descriptions. You type a description of what you want (for example, “a chart showing three overlapping circles labeled HR, Finance, and IT”), and the AI generates an image. Tools like DALL-E, Midjourney, and features within some enterprise platforms offer image generation. At Tulane, AI-generated images used externally must be disclosed.

Inferencecore

The process of an AI model generating an output based on your input. When you type a prompt and the AI responds, that response is the result of inference. Each inference consumes computational resources measured in tokens, which is why AI platforms track usage and cost.

K

Knowledge Baseusing

A structured collection of documents, data, or information that an AI system can reference when generating responses. Instead of relying only on its general training, an AI connected to a knowledge base can answer questions using your organization’s specific information. Example: connecting Tulane AI to a collection of HR policy documents so it can answer benefits questions accurately.

L

Large Language Model (LLM)core

The type of AI that powers tools like Tulane AI, ChatGPT, Claude, Copilot, and Gemini. LLMs are trained on massive amounts of text and learn to predict and generate language. They are effective for drafting, summarizing, analyzing, brainstorming, and coding, but they can also produce errors and hallucinations. “LLM” and “language model” are used interchangeably.

M

Machine Learning (ML)core

A branch of AI where systems learn from data instead of following manually written rules. Rather than programming a computer with every possible scenario, you give it examples and let it find patterns. Most modern AI tools are built on machine learning. Generative AI is one application of it.

Modelcore At Tulane

In AI, a model is the trained system that processes your inputs and generates outputs. Different models have different strengths, speeds, and costs. Tulane AI (nebulaONE/TulaneAI) provides access to multiple models, so the right model can be matched to the task.

Model Monitoring / Driftrisk

Tracking how an AI model performs over time to detect when it starts producing lower-quality or less accurate results. Models can “drift” as the data they encounter changes or as conditions shift. For example, a model trained on pre-pandemic procurement patterns might give poor recommendations in a post-pandemic budget environment. Monitoring ensures that AI tools continue to perform as expected and flags when retraining or replacement is needed.

Multimodalcore

An AI system that can work with more than one type of input or output: text, images, audio, video, or code. A multimodal model might accept an uploaded photo and a text question at the same time and respond with a written description of what is in the image. Many newer AI models are multimodal.

N

Natural Language Processing (NLP)core

The branch of AI that enables computers to understand, interpret, and generate human language. When you type a question to an AI tool in plain English and it understands what you are asking, that is NLP at work. NLP is the foundation that makes conversational AI tools possible.

nebulaONEtulane At Tulane

The enterprise AI platform selected for the COO division. Provides individual workspaces with controlled model access, token tracking, and governance visibility. Approved for all data classification levels (Levels 1 through 4). In staff-facing communications, nebulaONE is referred to as “Tulane AI.”

Neural Networkcore

A type of AI model inspired by the structure of the human brain, made of interconnected layers that process data to recognize patterns and make predictions. Neural networks are the building blocks behind image recognition, speech transcription, and language models. You do not need to understand how neural networks work. The term is helpful to recognize because it appears in AI product descriptions and news coverage.

O

Open Source (AI)skills At Tulane

AI models or tools whose underlying code is publicly available for anyone to use, modify, or build upon. Llama (Meta) and Mistral are examples of open-source language models. Open source does not automatically mean free, safe, or approved for Tulane business use. The same tool approval process applies regardless of whether a model is open source or proprietary.

Oracle AItulane, ops At Tulane

AI features embedded within Tulane’s Oracle ERP and HCM systems. These features are being activated through the DriveStream phased rollout and require a technical and data risk assessment before use. You may hear about Oracle AI in the context of procurement, HR workflows, or financial reporting.

P

Personal AI Assistantskills, tulane At Tulane

At Tulane, a targeted AI tool built to solve a specific problem in an employee’s workflow. Example: an assistant that drafts first versions of standard procurement responses using your department’s templates and past examples. Personal AI Assistants are one of the primary ways the CoE tracks measurable value from AI adoption.

