AI glossary Terms & Defintions: About Concepts, Methods and Limits

Search and algorithms

A* search

How it works

A* orders candidate paths using f = g + h. Optimality depends on assumptions about costs, the heuristic and how repeated states are handled.

Example

Find a low-cost route using straight-line distance as a lower bound.

Limits and common confusions

An inadmissible heuristic or incorrect duplicate handling can invalidate optimality guarantees.

Related: Search algorithm

Editorial review: 2026-10-14 · Link to this term

Knowledge and reasoning

Abductive reasoning

How it works

Abduction asks which hypotheses, together with background knowledge, could account for evidence. Several explanations may remain plausible, so it differs from deductively proving a conclusion.

Example

Infer a possible equipment fault from observed symptoms.

Limits and common confusions

An explanation consistent with evidence is not thereby confirmed as true.

Related: Knowledge representation and reasoning

Editorial review: 2026-10-14 · Link to this term

Evaluation

Ablation study

How it works

Compare otherwise controlled variants with a suitable baseline. Retraining may be necessary to distinguish a component's role during learning from its role during inference.

Example

Remove a retrieval component and compare grounded-answer accuracy.

Limits and common confusions

Interactions between components mean that one ablation does not isolate every causal contribution.

Related: Model evaluation

Editorial review: 2026-10-14 · Link to this term

Computing foundations

Abstract data type

How it works

The contract describes observable behavior without choosing a particular implementation. Different implementations can satisfy the same abstract interface.

Example

A stack supports pushing and popping in last-in, first-out order.

Limits and common confusions

An abstract specification does not determine performance or memory layout.

Related: Algorithm

Editorial review: 2026-10-14 · Link to this term

Knowledge and reasoning

Abstraction

How it works

Abstraction can reduce reasoning cost by grouping states or replacing complex objects with simpler features. Its usefulness depends on preserving distinctions needed for the task.

Example

Model a delivery location as a graph node instead of its complete geometry.

Limits and common confusions

Omitted details can invalidate a plan or prediction when assumptions change.

Related: Knowledge representation and reasoning

Editorial review: 2026-10-14 · Link to this term

Evaluation

Accuracy

In plain language

Accuracy counts correct answers divided by all answers.

Technical definition

For single-label classification it is the mean indicator of predicted-label equality.

How it works

Always state the dataset and label rules. Compare with simple baselines and inspect which classes are missed; one aggregate score can conceal important differences.

Example

Ninety correct labels out of one hundred examples gives 90% accuracy.

Limits and common confusions

If 99% of examples are negative, always predicting negative achieves 99% accuracy while missing every positive.

Related: Precision and recall · Classification

Editorial review: 2026-10-13 · Link to this term

Agents and systems

Action selection

How it works

A controller or policy maps available information to actions. Selection can use rules, planning, learned values or other mechanisms.

Example

Choose whether a robot should wait, turn or move.

Limits and common confusions

Choosing an action differs from successfully executing it in the environment.

Related: AI agent

Editorial review: 2026-10-14 · Link to this term

Models and learning

Activation function

How it works

Nonlinear activations let compositions of layers represent relationships beyond a single affine map. Different functions have different gradients and output ranges.

Example

Apply a rectified linear function after a weighted sum.

Limits and common confusions

Activation choice affects optimization and expressiveness but does not alone determine model quality.

Related: Machine learning

Editorial review: 2026-10-14 · Link to this term

Agents and systems

Adaptive AI

How it works

Adaptation may involve online updates, retraining, context changes or a controller selecting different actions. These mechanisms have different persistence and safety implications and should be described explicitly.

Example

Update a classifier when the operating environment changes.

Limits and common confusions

Adaptive does not mean that every interaction changes the model or that earlier skills are preserved.

Related: AI agent

Editorial review: 2026-10-14 · Link to this term

Models and learning

Adaptive neuro-fuzzy inference system

ANFIS

How it works

ANFIS connects fuzzy-rule computations to a trainable network structure. The original formulation uses a hybrid learning procedure for premise and consequent parameters.

Example

Fit fuzzy-rule parameters from input-output observations.

Limits and common confusions

Learned rules can become numerous or difficult to interpret; the method does not guarantee optimal generalization.

Related: Machine learning

Editorial review: 2026-10-14 · Link to this term

Search and algorithms

Admissible heuristic

How it works

This lower-bound property supports optimality arguments for informed search under the algorithm's other assumptions. It says nothing about how informative the estimate is.

Example

Zero is admissible when all remaining costs are nonnegative.

Limits and common confusions

Admissibility alone does not make search fast.

Related: Search algorithm

Editorial review: 2026-10-14 · Link to this term

Reliability and security

Adversarial example

In plain language

A carefully chosen change can mislead a model even when a person sees little relevant difference.

Technical definition

The cited work studies perturbations that increase classification loss and develops a gradient-based construction. Robustness claims depend on the allowed perturbations and adversary's capabilities.

How it works

A threat model states what the adversary can change and observe. Evaluation must distinguish random noise robustness from performance against deliberately optimized inputs.

Example

In an authorized test, slightly perturb an image and measure whether the classifier changes its label.

Limits and common confusions

Success against one model or perturbation set does not generalize automatically. A failed attack is not proof of robustness.

Related: Classification · Neural network · Prompt injection

Editorial review: 2026-10-13 · Link to this term

Agents and systems

Agent architecture

How it works

Architectures specify how information flows between mechanisms such as reactive control, planning and memory. Hierarchical designs separate decisions at different timescales.

Example

Combine a route planner with a fast obstacle-avoidance controller.

Limits and common confusions

An architecture alone does not determine the quality of its models or objectives.

Related: AI agent

Editorial review: 2026-10-14 · Link to this term

Agents and systems

Agent memory

How it works

Memory can be stored notes, structured state, logs or retrievable records. A later step may insert selected records into the model's context; this is distinct from retraining the model.

Example

Store task progress so an interrupted workflow can resume.

Limits and common confusions

Stored errors can persist, and retained personal information needs appropriate access and deletion controls.

Related: AI agent

Editorial review: 2026-10-14 · Link to this term

Agents and systems

Agentic AI

How it works

An agentic system combines a model with tools, observations and a control loop. The degree of autonomy varies; fixed workflows and dynamically chosen action sequences are different architectures.

Example

A research assistant chooses which document to retrieve next.

Limits and common confusions

Autonomy needs explicit permissions, stopping conditions and recovery; the label does not imply general intelligence.

Related: AI agent

Editorial review: 2026-10-14 · Link to this term

Foundations

AI agent

In plain language

An agent observes a situation and does something in response.

Technical definition

Its decision rule connects observations or internal state to actions under specified objectives.

How it works

Software actions can include requesting information or changing a file; physical actions can move a robot. Specify what the system can observe, which tools it can use and when a person must approve an action.

Example

A glossary assistant identifies an unsupported definition and requests a source search.

Limits and common confusions

Calling software an agent does not imply autonomy, reliable judgment or human intentions.

Related: Artificial intelligence

Editorial review: 2026-10-13 · Link to this term

Responsible AI

AI alignment

How it works

Objectives, feedback and safeguards can encourage desired behavior, but intended values may conflict or be incompletely specified. Alignment concerns actual behavior under relevant conditions, not merely a polite answer style.

Example

Check whether a travel agent respects a budget while booking.

Limits and common confusions

Agreement with a user is not proof of truthfulness, safety or alignment.

Related: Responsible AI

Editorial review: 2026-10-14 · Link to this term

Human-AI interaction

AI copilot

How it works

The term is a product metaphor rather than a technical architecture. Effective assistance makes capabilities and limits clear and lets users review, correct or decline consequential suggestions.

Example

Suggest an email draft that the sender reviews before sending.

Limits and common confusions

Calling a product a copilot does not establish meaningful human control.

Related: Human-computer interaction

Editorial review: 2026-10-14 · Link to this term

Agents and systems

AI engineering

How it works

The work spans data pipelines, models, evaluation, integration, deployment and maintenance. Product behavior depends on these surrounding components as well as model quality.

Example

Version a retrieval index and test it together with the answering model.

Limits and common confusions

A successful model experiment is not yet a reliable production service.

Related: AI agent

Editorial review: 2026-10-14 · Link to this term

Responsible AI

AI ethics

How it works

Ethical analysis asks who benefits, who bears risks, what rights are affected and whose judgments count. It includes decisions about whether an application should be built, not only how accurately it predicts.

Example

Examine consent and power imbalances before collecting training data.

Limits and common confusions

Ethical disagreement cannot always be resolved by optimizing one numerical score.

Related: Responsible AI

Editorial review: 2026-10-14 · Link to this term

Human-AI interaction

AI fluency

How it works

Fluency includes choosing suitable tasks, communicating requirements, evaluating results and taking responsibility for their use. It develops through practice and judgment rather than memorizing a particular interface.

Example

Check an assistant's cited source before using its summary.

Limits and common confusions

Familiarity with prompting alone does not establish subject expertise or make outputs reliable.

Related: Human-computer interaction

Editorial review: 2026-10-14 · Link to this term

Responsible AI

AI guardrails

How it works

Examples include permissions, input checks, output validation and escalation rules. Controls may sit around a model or be incorporated into its training; layered design addresses different failure modes.

Example

Block a tool call that targets a folder outside the user's authorized scope.

Limits and common confusions

Guardrails can fail or be bypassed; a prompt instruction alone is not a reliable security boundary.

Related: Responsible AI

Editorial review: 2026-10-14 · Link to this term

Responsible AI

AI safety

How it works

Safety involves identifying hazards, evaluating likely failures and applying controls throughout a system's life. Reliability, security and fairness contribute but are not interchangeable with safety.

Example

Require confirmation before an assistant changes a production database.

Limits and common confusions

Passing a benchmark does not establish safety for every deployment context.

Related: Responsible AI

Editorial review: 2026-10-14 · Link to this term

Human-AI interaction

AI slop

How it works

The phrase usually concerns generic, inaccurate or excessive content that adds little value. It describes a quality judgment rather than a detectable model architecture or a formal scientific category.

Example

Publish many repetitive pages that do not answer readers' questions.

Limits and common confusions

AI involvement alone does not determine quality, and appearance cannot reliably establish how content was made.

Related: Human-computer interaction

Editorial review: 2026-10-14 · Link to this term

Responsible AI

AI washing

How it works

The concern is the gap between what is represented and what the system demonstrably does. Claims about automation, predictive power or sophistication should be supported with relevant evidence.

Example

Advertise a rule-based feature as a uniquely capable learning system without substantiation.

Limits and common confusions

The term is descriptive; evaluating a particular claim requires the actual facts and context.

Related: Responsible AI

Editorial review: 2026-10-14 · Link to this term

Agents and systems

AI workflow

How it works

A workflow determines how inputs, model calls, tools and checks are connected. Routing or parallel steps may be predefined even when individual model outputs are unpredictable.

Example

Extract invoice fields, validate totals, then request human approval.

Limits and common confusions

A fixed workflow is not automatically an autonomous agent; errors can propagate between steps.

Related: AI agent

Editorial review: 2026-10-14 · Link to this term

Foundations

Algorithm

In plain language

An algorithm describes how to do something systematically, such as sorting names or searching a collection.

Technical definition

Its specification defines permitted inputs, operations and outputs. Deterministic and randomized procedures differ in whether randomness can affect their execution.

How it works

An algorithm is distinct from its implementation and from the data it processes. A learning algorithm produces model parameters; a prediction procedure uses those parameters.

Example

Binary search repeatedly reduces the candidate interval in a sorted list.

Limits and common confusions

Calling something an algorithm does not establish correctness, efficiency or intelligence. Those properties need separate analysis or evidence.

Related: Machine learning · Heuristic · Optimization

Editorial review: 2026-10-13 · Link to this term

Computing foundations

Algorithmic efficiency

How it works

Time and memory are common resources, but communication or energy may also matter. Comparisons require equivalent tasks and explicit assumptions.

Example

Choose a sparse representation to reduce memory and work.

Limits and common confusions

Improving one resource can increase another; efficiency is not a single universal score.

Related: Algorithm

Editorial review: 2026-10-14 · Link to this term

Data and evaluation

Algorithmic fairness

In plain language

It examines who benefits, who is harmed and whether relevant groups are treated appropriately.

