{"id":428,"date":"2026-06-01T18:20:42","date_gmt":"2026-06-01T18:20:42","guid":{"rendered":"https:\/\/sites.clarkson.edu\/ai-institute\/workbench\/tools\/specs\/"},"modified":"2026-06-19T15:33:31","modified_gmt":"2026-06-19T15:33:31","slug":"specs","status":"publish","type":"page","link":"https:\/\/sites.clarkson.edu\/ai-institute\/workbench\/tools\/specs\/","title":{"rendered":"Specs \u2014 AI Terms in Plain Language"},"content":{"rendered":"\n\u2190 Workbench<\/a>\n\n\n\n\n
\n\n
\n

Clarkson AI Institute · Workbench · Specs<\/p>\n

AI Terms in Plain Language<\/h1>\n

A working glossary for faculty, staff, students, administrators, and partners using AI at Clarkson. Terms are in lay language. Technical depth is available elsewhere.<\/p>\n\n\n\n

\nA<\/a>\nB<\/a>\nC<\/a>\nD<\/a>\nE<\/a>\nF<\/a>\nG<\/a>\nH<\/a>\nI<\/a>\nJ<\/a>\nK<\/a>\nL<\/a>\nM<\/a>\nN<\/a>\nO<\/a>\nP<\/a>\nQ<\/a>\nR<\/a>\nS<\/a>\nT<\/a>\nU<\/a>\nV<\/a>\nW<\/a>\nX<\/a>\nZ<\/a>\n<\/div>\n<\/div>\n\n
\n\n

A<\/h2>\n

Accuracy<\/strong> — How often an AI system gets the answer right.<\/p><\/div>\n

Active learning<\/strong> — A training method where the AI asks humans to label the examples it is most unsure about.<\/p><\/div>\n

Adversarial attack<\/strong> — A trick designed to make an AI system fail, such as slightly changing an image so the AI misidentifies it.<\/p><\/div>\n

Agent<\/strong> — An AI system that can take steps toward a goal, such as searching, planning, using tools, writing files, or completing tasks.<\/p><\/div>\n

Agentic AI<\/strong> — AI that does more than answer questions; it can act, plan, use tools, and pursue multi-step goals.<\/p><\/div>\n

AGI (Artificial General Intelligence)<\/strong> — A hypothetical AI that could perform most intellectual tasks humans can do. This does not currently exist in any settled, agreed-upon sense.<\/p><\/div>\n

AI (Artificial Intelligence)<\/strong> — Computer systems designed to do tasks that normally require human intelligence, such as recognizing images, understanding language, making predictions, or solving problems.<\/p><\/div>\n

AI alignment<\/strong> — The effort to make AI systems behave in ways that match human values, intentions, and safety needs.<\/p><\/div>\n

AI assistant<\/strong> — A conversational AI system that helps users write, search, summarize, code, plan, or reason.<\/p><\/div>\n

AI audit<\/strong> — A formal review of an AI system to check its safety, fairness, accuracy, security, legality, or reliability.<\/p><\/div>\n

AI benchmark<\/strong> — A test used to compare how well different AI systems perform.<\/p><\/div>\n

AI ethics<\/strong> — The study and practice of making AI fair, safe, accountable, transparent, and beneficial.<\/p><\/div>\n

AI governance<\/strong> — Rules, policies, processes, and oversight structures for controlling how AI is built and used.<\/p><\/div>\n

AI hallucination<\/strong> — When an AI confidently gives information that is false, made up, or unsupported.<\/p><\/div>\n

AI model<\/strong> — The trained system that produces predictions, answers, classifications, images, text, or other outputs.<\/p><\/div>\n

AI policy<\/strong> — Laws, rules, institutional guidelines, or public policies about AI.<\/p><\/div>\n

AI safety<\/strong> — Work focused on preventing AI from causing harm.<\/p><\/div>\n

AI system<\/strong> — The full setup around an AI model, including data, software, user interface, rules, tools, monitoring, and human oversight.<\/p><\/div>\n

Algorithm<\/strong> — A set of steps or instructions a computer follows to solve a problem.<\/p><\/div>\n

Algorithmic bias<\/strong> — Unfair or skewed results caused by data, design choices, assumptions, or social inequalities built into a system.<\/p><\/div>\n