Predictive Analyticsops

Using historical data and AI models to forecast future outcomes. Examples in a university operations context: predicting budget shortfalls based on spending patterns, forecasting employee turnover in specific departments, or estimating facilities maintenance needs based on equipment age and usage data.

Prescriptive Analyticsops

A step beyond predictive analytics. Prescriptive models do not just forecast what might happen. They recommend specific actions to improve the outcome. Example: rather than just predicting which departments will exceed their travel budget, a prescriptive model might recommend specific spending adjustments by category and timeline to stay on target.

Promptusing

The text, question, or instructions you give an AI tool to get a response. The quality and specificity of your prompt directly affects the quality of the output. Writing effective prompts is a core skill covered in CoE training.

Prompt Engineeringusing, skills

The practice of writing and refining prompts to get better, more consistent results from AI tools. This includes techniques like giving the AI a specific role (“You are a procurement specialist”), providing examples of what you want, breaking complex tasks into steps, and specifying format requirements. Prompt engineering is not programming. It is a skill anyone can develop.

Prompt Injectionrisk

A technique where hidden or misleading instructions are embedded in text that an AI tool processes, causing it to behave in unintended ways. This is one reason the AI Usage Guidelines require CoE review of custom AI skills and prompt packages downloaded from public sources. They can contain hidden instructions that compromise security or produce unreliable outputs.

Prompt Libraryusing At Tulane

A shared, curated collection of effective prompts that have been tested and documented. A prompt library allows teams to reuse what works instead of starting from scratch each time. The CoE is developing a prompt library for the COO division as part of the Phase 2 roadmap.

R

RAG (Retrieval-Augmented Generation)using

A method that makes AI more accurate by connecting it to a specific set of documents or data when it generates a response. Instead of relying only on what it learned during training, the AI retrieves relevant information from your source material first, then uses it to produce a better answer. This is how many enterprise AI tools provide answers based on your organization’s actual documents and data rather than general knowledge.

Responsible AIrisk At Tulane

The broad set of principles and practices that ensure AI systems are fair, transparent, safe, privacy-respecting, and aligned with institutional values. Responsible AI is not a single rule. It is the umbrella concept that ties all of the CoE’s governance activities together.

At Tulane, responsible AI is operationalized through the AI Usage Guidelines, the ethics checklist, the governance tier system, and the Human at the Helm principle.
Risk Tieringrisk At Tulane

Categorizing AI use cases by their potential impact and risk level to determine how much oversight they require. Higher-risk use cases (those involving sensitive data, financial decisions, or autonomous actions) get more review. Lower-risk use cases (personal productivity with public data) need less.

At Tulane, this concept is implemented through the four governance tiers (Tier 0 through Tier 3) defined in the AI Usage Guidelines.
RPA (Robotic Process Automation)ops

Software that automates repetitive, rules-based tasks by mimicking the clicks and keystrokes a human would perform. RPA is different from generative AI. It does not create new content or make judgment calls. It follows a script to move data between systems, fill forms, or process transactions. RPA is sometimes combined with AI to handle tasks that require both structured automation and intelligent decision-making. You may hear about RPA in the context of Oracle, procurement, or HR process improvements.

S

Sandboxtulane At Tulane

A controlled environment where you can experiment with AI tools without affecting real data or live systems. Tulane AI (nebulaONE/TulaneAI) provides individual sandboxes for each user so you can practice and test safely.

Sentiment Analysisops

AI’s ability to detect the overall tone or attitude in text, such as whether feedback is positive, negative, or neutral. Example: analyzing open-ended survey responses to identify themes and overall employee sentiment without reading every response individually.

Shadow AIrisk At Tulane

The use of unapproved AI tools for work purposes. Shadow AI creates data security risks because unapproved tools may not protect university data, may use your inputs to train their models, or may lack audit trails. Common example: pasting internal data into a free ChatGPT account. The CoE’s approach is to provide approved alternatives so people have a safe, sanctioned path.