Technical definition

Formal criteria measure different properties. Equality of opportunity compares true-positive rates across groups for a specified positive outcome; other criteria compare different rates or distributions.

How it works

Choose criteria with attention to the context, labels and affected people. A statistical constraint makes one objective explicit, rather than supplying a universal ethical definition.

Example

Compare missed-positive rates across groups in a clearly defined evaluation dataset.

Limits and common confusions

Biased labels or unrepresentative samples can undermine an audit. Satisfying one metric does not demonstrate overall fairness.

Related: Precision and recall · Classification · Probability calibration

Editorial review: 2026-10-13 · Link to this term

Computing foundations

Analysis of algorithms

How it works

Analysis relates behavior to input size, assumptions and sometimes input distributions. It complements experiments by explaining why performance scales in a particular way.

Example

Count comparisons in a sorting procedure.

Limits and common confusions

An analysis under unrealistic assumptions can mispredict deployed performance.

Related: Algorithm

Editorial review: 2026-10-14 · Link to this term

Knowledge and reasoning

Answer set programming

ASP

How it works

Rules and constraints describe acceptable solutions. An answer-set solver searches for stable models, rather than simply following Prolog's execution strategy.

Example

Encode scheduling choices and constraints, then ask a solver for valid schedules.

Limits and common confusions

Grounding can be large, and the chosen stable-model semantics must match the intended problem.

Related: Knowledge representation and reasoning

Editorial review: 2026-10-14 · Link to this term

Search and algorithms

Ant colony optimization

ACO

How it works

Artificial agents construct candidate solutions using learned desirability signals and heuristics. Evaporation and reinforcement balance exploration with repeated use of promising components.

Example

Search for low-cost tours through a set of locations.

Limits and common confusions

The biological metaphor does not guarantee a global optimum or superiority over other methods.

Related: Search algorithm

Editorial review: 2026-10-14 · Link to this term

Human interaction

Anthropomorphism

In plain language

A fluent voice or friendly message can make software seem as though it has human feelings or intentions.

Technical definition

In AI interaction, conversational framing and person-like presentation can influence users' interpretations and trust. Behavioral appearance and evidence about internal capacities should be assessed separately.

How it works

Human-like design can make an interface approachable, but explanations should describe actual capabilities and limitations. The cited paper discusses opportunities and risks rather than providing a consciousness test.

Example

A chatbot says it is worried; the wording alone does not establish that it experiences worry.

Limits and common confusions

Avoid treating a social impression as proof of understanding, concern or reliability.

Related: Chatbot · Artificial intelligence

Editorial review: 2026-10-13 · Link to this term

Search and algorithms

Anytime algorithm

How it works

The contract specifies when an initial result is available and how quality changes over time. Such methods support decisions under limited computation budgets.

Example

Return the best route found so far when a deadline expires.

Limits and common confusions

A usable partial result is not necessarily optimal or accompanied by a tight error bound.

Related: Search algorithm

Editorial review: 2026-10-14 · Link to this term

Software and openness

Application programming interface

API

How it works

An API specifies operations, inputs, outputs and relevant behavior. It may be a local library interface or a network service; HTTP APIs are only one kind.

Example

Call a documented function to retrieve a record by identifier.

Limits and common confusions

A documented API still requires error handling, version awareness and appropriate authorization.

Related: Open-source software

Editorial review: 2026-10-14 · Link to this term

Search and algorithms

Approximate string matching

How it works

Insertions, deletions and substitutions can define edit distance. Search may target whole records or substrings, with different algorithms and guarantees.

Example

Find a misspelled name within a database.

Limits and common confusions

Textual similarity does not establish semantic equivalence or identity.

Related: Search algorithm

Editorial review: 2026-10-14 · Link to this term

Models and learning

Approximation error

How it works

In learning, the chosen hypothesis family may be unable to express the true relationship even with abundant data. This differs from estimation error due to limited observations.

Example

A straight line cannot exactly represent a curved relation.

Limits and common confusions

Observed prediction error also includes other sources; it cannot all be attributed to approximation.

Related: Machine learning

Editorial review: 2026-10-14 · Link to this term

Foundations

Artificial general intelligence

AGI

In plain language

The emphasis is flexible competence beyond one narrow task.

Technical definition

Proposed frameworks distinguish breadth, performance and autonomy rather than using one score.

How it works

Definitions and thresholds vary; state the chosen criteria before assessing a claim.

Example

Evaluate transfer across unfamiliar tasks rather than only repeated versions of one benchmark.

Limits and common confusions

A high benchmark score alone does not settle whether a system is AGI.

Related: Narrow AI · Artificial intelligence

Editorial review: 2026-10-13 · Link to this term

Models and learning

Artificial immune system

AIS

How it works

Examples use abstractions of recognition, diversity, clonal selection or immune memory for learning and optimization. The computational objective is separate from biological fidelity.

Example

Detect unusual patterns using an immune-inspired recognition scheme.

Limits and common confusions

Biological inspiration alone does not establish detection accuracy or robustness.

Related: Machine learning

Editorial review: 2026-10-14 · Link to this term

Foundations

Artificial intelligence

AI

In plain language

AI includes systems that learn, reason, plan or perceive.

Technical definition

An agent-centered account studies computational action under goals, observations and resource constraints.

How it works

Learning is one approach within AI. A rule-based planner can also be an AI system. Evaluate a system against its task and operating conditions rather than assuming that fluent output proves understanding.

Example

A warehouse planner chooses a route through available aisles.

Limits and common confusions

Task success does not establish consciousness or unlimited competence.

Related: AI agent · Machine learning

Editorial review: 2026-10-13 · Link to this term

History and concepts

Artificial superintelligence

ASI

In plain language

The idea concerns broad superiority, not merely being faster at one calculation or game.

Technical definition

Good's historical essay considers an ultraintelligent machine able to improve machine design and speculates about further capability growth. This is an argument about possibilities and assumptions.

How it works

Separate the definition, feasibility argument and predicted timeline. They are different claims requiring different evidence; the historical essay is not a demonstration of a realized system.

Example

A thought experiment asks whether superior machine-design ability could accelerate subsequent technical improvements.

Limits and common confusions

There is no universally accepted measurement threshold here. Do not treat speculative mechanisms or dates as established empirical facts.

Related: Artificial general intelligence · Narrow AI · Deep Blue

Editorial review: 2026-10-13 · Link to this term

Computing foundations

Asymptotic computational complexity

How it works

Asymptotic analysis suppresses lower-order terms and constant factors to compare scaling. The resource can be time, space or another explicitly defined quantity.

Example

Describe an algorithm's large-input growth with a tight asymptotic bound.

Limits and common confusions

The asymptotically better method may be slower for small practical inputs.

Related: Algorithm

Editorial review: 2026-10-14 · Link to this term

Architectures

Attention

Attention mechanism

In plain language

A query controls how information from available values is combined.

Technical definition

In scaled dot-product attention, softmax-normalized query–key scores weight value vectors.

How it works

Scaled dot-product attention compares a query with keys, divides the scores by the square root of the key dimension, and normalizes them with softmax. The resulting weights combine value vectors. Queries, keys and values are learned representations, not necessarily words or human-readable concepts.

Example

A translation model can weight several source tokens while producing one target token.

Limits and common confusions

A large weight is not, by itself, a causal explanation of a model decision. Attention is a computation, not human awareness.

Related: Transformer

Editorial review: 2026-10-13 · Link to this term

Responsible AI

Auditable AI

How it works

Versioned data and model records, evaluation results, decisions and access-controlled logs help reviewers trace what occurred. Model cards document intended uses and limitations but are only one component of an audit trail.

Example

Reconstruct which model version produced a disputed recommendation.

Limits and common confusions

A log is evidence of recorded events, not proof that a system is fair or correct.

Related: Responsible AI

Editorial review: 2026-10-14 · Link to this term

Human interaction

Augmented intelligence

Human-AI assistance

In plain language

The person and software work together, with useful controls for review and correction.

Technical definition

This describes a human-AI interaction arrangement rather than a distinct learning algorithm. Assistance should communicate capabilities, uncertainty and opportunities to correct errors.

How it works

The linked human-AI guidelines support designing such interactions. Evaluate the combined workflow: a model's isolated score does not establish that users perform better with it.

Example

An editor receives proposed tags, checks their evidence and accepts or changes them before publication.

Limits and common confusions

A nominal human approval step may provide little oversight if time, information or authority to challenge the result is absent.

Related: AI agent · Anthropomorphism · Machine learning

Editorial review: 2026-10-13 · Link to this term

Models and learning

Autoencoder

How it works

An encoder produces a code and a decoder attempts reconstruction. Bottlenecks, noise or other constraints encourage useful structure rather than unrestricted copying; the code need not have fewer dimensions than the input.

Example

Learn representations by reconstructing deliberately corrupted images.

Limits and common confusions

Low reconstruction error alone does not establish useful features or meaningful generative sampling.

Related: Machine learning

Editorial review: 2026-10-14 · Link to this term

Computing foundations

Automata theory

How it works

Finite automata, pushdown automata and Turing machines model different computational capabilities. Their state and memory rules determine expressive power.

Example

Use a finite-state machine to recognize a simple input pattern.

Limits and common confusions

Finite-state models cannot recognize every formal language.

Related: Algorithm

Editorial review: 2026-10-14 · Link to this term

Models and learning

Automated machine learning

AutoML

How it works

AutoML can search preprocessing, model families and hyperparameters under resource constraints. It does not remove the need to define targets, appropriate data and deployment requirements.

Example

Compare candidate pipelines within a fixed computation budget.

Limits and common confusions

Automation can optimize the wrong metric or exploit data leakage if the task setup is flawed.

Related: Machine learning

Editorial review: 2026-10-14 · Link to this term

Knowledge and reasoning

Automated planning

How it works

A planning problem describes states, actions, their conditions and effects, and a goal. Classical planning assumes a relatively well-specified world; richer approaches handle uncertainty and partial observation.

Example

Plan the steps needed to move packages between locations.

Limits and common confusions

A valid plan in a model can fail when the model omits real constraints or effects.

Related: Knowledge representation and reasoning

Editorial review: 2026-10-14 · Link to this term

Knowledge and reasoning

Automated reasoning

How it works

The task may involve theorem proving, satisfiability, constraint solving or probabilistic inference. Soundness and completeness are properties of specified systems, not of every generated explanation.

Example

Check whether a proposed conclusion follows from encoded premises.

Limits and common confusions

Correct formal inference cannot repair false premises or a misrepresented question.

Related: Knowledge representation and reasoning

Editorial review: 2026-10-14 · Link to this term

Language and speech

Automatic speech recognition

Speech-to-text · ASR

In plain language

Speech-to-text estimates what was said from the audio signal.

Technical definition

An ASR system maps acoustic observations to text, often combining learned acoustic and linguistic regularities. Word error rate measures substitutions, deletions and insertions against a reference.

How it works

The reference transcript and normalization rules affect the measured error rate. Evaluate relevant microphones, background noise, accents and vocabulary rather than only a clean benchmark.

Example

Generate an editable transcript from a recorded interview.

Limits and common confusions

Transcription can omit or invent words, especially with noise. Identifying words is distinct from identifying the speaker.

Related: Speaker recognition · Natural language processing

Editorial review: 2026-10-13 · Link to this term

Agents and systems

Automation

How it works

Automation can be rule-based, mechanical or learned; it is not synonymous with AI. Delegating a task changes how people supervise, recover from failures and handle exceptions.

Example

Automatically route validated records to a processing queue.

Limits and common confusions

Automating a flawed process can scale its mistakes rather than improve it.

Related: AI agent

Editorial review: 2026-10-14 · Link to this term

Human-AI interaction

Automation bias

How it works

Users may accept incorrect advice or fail to notice issues the automation does not flag. Interface design, workload and expertise influence this behavior.

Example

Accept a wrong route because the navigation system recommended it.

Limits and common confusions

Showing an explanation alone may not reduce inappropriate reliance.

Related: Human-computer interaction

Editorial review: 2026-10-14 · Link to this term

Training

Backpropagation

In plain language

It measures how a small change in each adjustable value would affect the loss.

Technical definition

Reverse-mode differentiation applies the chain rule to local derivatives, accumulating contributions along every relevant path in a computational graph.