Annotation<\/strong> — Adding labels or notes to data so an AI can learn from it.<\/p><\/div>\n

API (Application Programming Interface)<\/strong> — A way for software programs to communicate with one another.<\/p><\/div>\n

Artificial neural network<\/strong> — A computer system loosely inspired by the brain, made of connected layers that learn patterns from data.<\/p><\/div>\n

Attention<\/strong> — A technique that helps an AI model decide which parts of the input matter most.<\/p><\/div>\n

Automation<\/strong> — Having a machine or computer do a task with little or no human involvement.<\/p><\/div>\n

Autonomous system<\/strong> — A system that can act on its own within some limits.<\/p><\/div>\n\n

B<\/h2>\n

Backpropagation<\/strong> — A training method that helps a neural network adjust itself after making mistakes.<\/p><\/div>\n

Base model<\/strong> — A general-purpose AI model before it has been specialized for a particular task or behavior.<\/p><\/div>\n

Batch<\/strong> — A group of examples processed together during training.<\/p><\/div>\n

Benchmark contamination<\/strong> — When an AI has seen the test questions during training, making its score look better than it really is.<\/p><\/div>\n

Bias<\/strong> — A systematic tilt in data, model behavior, or outcomes. Bias can be statistical, social, cultural, or political.<\/p><\/div>\n

Big data<\/strong> — Extremely large datasets that can be analyzed to find patterns.<\/p><\/div>\n

Black box<\/strong> — A system whose inner workings are hard for people to understand, even if its outputs are visible.<\/p><\/div>\n

Bot<\/strong> — A software program that performs tasks automatically. A chatbot is a bot designed for conversation.<\/p><\/div>\n\n

C<\/h2>\n

Calibration<\/strong> — How well an AI’s confidence matches reality. A well-calibrated AI is more confident when it is likely right and less confident when it is likely wrong.<\/p><\/div>\n

Chain of thought<\/strong> — Step-by-step reasoning used by a model internally or displayed externally. In many systems, the full private reasoning is not shown to users.<\/p><\/div>\n

Chatbot<\/strong> — A program designed to communicate through conversation.<\/p><\/div>\n

Classification<\/strong> — Sorting something into categories, such as spam\/not spam or benign\/malignant.<\/p><\/div>\n

Classifier<\/strong> — An AI system that assigns labels or categories.<\/p><\/div>\n

Cloud AI<\/strong> — AI that runs on remote servers rather than directly on your device.<\/p><\/div>\n

Clustering<\/strong> — Grouping similar items together without already knowing the categories.<\/p><\/div>\n

Code generation<\/strong> — AI writing computer code from a prompt or description.<\/p><\/div>\n

Computer vision<\/strong> — AI that works with images or video, such as recognizing faces, reading scans, or detecting objects.<\/p><\/div>\n

Confidence score<\/strong> — A number showing how sure the AI is about its answer.<\/p><\/div>\n

Context<\/strong> — The information available to the AI when it produces an answer.<\/p><\/div>\n

Context engineering<\/strong> — Designing the information given to an AI so it performs better.<\/p><\/div>\n

Context length \/ context window<\/strong> — How much information an AI model can consider at once.<\/p><\/div>\n

Conversational AI<\/strong> — AI built to communicate with people using natural language.<\/p><\/div>\n

Copilot<\/strong> — A kind of AI assistant that works alongside a person, often in coding, writing, or office software.<\/p><\/div>\n

Corpus<\/strong> — A collection of texts or other data used for training or analysis.<\/p><\/div>\n

Curation<\/strong> — Selecting, organizing, cleaning, or preparing information for use.<\/p><\/div>\n\n

D<\/h2>\n

Data<\/strong> — Information used by computers, such as text, numbers, images, audio, clicks, records, or sensor readings.<\/p><\/div>\n

Data augmentation<\/strong> — Creating modified versions of data to help a model learn better.<\/p><\/div>\n

Data drift<\/strong> — When real-world data changes over time, causing an AI system to become less accurate.<\/p><\/div>\n

Data labeling<\/strong> — Adding correct answers or tags to examples so a model can learn.<\/p><\/div>\n