System of Record (SOR)ops At Tulane

The authoritative system where core data is created and maintained. AI tools should pull from and respect these systems as the source of truth. AI outputs do not replace what is in the system of record.

For example, Oracle is Tulane’s system of record for financial and HR data. The student information system is the system of record for enrollment data.
System Prompt / Custom Instructionsusing At Tulane

Persistent instructions that shape how an AI tool behaves across an entire conversation or session. A system prompt might tell the AI “You are a Tulane HR specialist. Always reference Tulane policies. Never provide legal advice.” The AI then follows those instructions for everything it generates in that session. System prompts are how Personal AI Assistants are configured to behave consistently for a specific task.

T

Temperatureusing

A setting that controls how predictable or creative an AI model’s responses are. Lower temperature produces more consistent, focused output. Higher temperature produces more varied, creative output. If you are generating something that needs to be accurate, like a policy summary, lower temperature is better. For brainstorming or creative drafting, higher temperature can help. Some AI tools let you adjust this setting directly.

Tokencore At Tulane

A unit of text that AI models process. A token is roughly three-quarters of a word. Both your input (the prompt) and the AI’s output (the response) are measured in tokens. Token usage is tracked in Tulane AI (nebulaONE/TulaneAI) to monitor consumption and manage costs. Every time you use an AI tool, you are spending tokens, similar to how a phone plan tracks minutes.

Training Datacore

The information an AI model was exposed to when it was originally built. Training data shapes what the model knows and how it responds. A model trained mostly on English-language business documents will perform differently than one trained on scientific research. AI tools do not update their training data in real time, which is why they can be wrong about recent events or unfamiliar topics.

Transparency / Explainabilityrisk

How understandable an AI system’s decisions or outputs are to humans. This includes knowing what data the AI used, how it arrived at a result, and what its limitations are. Transparency matters most when AI outputs inform decisions that affect people or money. If you cannot explain why an AI tool recommended something, that is a signal to apply more human review before acting on it.

Tulane AItulane At Tulane

The staff-facing name for the nebulaONE enterprise AI platform. Used in newsletters, the AI webpage, training materials, and communications directed at COO division employees. When someone says “Tulane AI,” they mean the nebulaONE platform.

W

Workflow Automationops

Using AI to handle repetitive, multi-step processes that previously required manual effort. Example: automatically routing incoming requests, extracting key details from documents, and drafting initial responses. Workflow automation is distinct from using AI as a one-time assistant for a single task. It involves setting up AI to handle a recurring process on an ongoing basis.

Workflow Orchestrationops

Coordinating tasks and data across multiple systems and teams, often combining human steps with automated and AI-driven actions into a single end-to-end process. Workflow orchestration is broader than workflow automation. Automation handles one process. Orchestration connects multiple processes, systems, and decision points into a coordinated flow. You may hear this term in conversations about enterprise-scale AI projects that span HR, finance, and procurement.

Z

Zero-Shot / Few-Shot Promptingusing

Describes how many examples you give an AI model before asking it to perform a task. Zero-shot means you give no examples, just instructions. Few-shot means you include a small number of examples to guide the output. Providing even two or three examples of what you want generally produces better, more consistent results. This is a practical prompting technique covered in CoE training.

AI terminology evolves quickly. New terms emerge, existing terms shift meaning, and vendor-specific names come and go. This glossary focuses on terms that are stable enough to be useful and relevant enough to appear in your work, training, or conversations. It is not exhaustive. If a term is missing and you need it defined, contact the CoE.

 

Document
AI Glossary, Office of the COO
Version 
2.0
Date 
June 2026
Status 
Current
Owner 
AI Center of Excellence, Office of the COO
Maintained by 
Communications Committee

 

This document was developed with AI assistance. Content has been reviewed by the Communications Committee.