How it works

Backpropagation computes gradients; an optimizer uses them to update weights. Separating these roles helps distinguish differentiation errors from learning-rate or optimization problems.

Example

After a classifier makes an incorrect prediction, gradients indicate how its weights affect that example's loss.

Limits and common confusions

Gradients can vanish or explode. A computed gradient does not guarantee a good update or a global optimum.

Related: Neural network · Stochastic gradient descent · Weight initialization

Editorial review: 2026-10-13 · Link to this term

Models and learning

Backpropagation through structure

BPTS

How it works

A model can apply shared transformations over a tree or another supported structure, then propagate derivatives through those applications. The computational graph determines valid dependencies.

Example

Train a recursive model over a parsed expression tree.

Limits and common confusions

The structure and parameter sharing must be specified; arbitrary cyclic computations need additional treatment.

Related: Machine learning

Editorial review: 2026-10-14 · Link to this term

Models and learning

Backpropagation through time

BPTT

How it works

Repeated uses of shared parameters become nodes in an unrolled computation graph. Gradients accumulate across steps; truncated variants limit how far gradients propagate.

Example

Train a recurrent model on a sequence of observations.

Limits and common confusions

Long dependencies can cause vanishing or exploding gradients and substantial memory cost.

Related: Machine learning

Editorial review: 2026-10-14 · Link to this term

Knowledge and reasoning

Backward chaining

How it works

A goal-driven procedure reduces a query to subgoals. Search order and cycle handling matter, particularly in logic-programming implementations.

Example

To prove eligibility, check each required condition.

Limits and common confusions

Failure of a particular search does not always establish that the goal is logically false.

Related: Knowledge representation and reasoning

Editorial review: 2026-10-14 · Link to this term

Vision and generation

Bag-of-visual-words model

How it works

Descriptors are assigned to a learned visual vocabulary, producing a histogram-like representation. The analogy to text words is representational rather than linguistic.

Example

Retrieve images using similar distributions of local visual features.

Limits and common confusions

Basic versions discard detailed spatial arrangement and depend on the visual vocabulary.

Related: Computer vision

Editorial review: 2026-10-14 · Link to this term

Language and prompting

Bag-of-words model

How it works

Tokenization and weighting determine the features. The representation is useful for simple classification and retrieval baselines but loses much sequential structure.

Example

Represent a document by its word-frequency vector.

Limits and common confusions

Texts with different meanings can have the same word counts.

Related: Natural language processing

Editorial review: 2026-10-14 · Link to this term

Models and learning

Batch normalization

How it works

Training often uses minibatch statistics; inference commonly uses accumulated estimates. The exact normalization axes and handling of statistics depend on the implementation.

Example

Normalize hidden activations during neural-network training.

Limits and common confusions

Small or unrepresentative batches and train-inference mismatch can cause problems.

Related: Machine learning

Editorial review: 2026-10-14 · Link to this term

Probability and decisions

Bayesian network

How it works

Nodes represent random variables and conditional distributions describe each variable given its parents. The graph encodes conditional-independence assumptions that support probabilistic inference.

Example

Relate equipment faults and observed symptoms probabilistically.

Limits and common confusions

An arrow is not automatically a causal claim; causal use needs additional assumptions.

Related: Probabilistic model

Editorial review: 2026-10-14 · Link to this term

Data and evaluation

Benchmark

In plain language

A benchmark creates a shared test, but its scope determines what the score means.

Technical definition

Protocols specify inputs, scoring and evaluation conditions. Multi-metric evaluations such as HELM examine dimensions beyond accuracy, exposing trade-offs across tasks.

How it works

Useful reporting includes the model version, settings, sampling procedure and uncertainty. A benchmark is evidence about tested scenarios, not a complete description of a system.

Example

Compare two models on accuracy, calibration and runtime using the same held-out tasks.

Limits and common confusions

Repeated tuning to a benchmark or contamination can inflate scores. A ranking may reverse on a different task mix.

Related: Accuracy · Cross-validation · Data leakage

Editorial review: 2026-10-13 · Link to this term

Responsible AI

Bias in AI

How it works

Bias can enter through sampling, measurement, objectives, institutional practices and human judgments. Statistical estimation bias and social bias are related vocabulary with different meanings; the relevant definition must be stated.

Example

A speech dataset underrepresents some accents.

Limits and common confusions

Removing protected attributes does not remove proxies or ensure fairness.

Related: Responsible AI

Editorial review: 2026-10-14 · Link to this term

Models and learning

Bias-variance trade-off

How it works

For appropriate loss and assumptions, expected error can be decomposed into bias, variance and noise terms. The familiar U-shaped curve is not universal across modern training regimes.

Example

Compare a rigid predictor with one that changes greatly across resampled datasets.

Limits and common confusions

Statistical bias here is different from social bias or unfair treatment.

Related: Machine learning

Editorial review: 2026-10-14 · Link to this term

Data and evaluation

Big data

In plain language

The challenge may be how much data arrives, how quickly it arrives or how varied it is.

Technical definition

NIST identifies volume, velocity, variety and variability as important drivers. Whether a workload needs a big-data architecture depends on its requirements and resource constraints.

How it works

The label describes an engineering problem rather than a fixed file-size threshold. Streaming and distributed processing address different bottlenecks from simply storing a large archive.

Example

Continuously process heterogeneous sensor events fast enough to support an operational dashboard.

Limits and common confusions

More data does not guarantee useful or representative evidence. Collection quality, sampling, privacy and computation costs still matter.

Related: Data mining · Machine learning

Editorial review: 2026-10-13 · Link to this term

Computing foundations

Big O notation

How it works

Writing f(n) = O(g(n)) means that a constant multiple of g eventually bounds f in magnitude under the usual definition. Algorithm analysis uses it to summarize scaling.

Example

A nested traversal of n by n elements takes O(n squared) operations.

Limits and common confusions

Big O is not an exact running time and does not by itself mean worst case.

Related: Algorithm

Editorial review: 2026-10-14 · Link to this term

Computing foundations

Binary tree

How it works

Left and right children define a recursive organization useful for expression trees, indexing and other algorithms. A binary search tree adds an ordering rule.

Example

Represent a binary arithmetic expression using operator nodes and operands.

Limits and common confusions

Not every binary tree is balanced or a search tree.

Related: Algorithm

Editorial review: 2026-10-14 · Link to this term

Models and learning

Boltzmann machine

How it works

Its distribution assigns probabilities through an energy function and a normalization factor. Learning generally involves expectations that may be expensive to compute.

Example

Model dependencies among binary variables through pairwise interactions.

Limits and common confusions

Sampling and partition-function computation can make training difficult.

Related: Machine learning

Editorial review: 2026-10-14 · Link to this term

Knowledge and reasoning

Boolean satisfiability problem

How it works

SAT solvers search for or rule out satisfying assignments. SAT is a decision problem; optimization variants additionally seek a best assignment under an objective.

Example

Check whether a set of configuration constraints can all hold.

Limits and common confusions

General SAT is NP-complete, although many practical instances are tractable.

Related: Knowledge representation and reasoning

Editorial review: 2026-10-14 · Link to this term

Models and learning

Boosting

How it works

Different boosting algorithms reweight examples or optimize a loss through additive updates. The procedure differs from independently fitting models and averaging them.

Example

Add weak learners to improve a classification ensemble.

Limits and common confusions

Noisy labels and excessive fitting can still impair generalization.

Related: Machine learning

Editorial review: 2026-10-14 · Link to this term

Models and learning

Bootstrap aggregating

How it works

Also called bagging, it can reduce variance when base predictors are unstable. Predictions are commonly averaged for regression or voted for classification.

Example

Average decision trees trained on different resampled datasets.

Limits and common confusions

Highly correlated base predictors limit the benefit of averaging.

Related: Machine learning

Editorial review: 2026-10-14 · Link to this term

Search and algorithms

Branching factor

How it works

Together with search depth, branching strongly influences the size of a search tree. Average, maximum and effective branching factors are different quantities.

Example

A game position may offer twenty legal moves.

Limits and common confusions

A constant branching approximation can poorly describe irregular search spaces.

Related: Search algorithm

Editorial review: 2026-10-14 · Link to this term

Search and algorithms

Brute-force search

How it works

Exhaustive enumeration can provide a simple baseline or correctness argument for finite problems. Its practicality depends on how many candidates must be examined.

Example

Try all assignments to a small set of Boolean variables.

Limits and common confusions

Candidate counts often grow exponentially; exhaustive does not mean computationally feasible.

Related: Search algorithm

Editorial review: 2026-10-14 · Link to this term

Models and learning

Capsule neural network

CapsNet

How it works

Routing mechanisms determine how information from lower-level capsules contributes to higher-level ones. The original dynamic-routing proposal used vector outputs to represent presence and instantiation properties.

Example

Represent an object's presence together with pose-related attributes.

Limits and common confusions

Capsules are an architectural family, not a proven general solution to viewpoint invariance.

Related: Machine learning

Editorial review: 2026-10-14 · Link to this term

Language and prompting

Chain-of-thought prompting

How it works

Worked examples or instructions can elicit a sequence of explanations before a final answer. These generated steps can improve some tasks, but they remain model outputs rather than a verified record of internal computation.

Example

Show a worked arithmetic example before posing a similar problem.

Limits and common confusions

Fluent intermediate steps may be wrong or unfaithful; check the result independently.

Related: Natural language processing

Editorial review: 2026-10-14 · Link to this term

Language and speech

Chatbot

In plain language

A chatbot provides a dialogue interface, whether its answers come from rules, retrieval or a generative model.

Technical definition

A dialogue system tracks conversational context and selects responses or actions. Turn-taking, clarification and repair help manage misunderstandings across an interaction.

How it works

The interface does not determine the underlying architecture. A narrow booking assistant and an open-ended language-model chat system have different capabilities and failure modes.

Example

A booking assistant asks for a missing date before offering available appointments.

Limits and common confusions

Conversational fluency can encourage unwarranted trust. Make uncertainty and action consequences clear, particularly when requests are ambiguous.

Related: Natural language processing · Large language model · AI agent

Editorial review: 2026-10-13 · Link to this term

Products and history

ChatGPT

In plain language

It presents a chat interface for interacting with AI models; the product name is distinct from any one underlying model.

Technical definition

The original research preview used a GPT-3.5-series model with supervised dialogue fine-tuning and reinforcement learning from human feedback, as described in its launch announcement.

How it works

This is historical context. Models, tools and availability evolve, so reproducible studies should record the actual model, date, settings and task rather than citing only the product name.

Example

Use a conversation to draft an explanation, then verify its claims against reliable sources.

Limits and common confusions

The launch report documented plausible but incorrect answers. Conversational polish should not be treated as factual certification.

Related: Chatbot · Large language model · Reinforcement learning from human feedback

Editorial review: 2026-10-13 · Link to this term

Models and learning

Classical machine learning

How it works

Examples include nearest-neighbor methods, decision trees and linear models. The grouping describes a practical contrast, not a strict scientific boundary; methods differ widely in complexity and interpretability.

Example

Compare a tree-based baseline with a neural model on tabular data.

Limits and common confusions

Classical methods are not necessarily simple, transparent or limited to small datasets.

Related: Machine learning

Editorial review: 2026-10-14 · Link to this term

Prediction tasks

Classification

In plain language

A classifier assigns labels such as landscape or portrait.

Technical definition

Binary, multiclass and multilabel problems differ in their allowed target structure.

How it works

A model may output scores or probabilities before a decision rule selects labels. Choosing a threshold is part of the deployed system and changes the types of errors it makes.

Example

One image receives both architecture and night-photography tags in a multilabel task.

Limits and common confusions

A class score is not automatically a calibrated probability.

Related: Supervised learning · Probability calibration

Editorial review: 2026-10-13 · Link to this term

Multimodal learning

CLIP

Contrastive Language–Image Pre-training

In plain language

CLIP learns to compare pictures with textual descriptions in a shared representation space.

Technical definition

A contrastive objective trains image and text encoders to align matching pairs relative to alternatives.

How it works

Contrastive training distinguishes matching image–text pairs from other pairs in a batch. At inference, image embeddings can be compared with embeddings of candidate textual labels. This enables a form of classification without training a new classifier on labeled examples for every target category.

Example

Compare an image with prompts such as 'a photograph of a bird' and 'a photograph of a bicycle'.