Data leakage<\/strong> — When information from the test set or future data accidentally gets into training, making performance look falsely strong.<\/p><\/div>\n

Data mining<\/strong> — Searching large datasets for patterns.<\/p><\/div>\n

Dataset<\/strong> — A structured collection of data.<\/p><\/div>\n

Decision boundary<\/strong> — The line or rule a model uses to separate one category from another.<\/p><\/div>\n

Deep learning<\/strong> — Machine learning using large neural networks with many layers.<\/p><\/div>\n

Deepfake<\/strong> — AI-generated or AI-altered media that makes someone appear to say or do something they did not.<\/p><\/div>\n

Deployment<\/strong> — Putting an AI system into real-world use.<\/p><\/div>\n

Diffusion model<\/strong> — A common type of image or media-generation model that learns to create outputs by reversing a process of adding noise.<\/p><\/div>\n

Dimensionality<\/strong> — The number of features or measurements used to represent something.<\/p><\/div>\n

Distillation<\/strong> — Training a smaller model to imitate a larger model.<\/p><\/div>\n

Domain adaptation<\/strong> — Adjusting an AI system so it works better in a new setting or field.<\/p><\/div>\n

Downstream task<\/strong> — A specific task a general model is later used for, such as summarization, translation, or diagnosis support.<\/p><\/div>\n\n

E<\/h2>\n

Edge AI<\/strong> — AI that runs on a local device, such as a phone, camera, car, or sensor, rather than in the cloud.<\/p><\/div>\n

Embedding<\/strong> — A way of turning words, images, or other data into numbers so a computer can compare meaning or similarity.<\/p><\/div>\n

Emergent behavior<\/strong> — A surprising ability or pattern that appears in a large AI system even though it was not directly programmed.<\/p><\/div>\n

Encoder<\/strong> — Part of a model that turns input into an internal representation.<\/p><\/div>\n

End-to-end learning<\/strong> — Training one system to go from raw input to final output without many hand-built intermediate steps.<\/p><\/div>\n

Ensemble<\/strong> — A system that combines several models to get better results.<\/p><\/div>\n

Evaluation (eval)<\/strong> — Testing an AI system to see how well it works.<\/p><\/div>\n

Explainability<\/strong> — The ability to explain why an AI system produced a particular result.<\/p><\/div>\n

Expert system<\/strong> — An older style of AI that uses human-written rules to imitate expert decision-making.<\/p><\/div>\n\n

F<\/h2>\n

False negative<\/strong> — When an AI says something is not present, but it actually is. Example: saying a disease is absent when it is present.<\/p><\/div>\n

False positive<\/strong> — When an AI says something is present, but it is not. Example: flagging a harmless email as spam.<\/p><\/div>\n

Feature<\/strong> — A piece of information used by a model, such as age, word frequency, pixel color, or purchase history.<\/p><\/div>\n

Feature engineering<\/strong> — Choosing or designing the information a model uses to make predictions.<\/p><\/div>\n

Few-shot learning<\/strong> — Giving a model a few examples in the prompt so it can imitate the pattern.<\/p><\/div>\n

Fine-tuning<\/strong> — Further training a general model on special data so it performs better for a particular task, style, field, or behavior.<\/p><\/div>\n

Foundation model<\/strong> — A large general-purpose model trained on broad data and adaptable to many tasks.<\/p><\/div>\n

Frontier model<\/strong> — A highly capable, cutting-edge AI model near the current limits of performance.<\/p><\/div>\n\n

G<\/h2>\n

GAN (Generative Adversarial Network)<\/strong> — A kind of AI system where two models compete: one creates fake examples, and the other tries to detect them.<\/p><\/div>\n

Generalization<\/strong> — A model’s ability to perform well on new examples, not just the data it was trained on.<\/p><\/div>\n

Generative AI<\/strong> — AI that creates new content, such as text, images, audio, video, code, or designs.<\/p><\/div>\n

Generative model<\/strong> — A model that produces new outputs rather than only classifying or predicting labels.<\/p><\/div>\n

GPU (Graphics Processing Unit)<\/strong> — A computer chip often used to train and run AI because it can do many calculations at once.<\/p><\/div>\n

Gradient descent<\/strong> — A method for gradually adjusting a model to reduce its errors.<\/p><\/div>\n

