AIAI key terms

The broad field of building computer systems that perform tasks normally requiring human intelligence, such as reasoning, learning, perception, and language use. It spans many techniques, not just machine learning.

A branch of AI where systems learn patterns from data rather than following hand-written rules. A trained model generalizes from examples to make predictions or decisions on new inputs.

A machine learning approach using neural networks with many layers. It automatically learns hierarchical features from raw data such as pixels or text, powering most modern speech, vision, and language systems.

A computing structure of interconnected nodes arranged in layers, loosely inspired by neurons. Each node applies weights and an activation function, and training adjusts those weights to reduce error.

The artifact produced by training: a set of learned parameters plus the logic needed to turn inputs into outputs. A model is what gets evaluated, deployed, and monitored in production.

The labeled or unlabeled examples used to fit a model's parameters. Its size, quality, and representativeness strongly determine how well the resulting model performs on real inputs.

A held-out portion of data used to tune model settings and compare candidate models during development. It is not used to fit parameters, helping detect overfitting before final testing.

Data held back entirely from training and tuning, used once at the end to estimate how a finished model will perform on unseen inputs. It provides the fairest performance estimate.

When a model memorizes training data, including its noise, and performs well on training examples but poorly on new ones. It typically signals too much model complexity or too little data.

When a model is too simple or poorly trained to capture the underlying pattern, so it performs badly on both training and new data. The fix is usually a richer model or better features.

A model's ability to perform well on inputs it did not see during training. It is the central goal of machine learning and is estimated using validation and test data.

Training on examples that pair inputs with correct answers, so the model learns to predict those answers for new inputs. Common tasks are classification and regression.

Training on data without correct answers, to discover structure such as clusters, groupings, or reduced representations. Clustering and dimensionality reduction are typical examples.

A supervised task predicting which category an input belongs to, such as spam versus not spam. The model outputs a class label, often with a confidence score per class.

A supervised task predicting a continuous numeric value, such as price or temperature. Performance is commonly measured with error metrics like mean absolute error.

A setting chosen before training that controls how learning proceeds, such as learning rate or tree depth. Unlike model parameters, hyperparameters are not learned from the data itself.

The practice of selecting, transforming, and creating input variables so a model can learn patterns more easily. Good features often matter more than model choice.

An individual measurable input variable used by a model, such as age, word count, or pixel value. Rows of features form the examples a model learns from.

Data where each example carries the correct answer or target value. Labels enable supervised learning but are often expensive, since humans usually must produce them.

The process of assigning correct answers to raw examples so they can train supervised models. Quality and consistency of labels directly affect model accuracy and fairness.

How fit a dataset is for its intended use, covering accuracy, completeness, consistency, timeliness, and relevance. Poor quality data produces unreliable models regardless of algorithm choice.

Detecting and correcting errors in data, such as duplicates, missing values, inconsistent formats, and outliers. It is usually a large share of preparation effort before modeling.

Rescaling numeric features to a common range or distribution so no variable dominates learning because of its units. It helps many algorithms converge faster and behave more predictably.

When information from the target or from the test period accidentally enters training features, inflating measured performance. The model looks accurate in testing but fails in production.

AI systems that create new content such as text, images, audio, or code rather than only labeling or scoring existing inputs. Most modern examples are built on large neural networks trained on vast data.

A neural network trained on enormous text corpora to predict and generate language. It can answer questions, summarize, and write code, but has no guaranteed factual grounding.

The instruction or input text given to a generative model to steer its output. Clear, specific prompts with context and format guidance usually produce more useful results.

The practice of designing and refining model inputs to reliably obtain desired outputs. Techniques include giving examples, specifying format, and breaking complex tasks into steps.

When a generative model produces fluent output that is factually wrong or unsupported, stated with false confidence. It happens because the model predicts plausible text rather than verifying facts.

A pattern where relevant documents are retrieved from a knowledge source and supplied to a language model so its answer is grounded in that content. It reduces hallucination and keeps answers current.

Tying a model's output to verifiable sources or supplied context so claims can be checked. Grounded responses cite or reflect retrieved evidence instead of relying only on learned patterns.

A small unit of text, such as a word fragment, that a language model processes. Prompts and outputs are measured in tokens, which drives context limits and API cost.

The maximum amount of text, measured in tokens, that a language model can consider at once. Content beyond it is dropped, so long inputs must be trimmed or summarized.

Systematic unfair skew in data, models, or decisions that disadvantages particular groups. In AI it can arise from unrepresentative training data or from choices made during design.

The goal that an AI system treats people and groups equitably, without unjustified disparities in outcomes or error rates. It is measured with metrics and requires explicit definition per context.

The degree to which people can understand why an AI system produced a given output. It supports trust, accountability, debugging, and regulatory compliance, especially in high-stakes decisions.

A design where people review, approve, or correct AI outputs before they take effect. It is a key control for high-risk uses where automated errors carry serious consequences.

The policies, roles, and controls that manage how AI models are approved, documented, monitored, and retired. It ensures accountability and compliance across the model lifecycle.

Protecting personal information from unauthorized collection, use, or disclosure. In AI it covers consent, minimization, retention limits, and safeguards applied to training and inference data.

Removing or altering identifying details so individuals cannot reasonably be re-identified from a dataset. Done well, it lets organizations use data while reducing privacy risk.

Deliberately crafted inputs designed to make an AI model misbehave, such as misclassifying an image or leaking instructions. Defenses include input validation, monitoring, and robust training.

An attack where malicious text in a model's input or retrieved content hijacks its instructions, causing unintended actions or data exposure. It is a leading risk for LLM applications.

The gradual decline in a deployed model's performance as real-world data diverges from what it learned. Monitoring inputs and outcomes triggers retraining or replacement when drift is detected.

Using a trained model to produce predictions or generated outputs on new inputs. It is the production phase, distinct from the training phase that created the model.

Moving a trained model into a live environment where applications can call it, often behind an API. Deployment includes packaging, scaling, access control, and rollback planning.

Practices and tooling for reliably building, shipping, and maintaining machine learning systems. It extends DevOps with data versioning, experiment tracking, and continuous model monitoring.

Continuously tracking a deployed model's inputs, outputs, errors, and business outcomes to catch failures, drift, or misuse. Alerts prompt investigation, retraining, or rollback.

A specific business problem where AI could add value, defined by the users, the decision or task, and the measurable outcome. Viable use cases have clear data and success criteria.

A small, time-boxed build that tests whether an AI approach can meet a defined need with available data. It reduces uncertainty before larger investment in a full solution.

The financial benefit of an AI initiative relative to its cost, including data work, compute, integration, and ongoing operations. It justifies prioritization among candidate use cases.