Limits and common confusions

Scores depend on wording and candidate labels. Similarity is not a calibrated probability, exhaustive scene understanding or a certificate that an image description is true.

Related: Attention

Editorial review: 2026-10-13 · Link to this term

Software and openness

Closed-source software

How it works

Users may receive a binary or access a hosted service while implementation access and modification remain restricted. The phrase describes an access model, not the quality of the software.

Example

Use a hosted prediction API without receiving its implementation.

Limits and common confusions

Closed source does not by itself imply better security, worse accuracy or a particular business model.

Related: Open-source software

Editorial review: 2026-10-14 · Link to this term

Exploratory analysis

Clustering

In plain language

Clustering helps explore collections without preassigning every item a category.

Technical definition

K-means minimizes within-cluster squared Euclidean distances; density-based methods use different assumptions.

How it works

Choose representations and scaling before interpreting groups. K-means alternates assignments and centroid updates and can settle in different local solutions. Compare stability and domain usefulness rather than treating one partition as ground truth.

Example

Cluster color palettes to find recurring combinations, then inspect representative images.

Limits and common confusions

Spherical-distance assumptions can misrepresent elongated or unevenly dense groups.

Related: Unsupervised learning · Embedding

Editorial review: 2026-10-13 · Link to this term

Search and algorithms

Combinatorial optimization

How it works

The solution space may contain assignments, subsets, sequences or graphs. Exact methods and approximations trade computational effort against solution guarantees.

Example

Choose a low-cost assignment of jobs to machines.

Limits and common confusions

A locally best solution can be far from globally optimal.

Related: Search algorithm

Editorial review: 2026-10-14 · Link to this term

Knowledge and reasoning

Commonsense knowledge

How it works

Such knowledge often includes defaults, context-sensitive categories and exceptions. Representing it requires more than storing isolated facts.

Example

Expect an unsupported object near Earth to fall unless something holds it.

Limits and common confusions

Everyday expectations vary across settings and cultures and are not universal laws.

Related: Knowledge representation and reasoning

Editorial review: 2026-10-14 · Link to this term

Knowledge and reasoning

Commonsense reasoning

How it works

Reasoning may combine deduction, plausible explanations and defeasible assumptions. Conclusions should remain revisable when new evidence appears.

Example

Infer a likely explanation for a wet pavement while considering rain and sprinklers.

Limits and common confusions

A plausible explanation is not necessarily the actual cause.

Related: Knowledge representation and reasoning

Editorial review: 2026-10-14 · Link to this term

Computing foundations

Computational complexity theory

How it works

It classifies problems by resource bounds and compares their difficulty through reductions. This differs from analyzing just one implementation.

Example

Compare polynomial-space and polynomial-time problem classes.

Limits and common confusions

A complexity classification depends on the computation model and exact problem formulation.

Related: Algorithm

Editorial review: 2026-10-14 · Link to this term

Models and learning

Computational learning theory

How it works

Results relate hypothesis classes, data, error guarantees and computational cost. Guarantees require explicit assumptions about sampling and the learning setting.

Example

Analyze how many examples are needed for a class of predictors.

Limits and common confusions

A theorem's guarantee need not hold when real data violate its assumptions.

Related: Machine learning

Editorial review: 2026-10-14 · Link to this term

Computing foundations

Computational problem

How it works

Decision problems ask yes-or-no questions; search and optimization problems ask for objects or best solutions. A problem is distinct from an algorithm that solves it.

Example

Ask for the shortest route for any valid weighted graph input.

Limits and common confusions

One solved example does not establish a method for the entire problem family.

Related: Algorithm

Editorial review: 2026-10-14 · Link to this term

Computing infrastructure

Compute in AI

How it works

The term may refer to hardware capacity, operations, accelerator time or a budget. Useful comparisons state units and include relevant memory and communication constraints.

Example

Report accelerator-hours together with hardware and workload settings.

Limits and common confusions

Operation count, elapsed time, energy and financial cost are different measurements.

Editorial review: 2026-10-14 · Link to this term

Vision and applications

Computer vision

CV

In plain language

It helps software detect objects, estimate motion or reconstruct aspects of a scene from visual observations.

Technical definition

Tasks range from image formation and geometric estimation to recognition and reconstruction. Methods combine modeling assumptions, numerical algorithms and learned representations.

How it works

A two-dimensional image only partially constrains the world that produced it. Lighting, viewpoint and occlusion make inference necessary rather than turning pixels directly into complete knowledge.

Example

Estimate camera movement across frames and locate visible objects.

Limits and common confusions

A visually plausible result can be wrong about geometry, identity or context. Evaluate the specific task and imaging conditions.

Related: Image classification · Object detection · Convolutional neural network

Editorial review: 2026-10-13 · Link to this term

Evaluation

Concept drift

How it works

A narrower statistical usage concerns a changing conditional target distribution. It differs from a change only in input frequencies, though operational terminology varies.

Example

The relation between browsing behavior and purchase likelihood changes.

Limits and common confusions

Detecting an input shift alone does not establish concept drift.

Related: Model evaluation

Editorial review: 2026-10-14 · Link to this term

Vision and generation

Conditional generation

How it works

The model represents or approximates how outputs depend on inputs such as a class label, prompt, image or partial sequence. Conditioning steers generation without guaranteeing every requested constraint.

Example

Generate a scene conditioned on a text description and a sketch.

Limits and common confusions

A requested condition can be ignored or inconsistently represented in the output.

Related: Computer vision

Editorial review: 2026-10-14 · Link to this term

Models and learning

Connectionism

How it works

In AI and cognitive modeling, connectionist systems use distributed numerical representations and interactions between units. Artificial neural networks are the principal example, but they are not detailed replicas of biological brains.

Example

Represent a category through activity across many units.

Limits and common confusions

Connectionism is a broad approach, not a claim that a network thinks like a human.

Related: Machine learning

Editorial review: 2026-10-14 · Link to this term

Search and algorithms

Consistent heuristic

How it works

For an edge from n to n', consistency requires h(n) to be at most its edge cost plus h(n'). With suitable goal values, this supports safe graph-search pruning.

Example

Straight-line distance is consistent when edge costs are at least geometric distances.

Limits and common confusions

An admissible heuristic need not be consistent.

Related: Search algorithm

Editorial review: 2026-10-14 · Link to this term

Knowledge and reasoning

Constraint logic programming

How it works

Instead of resolving every relation by enumeration, a constraint solver maintains restrictions such as numeric bounds. Its guarantees depend on the domain and solver.

Example

Constrain start times using arithmetic inequalities inside a logical program.

Limits and common confusions

Constraint consistency does not always imply a complete solution exists.

Related: Knowledge representation and reasoning

Editorial review: 2026-10-14 · Link to this term

Knowledge and reasoning

Constraint programming

How it works

Solvers combine search with constraint propagation to find satisfying assignments or optimize an objective. The representation influences which reductions and search strategies are effective.

Example

Assign rooms and times without scheduling conflicts.

Limits and common confusions

Expressing the constraints does not make a difficult problem easy.

Related: Knowledge representation and reasoning

Editorial review: 2026-10-14 · Link to this term

Language models

Context window

In plain language

Only a limited amount of conversation, instructions and retrieved material can be provided at once.

Technical definition

A model or serving system imposes token limits; input and output accounting depends on that implementation. Available context length differs from reliable use of every included fact.

How it works

Lost in the Middle studied position-sensitive retrieval performance in long inputs. Its findings motivate testing relevant information at different positions instead of treating a large limit as proof of comprehension.

Example

Place the same answer-bearing passage at the beginning, middle and end of a long evaluation document.

Limits and common confusions

Adding more context can add distraction. The cited experiments do not establish identical behavior for every later model.

Related: Token · Attention · Retrieval-augmented generation

Editorial review: 2026-10-13 · Link to this term

Models and learning

Continual learning

How it works

A changing training stream can overwrite information needed for previous tasks. Methods such as replay or constraints on parameter changes address this stability-versus-adaptation problem.

Example

Learn new object classes without losing earlier recognition skills.

Limits and common confusions

Avoiding catastrophic forgetting is an objective, not a property guaranteed by the label.

Related: Machine learning

Editorial review: 2026-10-14 · Link to this term

Models and learning

Contrastive learning

How it works

A training objective encourages selected positive pairs to have compatible representations relative to negatives. How pairs are defined determines which similarities the representation learns.

Example

Treat two suitable crops of one image as a positive pair.

Limits and common confusions

Poor pairing or false negatives can suppress distinctions needed for a later task.

Related: Machine learning

Editorial review: 2026-10-14 · Link to this term

Language and prompting

Conversational AI

How it works

A conversational system may combine language understanding, dialogue state, response selection or generation, and speech components. It can be scripted, learned or hybrid and need not be a large language model.

Example

Maintain a booking conversation while collecting required details.

Limits and common confusions

A natural-sounding exchange does not imply reliable task completion or human-like understanding.

Related: Natural language processing

Editorial review: 2026-10-14 · Link to this term

Architectures

Convolutional neural network

CNN

In plain language

The same learned detector can inspect many image locations instead of learning an unrelated detector for every pixel.

Technical definition

Convolutional layers combine local receptive fields with shared parameters. Stride, padding and filter dimensions determine output geometry; pooling may reduce spatial resolution.

How it works

Weight sharing encodes a useful spatial assumption and reduces parameter requirements. A classification head or detection system can build on the resulting feature maps.

Example

A filter responds to an edge wherever that pattern appears within its usable receptive field.

Limits and common confusions

A CNN is not automatically invariant to all translations, rotations or scale changes. Boundary handling and sampling matter.

Related: Neural network · YOLO · Transfer learning

Editorial review: 2026-10-13 · Link to this term

Evaluation

Cross-validation

In plain language

Repeated held-out tests show how sensitive a result is to the available examples.

Technical definition

In k-fold cross-validation each fold is held out once while the procedure is trained on the others.

How it works

Split design must match deployment. Keep related observations together and respect temporal order where future prediction is intended. Fit preprocessing inside each training fold. Nested validation can separate hyperparameter selection from performance estimation.

Example

Hold out entire photographers when evaluating generalization to unseen photographers.

Limits and common confusions

Fold scores are not independent measurements, and careless random splitting can leak information.

Related: Data leakage · Overfitting

Editorial review: 2026-10-13 · Link to this term

Search and algorithms

Crossover in evolutionary algorithms

How it works

The recombination rule depends on the representation, such as strings, permutations or trees. It aims to reuse useful structures from different candidates.

Example

Combine portions of two encoded candidate solutions.

Limits and common confusions

Naive recombination can break constraints or useful dependencies between components.

Related: Search algorithm

Editorial review: 2026-10-14 · Link to this term

Products and history

DALL·E

DALL-E · DALL E

In plain language

A text description guides a generated image rather than selecting one fixed photograph.

Technical definition

The original DALL·E modeled text and image tokens autoregressively with a transformer. That architecture should not be assumed for every later model sharing the name.

How it works

The historical announcement illustrates compositional text-to-image generation. When documenting an image, record the exact model and prompt if available, along with edits and generation date.

Example

Describe an imaginary ceramic animal and generate candidate visual interpretations.

Limits and common confusions

Generated images can misrepresent requested details and are not documentary evidence. Product names do not establish provenance or current availability.

Related: Transformer · Diffusion model · CLIP

Editorial review: 2026-10-13 · Link to this term

Data and representation

Data augmentation

How it works

Transformations encode assumptions about valid variation. Cropping, color changes or paraphrasing can help learning when they preserve the target, but augmentation must match the task.

Example

Rotate a scanned symbol only when rotation leaves its label unchanged.

Limits and common confusions

A transformation can silently change the correct label or erase critical evidence.

Related: Dataset

Editorial review: 2026-10-14 · Link to this term

Evaluation

Data leakage

In plain language

Leakage gives a model an unfair preview of the answers or evaluation data.

Technical definition

It can enter through preprocessing, feature construction, duplicate examples or selection decisions.

How it works

Define the prediction time and data boundaries first. Separate evaluation data before learned transformations, and fit those transformations only on training portions. A pipeline helps apply the same procedure consistently.

Example

Feature selection on the complete dataset leaks held-out labels into subsequent evaluation.

Limits and common confusions

A pipeline cannot repair a dataset whose features already contain future outcomes.