Ground truth<\/strong> — The correct answer used to train or test an AI system.<\/p><\/div>\n

Grounding<\/strong> — Connecting an AI’s answer to reliable sources, facts, data, documents, or real-world evidence.<\/p><\/div>\n

Guardrails<\/strong> — Rules or systems that try to keep AI behavior safe and appropriate.<\/p><\/div>\n\n

H<\/h2>\n

Hallucination<\/strong> — A false or unsupported AI output that sounds plausible.<\/p><\/div>\n

Human-in-the-loop<\/strong> — A system design where humans review, approve, correct, or supervise AI decisions.<\/p><\/div>\n

Human-on-the-loop<\/strong> — Humans monitor the AI system from above, but do not approve every individual action.<\/p><\/div>\n

Human-out-of-the-loop<\/strong> — An AI system acts without meaningful human review.<\/p><\/div>\n

Hyperparameter<\/strong> — A setting chosen before or during training, such as learning rate, model size, or batch size.<\/p><\/div>\n\n

I<\/h2>\n

Image generation<\/strong> — AI creating images from text, sketches, reference images, or other inputs.<\/p><\/div>\n

Inference<\/strong> — Using a trained model to produce an answer, prediction, or output.<\/p><\/div>\n

Instruction tuning<\/strong> — Training a model to follow human instructions more usefully.<\/p><\/div>\n

Interpretability<\/strong> — Understanding what parts of a model or input led to a certain output.<\/p><\/div>\n

Iteration<\/strong> — Repeatedly improving a system through cycles of testing and revision.<\/p><\/div>\n\n

J<\/h2>\n

Jailbreak<\/strong> — A prompt or attack designed to bypass an AI system’s safety rules.<\/p><\/div>\n

JSON mode<\/strong> — A setting that makes a model output structured data in JSON format, useful for software systems.<\/p><\/div>\n\n

K<\/h2>\n

Knowledge base<\/strong> — A collection of information an AI system can use or search.<\/p><\/div>\n

Knowledge cutoff<\/strong> — The latest date through which a model’s built-in training knowledge is likely to be current.<\/p><\/div>\n

Knowledge graph<\/strong> — A structured map of entities and relationships, such as people, places, events, and concepts.<\/p><\/div>\n\n

L<\/h2>\n

Label<\/strong> — The answer attached to a training example.<\/p><\/div>\n

Labeled data<\/strong> — Data that includes correct answers or categories.<\/p><\/div>\n

Language model<\/strong> — An AI model trained to understand, predict, or generate language.<\/p><\/div>\n

LLM (Large Language Model)<\/strong> — A large AI model trained on massive amounts of text and often able to write, summarize, translate, code, and answer questions.<\/p><\/div>\n

Latency<\/strong> — How long it takes a system to respond.<\/p><\/div>\n

Learning rate<\/strong> — A setting that controls how quickly a model changes during training.<\/p><\/div>\n

Local model<\/strong> — An AI model running on your own device or server rather than through a cloud provider.<\/p><\/div>\n

Loss<\/strong> — A measure of how wrong a model is during training.<\/p><\/div>\n\n

M<\/h2>\n

Machine learning (ML)<\/strong> — A branch of AI where systems learn patterns from data rather than being programmed with every rule.<\/p><\/div>\n

Machine translation<\/strong> — AI translation from one language to another.<\/p><\/div>\n

Metadata<\/strong> — Information about information, such as author, date, file type, location, or source.<\/p><\/div>\n

Model card<\/strong> — A document describing what a model does, how it was trained, intended uses, limitations, and risks.<\/p><\/div>\n

Model collapse<\/strong> — A possible problem where future AI models trained heavily on AI-generated material become worse or less diverse.<\/p><\/div>\n

Model drift<\/strong> — When an AI system’s performance changes over time because the world, data, or user behavior changes.<\/p><\/div>\n

Model weights<\/strong> — The internal numbers learned during training that shape how a model behaves.<\/p><\/div>\n

Multimodal AI<\/strong> — AI that can work with more than one kind of input or output, such as text, images, audio, and video.<\/p><\/div>\n\n

N<\/h2>\n

Natural language<\/strong> — Ordinary human language, such as English, Spanish, or Mandarin.<\/p><\/div>\n