Related: Cross-validation · Supervised learning

Editorial review: 2026-10-13 · Link to this term

Data and evaluation

Data mining

In plain language

It looks for useful structure, such as recurring combinations or groups of similar records.

Technical definition

In the knowledge-discovery workflow, mining is the pattern-extraction step within broader selection, preparation, evaluation and interpretation of data.

How it works

A discovered pattern becomes useful knowledge only after its validity, novelty and relevance are assessed. Data preparation and domain interpretation are integral to that assessment.

Example

Find recurring combinations of materials in an archive, then investigate whether cataloguing conventions explain them.

Limits and common confusions

A frequent association is not necessarily causal. Searching many possibilities can expose chance patterns that require independent validation.

Related: Clustering · Classification · Big data

Editorial review: 2026-10-13 · Link to this term

Models and learning

Data-driven AI

How it works

Data-driven methods use observations to fit models or discover structure. They still depend on human choices about representation, objectives, collection and deployment, and can be combined with explicit knowledge.

Example

Learn a predictive relation from historical measurements.

Limits and common confusions

Learning from data does not remove assumptions or establish causality.

Related: Machine learning

Editorial review: 2026-10-14 · Link to this term

Knowledge and reasoning

Datalog

How it works

Facts and rules describe relations, and queries ask what follows. Restrictions support useful database evaluation properties; extensions with negation or other features need specified semantics.

Example

Define an ancestor relation recursively from parent facts.

Limits and common confusions

Different Datalog extensions differ in expressiveness, termination and treatment of negation.

Related: Knowledge representation and reasoning

Editorial review: 2026-10-14 · Link to this term

Data and representation

Dataset

How it works

A dataset has a scope, collection process and representation, whether files, tables or another structure. Documentation should describe provenance, intended uses, limitations and relevant permissions.

Example

A collection of images with labels and acquisition metadata.

Limits and common confusions

A dataset is not necessarily a random or representative sample of its target population.

Editorial review: 2026-10-14 · Link to this term

Models and learning

DBSCAN

DBSCAN

How it works

Neighborhood radius and minimum-neighbor settings distinguish core points, border points and noise. The algorithm does not require specifying the number of clusters.

Example

Find dense geographic groups while leaving isolated observations unclustered.

Limits and common confusions

One global density scale can struggle when cluster densities vary substantially.

Related: Machine learning

Editorial review: 2026-10-14 · Link to this term

Models and learning

Decision boundary

How it works

A boundary is induced by a model's decision rule and threshold. It can be linear or nonlinear and may change when the threshold changes.

Example

Draw the curve where two class scores are equal.

Limits and common confusions

The boundary is learned from assumptions and data, not necessarily a natural separation in reality.

Related: Machine learning

Editorial review: 2026-10-14 · Link to this term

Probability and decisions

Decision theory

How it works

Expected-utility models combine probabilities with utilities to compare options. They separate beliefs about what may happen from judgments about what is valuable.

Example

Compare actions by expected benefit under uncertain outcomes.

Limits and common confusions

Utilities and probability estimates can be incomplete, disputed or incorrect.

Related: Probabilistic model

Editorial review: 2026-10-14 · Link to this term

Architectures

Decision tree

In plain language

A tree asks a sequence of questions before making a prediction.

Technical definition

Common algorithms greedily choose splits that reduce impurity or prediction loss.

How it works

Trees support classification and regression. Depth limits, minimum leaf sizes and pruning control complexity. A small tree can expose a readable decision path, while a very large tree becomes difficult to inspect.

Example

A toy image classifier first tests brightness, then color variation.

Limits and common confusions

Small data changes may produce different trees; an interpretable path is not evidence of causation.

Related: Classification · Regression · Random forest

Editorial review: 2026-10-13 · Link to this term

Software and openness

Declarative programming

How it works

Logic queries and constraints describe conditions for solutions, while an underlying engine chooses how to evaluate them. The boundary with imperative programming is not absolute.

Example

State which rows satisfy a query rather than manually iterating over them.

Limits and common confusions

Declarative syntax does not remove computational cost or guarantee efficient execution.

Related: Open-source software

Editorial review: 2026-10-14 · Link to this term

Knowledge and reasoning

Deductive classifier

How it works

Unlike a statistical classifier, it derives consequences under a formal knowledge model. Description-logic systems commonly use such classification services.

Example

Infer that an explicitly described entity belongs to a defined class.

Limits and common confusions

A missing assertion is not automatically false in open-world reasoning.

Related: Knowledge representation and reasoning

Editorial review: 2026-10-14 · Link to this term

History and concepts

Deep Blue

In plain language

It is a landmark example of strong performance in a narrowly defined intellectual task.

Technical definition

The system combined chess search and evaluation with specialized computing hardware. Its approach should not be confused with the later neural-network methods often called deep learning.

How it works

The match demonstrated the strength of an engineered chess system under the competition conditions. It did not establish general-purpose intelligence across unrelated tasks.

Example

Explore possible move sequences and evaluate positions before selecting a move.

Limits and common confusions

The word 'Deep' in the name is not evidence that the system used modern deep learning.

Related: Narrow AI · Heuristic · Deep learning

Editorial review: 2026-10-13 · Link to this term

Architectures

Deep learning

In plain language

Instead of supplying every useful feature manually, a deep model can learn intermediate representations along with the task.

Technical definition

Depth composes transformations so later representations depend on earlier ones. Training adjusts parameters to fit an objective, with supervised, self-supervised or other learning signals.

How it works

Depth is an architectural property, not a synonym for intelligence. Representations may capture reusable structure, but what they capture depends on the data, objective and optimization.

Example

An image model learns intermediate visual features before producing category scores.

Limits and common confusions

More layers do not automatically improve generalization. Data quality, computation and evaluation remain essential.

Related: Neural network · Self-supervised learning · Overfitting

Editorial review: 2026-10-13 · Link to this term

Vision and applications

Deepfake

Deep fake

In plain language

A familiar example changes a person's depicted face or performance in a video.

Technical definition

FaceForensics++ studies facial manipulation detection across several generation methods and compression conditions. It is a benchmark for particular manipulations, not a detector for every synthetic medium.

How it works

Separate the act of synthesis from the claim made about it: clearly labeled creative editing and deceptive impersonation have different contexts. Preserve provenance when assessing authenticity.

Example

Compare an original interview frame with a manipulated facial rendering in a labeled research dataset.

Limits and common confusions

A detector can fail on new generators or compression settings. Absence of a detected artifact does not prove authenticity.

Related: Computer vision · Generative adversarial network · Distribution shift

Editorial review: 2026-10-13 · Link to this term

Knowledge and reasoning

Description logic

DL

How it works

Description logics underpin many ontology languages and support tasks such as classification and consistency checking. Different members balance expressiveness and decidability.

Example

Describe a class using restrictions on its relationships to other classes.

Limits and common confusions

A claim about one description logic may not hold for a more expressive variant.

Related: Knowledge representation and reasoning

Editorial review: 2026-10-14 · Link to this term

Reliability and security

Differential privacy

DP

In plain language

It limits what a released result can reveal about whether a particular contribution was included.

Technical definition

A randomized mechanism compares neighboring datasets using privacy parameters such as epsilon and, in approximate forms, delta. The adjacency definition specifies the protected unit.

How it works

Noise must be calibrated to the computation and privacy accounting. Repeated releases consume a composed privacy budget; simply adding arbitrary randomness is insufficient.

Example

Release a noisy aggregate count with a documented privacy mechanism and budget.

Limits and common confusions

Protection depends on parameter choices, implementation and the protected unit. It does not prevent learning true population-level patterns or remove every privacy risk.

Related: Machine learning · Big data

Editorial review: 2026-10-13 · Link to this term

Generative models

Diffusion model

Denoising diffusion probabilistic model · DDPM

In plain language

The model learns how to turn progressively corrupted data back toward plausible samples.

Technical definition

A learned reverse process approximates denoising transitions associated with a specified forward corruption process.

How it works

In a common formulation, training adds Gaussian noise to examples at different time steps. A neural network learns a denoising quantity, often the added noise. Sampling starts from noise and repeatedly applies learned reverse steps. Text conditioning can guide generation, but it is an additional mechanism rather than the definition of diffusion.

Example

A system synthesizes an image by repeatedly refining a noisy array under a caption's guidance.

Limits and common confusions

Sampling can require many evaluations. Results depend on the training distribution, conditioning and sampler; a visually convincing image is not evidence that its depicted event occurred.

Related: Generative adversarial network · Variational autoencoder

Editorial review: 2026-10-13 · Link to this term

Search and algorithms

Dijkstra's algorithm

How it works

It repeatedly finalizes the smallest tentative distance and relaxes outgoing edges. A priority queue commonly manages the frontier.

Example

Compute cheapest routes from one station in a weighted network.

Limits and common confusions

Negative edge weights invalidate the usual correctness argument.

Related: Search algorithm

Editorial review: 2026-10-14 · Link to this term

Data and evaluation

Dimensionality reduction

In plain language

It compresses a many-feature representation into a smaller one for analysis, visualization or modeling.

Technical definition

Methods optimize different criteria, such as reconstruction, variance retention or local neighborhood relationships. Linear projection and nonlinear embeddings need not preserve the same properties.

How it works

Choose the method according to what must be preserved. A two-dimensional visualization is a representation with distortions, not a complete picture of the original geometry.

Example

Project correlated measurements onto a few principal components before inspecting broad variation.

Limits and common confusions

Apparent clusters can depend on parameters and projection artifacts. Reduced dimensions can lose information relevant to a downstream task.

Related: Principal component analysis · Embedding · Clustering

Editorial review: 2026-10-13 · Link to this term

Training and optimization

Direct preference optimization

DPO

In plain language

DPO makes preferred answers more likely relative to alternatives and a reference model.

Technical definition

Its loss uses policy-to-reference log-probability ratios derived from a regularized preference-learning formulation.

How it works

The original method avoids separately training an explicit reward model and running an online reinforcement-learning loop during fine-tuning. It still uses a defined preference dataset and objective. Its relationship to reward optimization follows mathematical assumptions rather than meaning that the model has acquired a universal notion of good behavior.

Example

Train on response pairs where editors prefer a well-supported explanation to an unsupported confident claim.

Limits and common confusions

Biased or inconsistent comparisons can train unwanted preferences. Better preference scores do not automatically establish factual correctness or robust behavior outside the training distribution.

Related: Fine-tuning · Reinforcement learning from human feedback

Editorial review: 2026-10-13 · Link to this term

Agents and systems

Distributed artificial intelligence

DAI

How it works

Agents may cooperate, coordinate or compete with limited information. Communication and task allocation are part of the problem, not incidental implementation details.

Example

Coordinate several agents that each observe part of an environment.

Limits and common confusions

Communication delays, conflicting objectives and partial knowledge can undermine coordination.

Related: AI agent

Editorial review: 2026-10-14 · Link to this term

Data and evaluation

Distribution shift

In plain language

A model may face different examples from the ones it learned to handle.

Technical definition

Shifts can affect inputs, label proportions or relationships between inputs and outcomes. WILDS evaluates naturally occurring shifts across settings such as locations and time.

How it works

Evaluate on relevant new environments and monitor performance after deployment. A random split within one dataset may miss the variation that matters operationally.

Example

A wildlife classifier trained on one camera location encounters different backgrounds and illumination elsewhere.

Limits and common confusions

A detected input change does not by itself quantify prediction harm, and stable aggregate statistics can hide important subgroup changes.

Related: Cross-validation · Transfer learning · Benchmark

Editorial review: 2026-10-13 · Link to this term

Models and learning

Double descent

How it works

In some settings, error peaks near an interpolation threshold before improving in a more overparameterized regime. The phenomenon depends on data, model and training choices.

Example

Compare models on both sides of the capacity needed to fit all training labels.

Limits and common confusions

Double descent is not a universal guarantee that ever-larger models improve.

Related: Machine learning

Editorial review: 2026-10-14 · Link to this term

Models and learning

Dropout

How it works

Sampling different masks discourages reliance on particular co-adapted activations. Implementations usually rescale activations and disable ordinary dropout during evaluation.

Example

Train a dense network with randomly masked hidden activations.

Limits and common confusions

Dropout is not always beneficial and must be used consistently between training and inference.

Related: Machine learning

Editorial review: 2026-10-14 · Link to this term

Models and learning

Eager learning

How it works

Eager methods invest computation during training so later predictions use the fitted model. This contrasts with lazy methods that defer more work until a query arrives.

Example

Fit a decision tree before classifying new records.

Limits and common confusions

Eager does not necessarily mean faster overall or less memory-intensive.

Related: Machine learning

Editorial review: 2026-10-14 · Link to this term

Models and learning

Early stopping

How it works

A common rule monitors validation performance and stops after improvement stalls. The selected checkpoint and stopping rule are part of model selection.

Example

Keep the checkpoint with the best validation loss.

Limits and common confusions

Repeated decisions based on a test set compromise its role as independent evaluation.

Related: Machine learning

Editorial review: 2026-10-14 · Link to this term

Models and learning

Echo state network

ESN

How it works

Input-driven reservoir dynamics create temporal features. The echo-state property concerns loss of dependence on initial conditions under suitable conditions, not merely random initialization.

Example

Train a linear readout from a recurrent reservoir for sequence prediction.

Limits and common confusions

Reservoir configuration and stability affect performance; not every random recurrent network has the required property.

Related: Machine learning

Editorial review: 2026-10-14 · Link to this term

Representations

Embedding

In plain language

An embedding represents something such as a word or image as coordinates a model can compare.

Technical definition

Similarity is defined in a representation space whose geometry depends on the learning objective.

How it works

Word embeddings can be learned from contextual co-occurrence. Contextual models can assign different vectors to the same word in different sentences. Distance comparisons are meaningful only relative to a specified representation and metric.

Example

Retrieve glossary passages near a question embedding, then verify the retrieved evidence.

Limits and common confusions

Nearby vectors do not prove factual equivalence. Representations can encode unwanted corpus associations.

Related: Retrieval-augmented generation · Clustering

Editorial review: 2026-10-13 · Link to this term

Agents and systems

Embodied agent

How it works

The body's capabilities and constraints influence what can be observed and done. Embodiment need not imply human-like form.

Example

A wheeled robot uses sensors and motors to navigate.

Limits and common confusions

Simulation performance may not transfer when real sensing and dynamics differ.

Related: AI agent

Editorial review: 2026-10-14 · Link to this term

Models and learning

Emergent capabilities

How it works

Some evaluations show sharp changes across model sizes. Research also shows that coarse scoring can turn gradual underlying improvements into apparently abrupt transitions; the metric and evidence matter.

Example

Exact-match accuracy rises suddenly while token-level error improves gradually.

Limits and common confusions

Emergence is a description needing a specified measurement, not proof of consciousness or a universal scaling threshold.

Related: Machine learning

Editorial review: 2026-10-14 · Link to this term

Learning methods

Ensemble learning

In plain language

Models can vote, average predictions or learn how to combine one another.

Technical definition

Bagging, boosting and stacking use different training and aggregation strategies.

How it works

Ensembles are useful when component errors leave room for improvement. Assess the complete combined procedure on held-out data rather than selecting a flattering component score.

Example

Average several independently trained regression models to stabilize estimates.

Limits and common confusions

Combining models with the same systematic error may preserve that error and adds computational cost.

Related: Random forest · Cross-validation

Editorial review: 2026-10-13 · Link to this term

Models and learning

Epoch

How it works

Training commonly uses multiple epochs, divided into minibatches. Streaming, resampling and distributed training require explicit conventions for what counts as a traversal.

Example

Process a finite dataset once before beginning another pass.

Limits and common confusions

More epochs can increase overfitting; one epoch is not one parameter update.

Related: Machine learning

Editorial review: 2026-10-14 · Link to this term

Models and learning

Error-driven learning

How it works

Different algorithms convert error into parameter updates in different ways. Gradient-based training and classical perceptron updates are examples, but not every learning objective uses explicit labeled targets.

Example

Update a predictor after observing its residual error.

Limits and common confusions

Reducing the chosen error can fail to improve the behavior actually desired.

Related: Machine learning

Editorial review: 2026-10-14 · Link to this term

Search and reasoning

Evolutionary algorithm

In plain language

Keep useful candidates, create variations and test the next generation.

Technical definition

A fitness evaluation guides selection; mutation and sometimes recombination explore alternatives. The representation and evaluation function determine the search space and selection pressure.

How it works

Evolutionary methods can search where useful derivatives are unavailable. In policy search, candidates may be controllers evaluated by interaction with a simulated environment.

Example

Evaluate candidate game-playing policies over several trials and vary stronger performers.

Limits and common confusions

Many evaluations may be expensive or noisy. Biological terminology does not imply biological realism or a guarantee of finding the best solution.

Related: Genetic algorithm · Optimization · Reinforcement learning

Editorial review: 2026-10-13 · Link to this term

Search and reasoning

Expert system

In plain language

It tries to support a particular expert task rather than demonstrate general intelligence.

Technical definition

Classic expert systems combine a domain knowledge base with an inference mechanism. Rule-based examples can record the rules supporting a recommendation or conclusion.

How it works

MYCIN is a historical research example in medical decision support. Its importance illustrates knowledge representation and reasoning; it should not be treated as a current clinical tool.

Example

An equipment troubleshooting system asks questions and narrows possible causes using explicit domain rules.

Limits and common confusions

Acquiring and maintaining expert knowledge is difficult. Performance within a narrow test setting does not establish reliability outside that scope.

Related: Knowledge base · Rule-based system · Narrow AI

Editorial review: 2026-10-13 · Link to this term

Reliability and security

Explainable AI

XAI · Explainable artificial intelligence

In plain language

An explanation should help its intended audience understand a result and its limitations.

Technical definition

NIST distinguishes providing explanations, making them meaningful, ensuring explanation accuracy and recognizing knowledge limits. These requirements concern more than producing fluent explanatory text.

How it works

Choose the explanation for the question and audience. A developer debugging features and a person affected by a decision may need different information.

Example

Show which input features influenced a prediction while stating the method's scope and uncertainty.

Limits and common confusions

A convincing explanation can be unfaithful to the actual computation. Explainability does not by itself establish fairness, causality or correctness.

Related: Model transparency: black, white and grey boxes · Classification · Hallucination in generative AI

Editorial review: 2026-10-13 · Link to this term

Data and representation

Feature

How it works

Features may be measured values, engineered attributes or learned numerical representations. Their construction determines which information a model can use.

Example

Represent a document by counts of selected words.

Limits and common confusions

A predictive feature can be a misleading proxy or leak target information.

Related: Dataset

Editorial review: 2026-10-14 · Link to this term

Data and evaluation

Feature extraction

In plain language

It turns material such as text or images into measurable inputs.

Technical definition

An extractor may compute counts, local descriptors or learned representations. Feature selection instead chooses among features already available.

How it works

Representations determine which distinctions a downstream model can easily use. A text count vector records word occurrence but may discard order; other representations retain different information.

Example

Convert descriptions into word-count vectors before grouping related records.

Limits and common confusions

Fit data-dependent preprocessing only on permitted training data. Extraction can discard important context or leak information if designed carelessly.

Related: Embedding · Data leakage · Classification

Editorial review: 2026-10-13 · Link to this term

Data and representation

Feature selection

How it works

Methods may use relevance scores, model fitting or search over subsets. Selection must occur inside the training-validation process to avoid leakage.

Example

Select informative measurements within each cross-validation training fold.

Limits and common confusions

A selected feature can be predictive without being causal.

Related: Dataset

Editorial review: 2026-10-14 · Link to this term

Models and learning

Federated learning

How it works

Participants compute local updates and a coordinating procedure combines them into a shared model. Communication cost, uneven participant data and unreliable devices affect the result.

Example

Several hospitals contribute model updates without pooling patient files.

Limits and common confusions

Updates can still leak information; federated learning alone is not a privacy guarantee.

Related: Machine learning

Editorial review: 2026-10-14 · Link to this term

Agents and systems

Feedback loop in AI systems

How it works

Recommendations, user responses and later training data can become coupled. Feedback may be intentional or hidden and can reinforce initial biases or change the environment being modeled.

Example

Recommended items receive more clicks and then appear more popular in training data.

Limits and common confusions

Feedback does not necessarily improve performance; an apparent signal can be created by the system itself.

Related: AI agent

Editorial review: 2026-10-14 · Link to this term

Learning methods

Few-shot learning

In plain language

A system draws on previous learning to make a few new examples useful.

Technical definition

Meta-learning can train across tasks for rapid adaptation; MAML learns an initialization suited to a few gradient steps. In-context demonstrations instead guide predictions without necessarily changing weights.

How it works

These mechanisms share a low-example setting but should not be conflated. Report the number of classes, examples per class and adaptation procedure.

Example

Provide a handful of labeled images for each new material category and evaluate on separate images.

Limits and common confusions

Prior training can contain extensive related data. Few-shot performance depends on task similarity, example selection and evaluation design.

Related: Zero-shot learning · Fine-tuning · Transfer learning

Editorial review: 2026-10-13 · Link to this term

Language and prompting

Few-shot prompting

How it works

Examples specify a pattern or task within the current context. Ordinary in-context demonstrations do not update the model's weights, unlike many few-shot learning algorithms.

Example

Show two input-label pairs before asking for a third label.

Limits and common confusions

Example ordering and choice can alter results; this is not evidence of durable learning.

Related: Natural language processing

Editorial review: 2026-10-14 · Link to this term

Learning methods

Fine-tuning

In plain language

Fine-tuning changes learned parameters using additional examples.

Technical definition

Optimization starts from pretrained parameters rather than a fresh initialization.

How it works

BERT demonstrated transferring pretrained representations to downstream language tasks using task-specific training. What changes depends on the method: all parameters, selected layers or additional parameters may be trained.

Example

Adapt a text encoder to classify glossary questions using labeled examples.

Limits and common confusions

Fine-tuning can overfit or degrade other capabilities; adding examples to a prompt does not itself update weights.

Related: Supervised learning · Transformer

Editorial review: 2026-10-13 · Link to this term

Knowledge and reasoning

First-order logic

How it works

It expresses statements about objects and relations using a domain and interpretation. Quantification is over individuals rather than over arbitrary predicates or functions.

Example

State that every artwork has at least one creator in a model.

Limits and common confusions

General first-order validity is undecidable; richer expression increases reasoning difficulty.

Related: Knowledge representation and reasoning

Editorial review: 2026-10-14 · Link to this term

Knowledge and reasoning

Fluent in planning

How it works

Fluents let action models express changing conditions while keeping other domain knowledge fixed. Their formal representation depends on the planning language.

Example

A door's open-or-closed status changes after an action.

Limits and common confusions

A fluent is not necessarily a Boolean variable in every formalism.

Related: Knowledge representation and reasoning

Editorial review: 2026-10-14 · Link to this term

Computing foundations

Formal language

How it works

Grammars and automata describe different families of languages. Formal languages support parsing, logic and programming-language analysis.

Example

Define the strings accepted by a parser for arithmetic expressions.

Limits and common confusions

A formal language need not resemble a natural spoken language.

Related: Algorithm

Editorial review: 2026-10-14 · Link to this term

Knowledge and reasoning

Forward chaining

How it works

A data-driven inference procedure repeatedly adds conclusions whose premises are established. Its behavior depends on the rule language and termination strategy.

Example

From facts about a device, derive all applicable warnings.

Limits and common confusions

Unrestricted rule systems may not terminate or may produce unintended consequences from bad premises.

Related: Knowledge representation and reasoning

Editorial review: 2026-10-14 · Link to this term

Models and learning

Foundation model

How it works

Pretraining supplies reusable capabilities; adaptation can use prompting, fine-tuning or other methods. A foundation model may handle language, images or several modalities, and downstream systems can inherit its limitations.

Example

Adapt a pretrained vision model to classify plant diseases.

Limits and common confusions

Broad usefulness does not imply universal competence or safe behavior without application-specific evaluation.

Related: Machine learning

Editorial review: 2026-10-14 · Link to this term

Knowledge and reasoning

Frame in knowledge representation

How it works

Slots hold values or references to other frames. Defaults can represent typical expectations that are revised when more specific information arrives.

Example

Represent a restaurant visit with participants, roles and expected events.

Limits and common confusions

Typical expectations are defeasible and can encode cultural assumptions.