NLP (Natural Language Processing)<\/strong> — AI that works with human language.<\/p><\/div>\n

Neural network<\/strong> — A model made of connected artificial neurons that learn patterns from data.<\/p><\/div>\n

Noise<\/strong> — Randomness, errors, irrelevant information, or messy variation in data.<\/p><\/div>\n

Non-deterministic<\/strong> — Not always giving the exact same answer every time.<\/p><\/div>\n\n

O<\/h2>\n

Objective function<\/strong> — The goal a model is trained to optimize.<\/p><\/div>\n

One-shot learning<\/strong> — Giving a model one example and asking it to follow the pattern.<\/p><\/div>\n

Open model<\/strong> — A model whose code, weights, or design may be publicly available, depending on the meaning of “open.”<\/p><\/div>\n

Open-source AI<\/strong> — AI software released with permissions that let others inspect, use, modify, or share it.<\/p><\/div>\n

Optimization<\/strong> — The process of improving a model by reducing errors or improving performance.<\/p><\/div>\n

Output<\/strong> — What the AI produces.<\/p><\/div>\n

Overfitting<\/strong> — When a model memorizes training examples too closely and performs poorly on new examples.<\/p><\/div>\n\n

P<\/h2>\n

Parameter<\/strong> — A number inside a model that helps determine its behavior.<\/p><\/div>\n

Pattern recognition<\/strong> — Finding meaningful regularities in data.<\/p><\/div>\n

Personalization<\/strong> — Adapting AI behavior to a specific user, organization, style, or context.<\/p><\/div>\n

Pipeline<\/strong> — A sequence of steps that data or tasks move through.<\/p><\/div>\n

Pretraining<\/strong> — The first large training stage where a model learns general patterns from broad data.<\/p><\/div>\n

Precision<\/strong> — Of the items the AI marked as positive, how many were actually positive.<\/p><\/div>\n

Prediction<\/strong> — An AI system’s estimate of what is likely to be true or happen.<\/p><\/div>\n

Predictive AI<\/strong> — AI used to forecast or classify rather than generate new content.<\/p><\/div>\n

Prompt<\/strong> — The instruction or input given to an AI system.<\/p><\/div>\n

Prompt engineering<\/strong> — Writing or structuring prompts to get better AI outputs.<\/p><\/div>\n

Prompt injection<\/strong> — An attack where hidden or malicious instructions try to override the user’s or developer’s intended instructions.<\/p><\/div>\n

Provenance<\/strong> — The origin or history of data, content, or an AI output.<\/p><\/div>\n\n

Q<\/h2>\n

Quantization<\/strong> — A technique that makes models smaller or faster by using less precise numbers.<\/p><\/div>\n

Query<\/strong> — A question or request sent to a search system, database, or AI system.<\/p><\/div>\n\n

R<\/h2>\n

RAG (Retrieval-Augmented Generation)<\/strong> — A method where an AI searches documents or databases before answering, so its answer can be based on supplied sources rather than memory alone.<\/p><\/div>\n

Recall<\/strong> — Of all the truly positive items, how many the AI successfully found.<\/p><\/div>\n

Recommendation system<\/strong> — AI that suggests items, such as movies, products, songs, articles, or posts.<\/p><\/div>\n

Red teaming<\/strong> — Testing an AI system by trying to make it fail or behave unsafely.<\/p><\/div>\n

Regression<\/strong> — Predicting a number, such as price, risk, temperature, or demand.<\/p><\/div>\n

Reinforcement learning<\/strong> — Training through rewards and penalties.<\/p><\/div>\n

RLHF (Reinforcement Learning from Human Feedback)<\/strong> — Training a model to behave better using human preferences or ratings.<\/p><\/div>\n

Representation<\/strong> — The internal way a model encodes information.<\/p><\/div>\n

Responsible AI<\/strong> — Building and using AI in ways that are ethical, safe, accountable, fair, and legally compliant.<\/p><\/div>\n

Retrieval<\/strong> — Finding relevant information from a database, document collection, or search index.<\/p><\/div>\n

Robotics<\/strong> — The field of building machines that can sense, move, and act in the physical world.<\/p><\/div>\n\n