Related: Knowledge representation and reasoning

Editorial review: 2026-10-14 · Link to this term

Knowledge and reasoning

Frame language

How it works

Implementations differ in inheritance, defaults and attached procedures. A frame language is not the same thing as a formal logical calculus.

Example

Define object classes with expected attributes and inherited slot values.

Limits and common confusions

Inheritance conflicts and default exceptions require explicit semantics.

Related: Knowledge representation and reasoning

Editorial review: 2026-10-14 · Link to this term

Knowledge and reasoning

Frame problem

How it works

A useful action theory must distinguish effects from unaffected facts without enumerating every unchanged fact for every action. This is narrower than all difficulties of commonsense reasoning.

Example

Moving a cup changes its location but ordinarily not its color.

Limits and common confusions

A solution depends on the action formalism and does not solve every real-world exception.

Related: Knowledge representation and reasoning

Editorial review: 2026-10-14 · Link to this term

Agents and systems

Fuzzy control system

How it works

Linguistic rules encode graded responses and are combined through a chosen inference procedure. Such controllers can incorporate operational knowledge without requiring a complete analytic plant model.

Example

Adjust a control signal using fuzzy categories of error and error change.

Limits and common confusions

Controller stability and performance still require analysis or testing.

Related: AI agent

Editorial review: 2026-10-14 · Link to this term

Search and reasoning

Fuzzy logic

In plain language

A value can partly belong to a category such as 'warm' instead of crossing one sharp boundary.

Technical definition

Fuzzy sets assign membership grades, commonly in the interval from zero to one. Operations on these grades support forms of approximate reasoning; the cited paper establishes the foundational set concept.

How it works

The membership function represents the chosen meaning of a vague category. It must be specified or learned for the intended application.

Example

A temperature may have partial membership in both 'cool' and 'warm' categories.

Limits and common confusions

A membership grade is not automatically a probability. Vagueness about a category differs from uncertainty about whether an event occurred.

Related: Rule-based system · Probability calibration

Editorial review: 2026-10-13 · Link to this term

Knowledge and reasoning

Fuzzy rule

How it works

A fuzzy inference system combines rule activations according to specified operators and may convert the result to a crisp output. Different inference schemes use different semantics.

Example

If temperature is high, increase cooling strongly.

Limits and common confusions

The rule's meaning depends on membership functions and combination operators.

Related: Knowledge representation and reasoning

Editorial review: 2026-10-14 · Link to this term

Knowledge and reasoning

Fuzzy set

How it works

A membership function maps elements to values, commonly in the interval from zero to one. Membership grades represent vagueness under the chosen interpretation.

Example

Represent how strongly a temperature belongs to the category warm.

Limits and common confusions

A membership degree is not automatically a probability that an event occurs.

Related: Knowledge representation and reasoning

Editorial review: 2026-10-14 · Link to this term

Probability and decisions

Game theory

How it works

A game specifies players, actions, information and payoffs. Different equilibrium and solution concepts answer different questions about strategic behavior.

Example

Analyze choices in an auction or coordination game.

Limits and common confusions

A mathematical equilibrium is not proof that real participants will behave that way.

Related: Probabilistic model

Editorial review: 2026-10-14 · Link to this term

Evaluation

Generalization error

How it works

Held-out measurements estimate this quantity under assumptions about sampling and independence. It is distinct from the observed training loss and from the generalization gap.

Example

Estimate classification error on an independent test sample.

Limits and common confusions

A small or unrepresentative test set gives an uncertain or biased estimate.

Related: Model evaluation

Editorial review: 2026-10-14 · Link to this term

Evaluation

Generalization in machine learning

How it works

Evaluation specifies which new cases matter and how they differ from training data. Generalization within one distribution differs from transfer to a new environment.

Example

Evaluate a classifier on independently collected examples.

Limits and common confusions

Good training performance is not evidence of generalization by itself.

Related: Model evaluation

Editorial review: 2026-10-14 · Link to this term

Generative models

Generative adversarial network

GAN

In plain language

One network generates examples while another learns to distinguish them from training examples.

Technical definition

Generator and discriminator parameters are optimized through coupled adversarial objectives.

How it works

The generator maps sampled inputs to synthetic examples. The discriminator learns to distinguish real examples from generated ones, supplying a training signal to the generator. The original formulation is a minimax game; practical variants modify losses, architectures and regularization to improve training.

Example

A generator learns to produce texture patches while a discriminator critiques whether they resemble patches from a reference collection.

Limits and common confusions

Training can be unstable, and mode collapse can reduce variety. Realistic output does not prove that the learned distribution covers all relevant examples or avoids memorization.

Related: Diffusion model · Variational autoencoder

Editorial review: 2026-10-13 · Link to this term

Vision and generation

Generative AI

GenAI

How it works

Generative models learn or represent patterns in data and use them to construct outputs, often conditioned on an instruction or other input. Sampling methods and training objectives vary across model families.

Example

Generate an illustrative image from a written description.

Limits and common confusions

Novel-looking output may reproduce training material, contain factual errors or lack originality in a legal sense.

Related: Computer vision

Editorial review: 2026-10-14 · Link to this term

Language and prompting

Generative pretrained Transformer

GPT

How it works

The original GPT work trained a Transformer language model on unlabeled text and then adapted it to supervised tasks. Later GPT systems differ substantially; the name should not be treated as a description of every language model.

Example

Pretrain next-token prediction before adapting to a classification task.

Limits and common confusions

GPT names a model lineage and approach, not every Transformer or every conversational product.

Related: Natural language processing

Editorial review: 2026-10-14 · Link to this term

Human-AI interaction

Generative user interface

GenUI

How it works

A model may construct a form, visualization or control arrangement at runtime instead of only filling a fixed screen with text. This is distinct from using AI to design a static interface during development.

Example

Generate an interactive comparison table for a user's current question.

Limits and common confusions

Generated controls still need accessibility, consistency and safe execution boundaries.

Related: Human-computer interaction

Editorial review: 2026-10-14 · Link to this term

Vision and generation

Generator in generative modeling

How it works

In a GAN, the generator is trained against a discriminator's feedback. Other generative systems can produce samples without a GAN discriminator, so the surrounding architecture matters.

Example

Map a sampled latent vector to a synthetic image.

Limits and common confusions

The generator's output is not necessarily novel, accurate or representative of the full data distribution.

Related: Computer vision

Editorial review: 2026-10-14 · Link to this term

Search and reasoning

Genetic algorithm

In plain language

Candidate solutions are combined or mutated to create a new population.

Technical definition

A chromosome encodes a solution; crossover exchanges components and mutation modifies them. Fitness-based selection biases which candidates contribute to later generations.

How it works

Encoding matters: recombining useful pieces helps only if those pieces remain meaningful in the resulting candidate. Constraint handling may repair or reject infeasible offspring.

Example

Represent a proposed timetable as assignments and search for versions with fewer conflicts.

Limits and common confusions

Poor encodings and premature loss of diversity can limit results. Population size and variation rates require task-specific evaluation.

Related: Evolutionary algorithm · Optimization

Editorial review: 2026-10-13 · Link to this term

Search and algorithms

Genetic operator

How it works

Mutation, crossover and selection shape how a population changes. Operators must be compatible with the representation and constraints of the problem.

Example

Modify a candidate schedule while preserving required assignments.

Limits and common confusions

An operator that destroys feasibility can waste most search effort.

Related: Search algorithm

Editorial review: 2026-10-14 · Link to this term

Search and algorithms

Glowworm swarm optimization

How it works

Agents update a signal analogous to brightness and move toward better-signaled neighbors within adaptive local neighborhoods. This can support splitting into groups around different optima.

Example

Search for several spatial signal sources.

Limits and common confusions

Results depend on neighborhood dynamics and objective properties; all optima are not guaranteed in every setting.

Related: Search algorithm

Editorial review: 2026-10-14 · Link to this term

Models and learning

Gradient boosting

How it works

Each step approximates a direction of improvement in function space, commonly using regression trees fitted to negative loss gradients.

Example

Improve a predictor by adding small trees with a learning-rate factor.

Limits and common confusions

Tree depth, learning rate and iteration count need validation; the name does not imply neural-network training.

Related: Machine learning

Editorial review: 2026-10-14 · Link to this term

Computing foundations

Graph data type

How it works

Operations can include adding edges, listing neighbors and accessing labels. Adjacency lists and matrices are alternative implementations with different trade-offs.

Example

Store a sparse relation network using adjacency lists.

Limits and common confusions

The abstract graph and a particular memory representation are different objects.

Related: Algorithm

Editorial review: 2026-10-14 · Link to this term

Data and representation

Graph database

GDB

How it works

Graph models emphasize relationships and traversal. RDF stores and property-graph databases use different data models and query conventions.

Example

Store people, organizations and the relationships connecting them.

Limits and common confusions

Using a graph does not itself provide an ontology or correct entity resolution.

Related: Dataset

Editorial review: 2026-10-14 · Link to this term

Computing foundations

Graph in mathematics

How it works

Graphs can be directed or undirected and may carry labels or weights. The selected formal definition determines whether loops or repeated edges are allowed.

Example

Represent cities as vertices and roads as edges.

Limits and common confusions

A graph representation needs explicit meanings for its vertices and edges.

Related: Algorithm

Editorial review: 2026-10-14 · Link to this term

Computing foundations

Graph theory

How it works

It studies properties such as connectivity, paths, cycles and coloring. AI uses these ideas in search, graphical models and knowledge representations.

Example

Ask whether a network remains connected after an edge is removed.

Limits and common confusions

Graph-theoretic properties do not supply the domain meaning of a modeled relationship.

Related: Algorithm

Editorial review: 2026-10-14 · Link to this term

Search and algorithms

Graph traversal

How it works

Breadth-first and depth-first traversal use different exploration orders. Tracking visited vertices prevents repeated processing in graphs with cycles.

Example

Visit all entities reachable from a starting record.

Limits and common confusions

A traversal order alone does not optimize a weighted route.

Related: Search algorithm

Editorial review: 2026-10-14 · Link to this term

Computing infrastructure

Graphics processing unit

GPU

How it works

Many AI operations can be organized into parallel arithmetic over arrays. Effective acceleration also depends on memory transfers, available memory, kernel implementation and workload structure.

Example

Run batched matrix multiplications during neural-network training.

Limits and common confusions

A GPU does not accelerate every algorithm and may be limited by data movement rather than arithmetic.

Related: Compute in AI

Editorial review: 2026-10-14 · Link to this term

Language and prompting

Grounding in AI answers

How it works

In retrieval-based answering, grounding involves using retrieved records to support claims. Other fields use grounding for linking symbols to perception or action; the intended sense should be identified.

Example

Answer a policy question using the current policy document.

Limits and common confusions

A citation or retrieved passage does not prove that every generated claim is supported.

Related: Natural language processing

Editorial review: 2026-10-14 · Link to this term

Reliability

Hallucination in generative AI

In plain language

A fluent answer can contain details that do not have a reliable basis.

Technical definition

Evaluation should distinguish factual correctness from faithfulness to a supplied source and define the evidence standard.

How it works

Language generation can reproduce misconceptions as well as invent plausible details. TruthfulQA investigates one part of this problem: answering questions that invite common false beliefs. Other tasks need different tests, such as checking whether a summary's claims follow from its source. Retrieval can supply evidence, but the generated claims must still be checked.

Example

A glossary assistant supplies a plausible paper title that cannot be found in the cited source.

Limits and common confusions

The term is broad and does not imply human perception or deliberate deception. A benchmark score cannot establish that every future answer is reliable.

Related: Large language model · Retrieval-augmented generation

Editorial review: 2026-10-13 · Link to this term

Search and reasoning

Heuristic

In plain language

A heuristic helps decide which promising option to examine next.

Technical definition

In informed search, a heuristic function estimates the remaining cost to a goal. Particular guarantees depend on its properties and on the search algorithm.

How it works

An admissible cost heuristic never overestimates the optimal remaining cost. This precise condition is stronger than the everyday claim that a shortcut usually works well.

Example

Straight-line distance can guide route search when route cost is geographic distance.

Limits and common confusions

A helpful-looking estimate may mislead search. Optimality guarantees should not be assumed for arbitrary heuristics or cost definitions.

Related: Algorithm · Optimization

Editorial review: 2026-10-13 · Link to this term

Models and learning

Hidden layer

How it works

Hidden units transform intermediate representations during computation. Hidden describes their architectural position, not a special secrecy or consciousness property.

Example

Use intermediate layers to construct features for classification.

Limits and common confusions

A hidden unit's activation need not correspond to one understandable concept.

Related: Machine learning

Editorial review: 2026-10-14 · Link to this term

Human-AI interaction

Human in the loop

How it works

People may label training examples, review outputs or authorize actions. The timing, information available and actual power to intervene determine what oversight achieves.

Example

Require a qualified reviewer to approve a proposed publication.

Limits and common confusions

A nominal reviewer who lacks time, context or authority may provide little effective oversight.

Related: Human-computer interaction

Editorial review: 2026-10-14 · Link to this term

Human-AI interaction

Human-centered AI

How it works

It considers reliable automation together with meaningful human control and accountability. Evaluation includes the people and workflows affected, rather than model scores alone.

Example

Design an assistant so users can inspect evidence and reverse its changes.

Limits and common confusions

Human-centered is a design commitment that needs evidence, not a property established by branding.

Related: Human-computer interaction

Editorial review: 2026-10-14 · Link to this term

Human-AI interaction

Human-computer interaction

HCI

How it works

HCI examines usability, accessibility, feedback and the broader context of use. AI introduces uncertainty and adaptation, making clear expectations and recovery from errors especially important.

Example

Evaluate whether users can understand and correct an automated recommendation.

Limits and common confusions

A visually attractive interface can still be confusing, inaccessible or misleading.

Editorial review: 2026-10-14 · Link to this term

Knowledge and reasoning

Hybrid AI

How it works

In this glossary the label particularly covers combinations of learned models and explicit knowledge or reasoning. Neuro-symbolic systems are one family; the actual components should always be named.

Example

Combine a statistical classifier with a rule-based eligibility check.

Limits and common confusions

The term is broad and sometimes promotional; it does not specify an architecture or performance guarantee.

Related: Knowledge representation and reasoning

Editorial review: 2026-10-14 · Link to this term

Search and algorithms

Hyper-heuristic

How it works

The search operates at the level of heuristic choice or composition rather than only candidate solutions. Learning may adapt which low-level method is used.

Example

Choose among scheduling heuristics based on observed performance.

Limits and common confusions

Transfer to a new problem distribution requires evaluation rather than assumption.

Related: Search algorithm

Editorial review: 2026-10-14 · Link to this term

Models and learning

Hyperparameter

How it works

Examples include learning rate, tree depth and regularization strength. Such choices are often selected using validation data, rather than estimated by the model's ordinary parameter-fitting step.

Example

Choose a tree-depth limit using cross-validation.

Limits and common confusions

Tuning against test results leaks evaluation information and can overstate performance.

Related: Machine learning

Editorial review: 2026-10-14 · Link to this term

Models and learning

Hyperparameter optimization

How it works

Grid search, random search and model-based strategies explore configurations with different cost-quality trade-offs. The validation procedure is part of the optimization problem.

Example

Search learning rates using cross-validation.

Limits and common confusions

Repeated search can overfit validation data; a final independent evaluation remains useful.

Related: Machine learning

Editorial review: 2026-10-14 · Link to this term

Computing foundations

Hyperplane

How it works

In d-dimensional Euclidean space it can be written as a nontrivial linear equation. Linear classifiers use hyperplanes as decision boundaries.

Example

A line is a hyperplane in a two-dimensional plane.

Limits and common confusions

Not every decision boundary is a hyperplane, and a hyperplane need not be a classifier.

Related: Algorithm

Editorial review: 2026-10-14 · Link to this term

Vision and applications

Image classification

Image recognition

In plain language

It answers which categories are present according to a defined label set.

Technical definition

A classifier maps image features to class scores or probabilities. Single-label and multilabel settings differ; classification alone does not require bounding-box localization.

How it works

ImageNet's challenge helped standardize large-scale visual evaluation. Its labels and evaluation protocol define a particular recognition problem rather than all possible visual understanding.

Example

Label an archive image as containing a watercolor painting rather than a photograph.

Limits and common confusions

Scores depend on the training taxonomy. A closed-set classifier may assign a known label to an unfamiliar subject; high confidence is not proof of correctness.

Related: Classification · Object detection · Probability calibration

Editorial review: 2026-10-13 · Link to this term

Models and learning

Incremental learning

How it works

Incremental procedures update a maintained representation or parameters. They overlap with online and continual learning but do not necessarily address retention across changing tasks.

Example

Call a partial-fit operation on successive chunks of data.

Limits and common confusions

Support for incremental updates does not guarantee adaptation to drift or freedom from forgetting.

Related: Machine learning

Editorial review: 2026-10-14 · Link to this term

Knowledge and reasoning

Inference engine

How it works

It separates the method of deriving conclusions from particular domain facts and rules. Debugging requires checking both the engine's procedure and the knowledge supplied.

Example

Apply production rules to derive a recommendation.

Limits and common confusions

A correct engine can still produce misleading results from a flawed knowledge base.

Related: Knowledge representation and reasoning

Editorial review: 2026-10-14 · Link to this term

Agents and systems

Intelligent control

How it works

A controller may learn a dynamics model, select among strategies or reason over goals, while still needing explicit operational constraints.

Example

Combine a learned predictor with a controller that selects feasible actions.

Limits and common confusions

An AI component does not remove requirements for stability, robustness and safe operation.

Related: AI agent

Editorial review: 2026-10-14 · Link to this term

Responsible AI

Interpretability

How it works

Interpretability depends on the audience, representation and question being answered. An understandable model structure and a post-hoc explanation are different ways of supporting understanding.

Example

Inspect the conditions in a small decision tree.

Limits and common confusions

An intuitive explanation may still be incomplete or unfaithful to the system's actual computation.

Related: Responsible AI

Editorial review: 2026-10-14 · Link to this term

Knowledge and reasoning

Interpretation in logic

How it works

An interpretation supplies a domain and meanings for constants, predicates and functions, allowing formulas to be evaluated. Syntax alone does not fix these meanings.

Example

Interpret a predicate as the actual parent relation over people.

Limits and common confusions

Different interpretations can make the same formula true or false.

Related: Knowledge representation and reasoning

Editorial review: 2026-10-14 · Link to this term

Models and learning

Intrinsic motivation in reinforcement learning

How it works

Novelty, prediction error or information gain can supplement external task rewards. The chosen signal defines what the agent finds worth exploring.

Example

Reward visits to unfamiliar states in a simulated environment.

Limits and common confusions

A curiosity signal can favor irreducible noise or distract from the intended task.

Related: Machine learning

Editorial review: 2026-10-14 · Link to this term

Models and learning

k-means clustering

How it works

The usual objective minimizes within-cluster sums of squared Euclidean distances. Initialization and the chosen number of clusters affect the solution.

Example

Group numerical observations around k representative centroids.

Limits and common confusions

It favors certain cluster shapes and can converge to a local optimum.

Related: Machine learning

Editorial review: 2026-10-14 · Link to this term

Models and learning

k-nearest neighbors

How it works

For a new input, find k neighbors under a chosen distance and combine their targets by voting or averaging, possibly with distance weights. Feature scaling and the metric define what counts as nearby.

Example

Predict a flower class from neighboring measurements.

Limits and common confusions

Irrelevant dimensions and poorly scaled features can make distance misleading.

Related: Machine learning

Editorial review: 2026-10-14 · Link to this term

Models and learning

Kernel method

How it works

Suitable kernels represent inner products without explicitly constructing all transformed features. Support vector machines are a prominent application.

Example

Use a radial-basis kernel for a nonlinear classification boundary.

Limits and common confusions

Kernel and hyperparameter choices matter, and many methods scale poorly with sample count.

Related: Machine learning

Editorial review: 2026-10-14 · Link to this term

Knowledge and reasoning

Knowledge acquisition

How it works

Sources may include experts, documents, observations or learned patterns. The process involves eliciting assumptions and resolving ambiguities, not simply copying text.

Example

Interview a domain expert to formulate diagnostic rules.

Limits and common confusions

Expert reports can be incomplete, inconsistent or difficult to translate into formal rules.

Related: Knowledge representation and reasoning

Editorial review: 2026-10-14 · Link to this term

Search and reasoning

Knowledge base

In plain language

It stores information in a form that can support questions and reasoning.

Technical definition

A symbolic knowledge base uses a vocabulary and representation conventions to encode assertions. An ontology can define the intended categories and relationships used across contributors.

How it works

Knowledge sharing requires agreement about meanings, not merely matching field names. Provenance and revision practices help users assess where assertions came from and when they apply.

Example

Represent an artist, their works and the materials associated with each work.

Limits and common confusions

Stored assertions can be incomplete, inconsistent or outdated. A knowledge base is not automatically a source of verified truth.

Related: Rule-based system · Retrieval-augmented generation · Semantic analysis

Editorial review: 2026-10-13 · Link to this term

Learning methods

Knowledge distillation

In plain language

A smaller model can learn to imitate a larger model's useful behavior.

Technical definition

One formulation matches softened teacher output distributions, sometimes alongside ordinary labeled targets.

How it works

Relative probabilities can convey relationships between classes that a single hard label omits. The student learns from teacher outputs over selected examples; it does not literally copy the teacher's internal understanding. Distillation objectives and training data determine what behavior is transferred.

Example

Train a compact image classifier using an ensemble's probability distributions for each training image.

Limits and common confusions

The student may inherit teacher mistakes and may fail outside the distillation data. Compression can sacrifice rare-case performance even when average scores remain good.

Related: Ensemble learning · Classification · Probability calibration

Editorial review: 2026-10-13 · Link to this term

Knowledge and reasoning

Knowledge engineering

KE

How it works

The work translates domain understanding into representations and reasoning procedures, then checks behavior with experts and cases.

Example

Build and test rules for diagnosing faults in equipment.

Limits and common confusions

A logically consistent rule set can still be incomplete or wrong about its domain.

Related: Knowledge representation and reasoning

Editorial review: 2026-10-14 · Link to this term

Knowledge and reasoning

Knowledge graph

How it works

Identifiers connect statements about objects, and a schema or ontology can describe their meaning. Provenance, consistency and identity resolution influence how useful the graph is for search or reasoning.

Example

Link an artwork to its creator, date and exhibition.

Limits and common confusions

A graph's structure does not make its assertions true or complete.

Related: Knowledge representation and reasoning

Editorial review: 2026-10-14 · Link to this term

Knowledge and reasoning

Knowledge representation and reasoning

How it works

A representation determines which facts, relations and assumptions can be stated. A reasoning procedure determines what follows, with trade-offs among expressiveness, computational cost and available evidence.

Example

Represent ownership and infer which person can authorize a change.

Limits and common confusions

Conclusions are only as dependable as the facts, semantics and inference assumptions supplied.

Editorial review: 2026-10-14 · Link to this term

Knowledge and reasoning

Knowledge-based system

KBS

How it works

Represented facts and rules are used by inference procedures. The separation can make domain knowledge inspectable and maintainable.

Example

Apply documented troubleshooting rules to observed symptoms.

Limits and common confusions

Maintaining accurate knowledge and handling exceptions can be more difficult than implementing inference.

Related: Knowledge representation and reasoning

Editorial review: 2026-10-14 · Link to this term

Knowledge and reasoning

Knowledge-driven AI

How it works

Knowledge may be encoded as rules, constraints, ontologies or a structured knowledge base. It can coexist with statistical learning rather than forming an exclusive alternative to it.

Example

Apply documented eligibility rules to a structured application.

Limits and common confusions

Explicit knowledge can be incomplete, outdated or inconsistent; it does not guarantee transparency of the entire system.

Related: Knowledge representation and reasoning

Editorial review: 2026-10-14 · Link to this term

Language and prompting

Language model

How it works

Language models can score or generate sequences and supply representations for other tasks. Neural next-token prediction is one formulation; language modeling predates large Transformers.

Example

Estimate plausible continuations of a text prefix.

Limits and common confusions

Predictive fluency does not establish factual knowledge, understanding or calibrated confidence.

Related: Natural language processing

Editorial review: 2026-10-14 · Link to this term

Language models

Large language model

LLM