S<\/h2>\n

Safety filter<\/strong> — A system that blocks or changes unsafe, disallowed, or inappropriate outputs.<\/p><\/div>\n

Scalability<\/strong> — The ability of a system to keep working well as usage, data, or complexity grows.<\/p><\/div>\n

Semantic search<\/strong> — Search based on meaning rather than exact keywords.<\/p><\/div>\n

Semi-supervised learning<\/strong> — Training with a mix of labeled and unlabeled data.<\/p><\/div>\n

Sentiment analysis<\/strong> — AI that estimates emotion, attitude, or tone in text.<\/p><\/div>\n

SLM (Small Language Model)<\/strong> — A language model smaller than large frontier models, often cheaper and easier to run.<\/p><\/div>\n

Speech recognition<\/strong> — AI that turns spoken language into text.<\/p><\/div>\n

Synthetic data<\/strong> — Artificially generated data used for training, testing, or privacy protection.<\/p><\/div>\n

System prompt<\/strong> — High-level instructions that shape how an AI assistant behaves.<\/p><\/div>\n\n

T<\/h2>\n

Temperature<\/strong> — A setting that affects how random or creative a model’s output is. Lower temperature is usually more predictable; higher temperature is usually more varied.<\/p><\/div>\n

Test set<\/strong> — Data held back to evaluate how well a model performs after training.<\/p><\/div>\n

Text-to-image model<\/strong> — AI that creates images from written descriptions.<\/p><\/div>\n

Token<\/strong> — A chunk of text processed by a language model. A token may be a word, part of a word, punctuation, or another text fragment.<\/p><\/div>\n

Token limit<\/strong> — The maximum number of tokens a model can process or produce.<\/p><\/div>\n

Tool use<\/strong> — When an AI model calls outside tools, such as search, calculators, calendars, databases, code interpreters, or file systems.<\/p><\/div>\n

Training<\/strong> — The process by which a model learns patterns from data.<\/p><\/div>\n

Training data<\/strong> — The information used to teach a model.<\/p><\/div>\n

Transformer<\/strong> — A neural network architecture that made modern large language models possible by using attention to process relationships among tokens.<\/p><\/div>\n

Transparency<\/strong> — Making enough information available about an AI system for people to understand, evaluate, or challenge it.<\/p><\/div>\n

Turing Test<\/strong> — A famous proposed test of whether a machine can imitate human conversation well enough to be mistaken for human.<\/p><\/div>\n\n

U<\/h2>\n

Uncertainty<\/strong> — How unsure a model or system is.<\/p><\/div>\n

Underfitting<\/strong> — When a model is too simple or poorly trained to capture the patterns in the data.<\/p><\/div>\n

Unlabeled data<\/strong> — Data without correct answers attached.<\/p><\/div>\n

Unsupervised learning<\/strong> — Learning patterns from data without labeled answers.<\/p><\/div>\n

User prompt<\/strong> — The specific message or request from the user.<\/p><\/div>\n\n

V<\/h2>\n

Validation set<\/strong> — Data used during development to tune and compare models before final testing.<\/p><\/div>\n

Vector<\/strong> — A list of numbers that represents something, such as a word, sentence, image, or user preference.<\/p><\/div>\n

Vector database<\/strong> — A database designed to store and search embeddings.<\/p><\/div>\n

Vector search<\/strong> — Searching by mathematical similarity between embeddings.<\/p><\/div>\n

Vision-language model<\/strong> — A model that can work with both images and text.<\/p><\/div>\n\n

W<\/h2>\n

Weights<\/strong> — The learned internal values of a model.<\/p><\/div>\n

Workflow automation<\/strong> — Using AI and software tools to complete multi-step tasks.<\/p><\/div>\n

World model<\/strong> — A model’s internal representation of how things relate or behave. In AI discussion, this term can be technical, speculative, or metaphorical depending on context.<\/p><\/div>\n\n

X<\/h2>\n

XAI (Explainable AI)<\/strong> — Methods for making AI decisions more understandable to people.<\/p><\/div>\n\n

Z<\/h2>\n

Zero-shot learning<\/strong> — Asking a model to do a task without giving examples in the prompt.<\/p><\/div>\n\n<\/div>\n\n