AI Web glossaryAa

Web experiences where AI helps create or adapt content and interfaces.

A Web also used by agents that can discover tools and carry out authorized tasks.

A statistical model trained on large quantities of text and sometimes other modalities.

Neural network architecture built around attention mechanisms.

A method combining document retrieval with answer generation.

Compact numerical representation of text, images or other objects.

Search based on proximity between numerical representations.

Combination of multiple retrieval approaches, such as keyword and vector search.

Reassessment of the ranking of retrieved documents.

A piece of a document prepared for indexing or processing.

Amount of tokens a model can process within an exchange.

Unit of text or another representation processed by a model.

Instruction or set of information supplied to an AI system.

Attempt to redirect a system through instructions embedded in untrusted data.

Plausible generated answer that is factually wrong or unsupported.

Reference to a source; on its own it does not guarantee accuracy.

Information describing a piece of content’s origin and certain transformations.

Provenance information associated with certain images and other media.

Set of specifications for technical provenance of digital content.

Open protocol for connecting AI applications to tools and resources.

Protocol for communication and collaboration between independent agent systems.

Proposal for exposing some Web functionality to agents.

System that plans or sequences actions within defined boundaries.

External function an agent may call with authorized parameters.

System of multiple coordinated or interoperable agents.

Step where a person reviews or approves a system’s decision.

Web API for GPU graphics and compute operations.

Web API designed for neural network operations.

Portable executable format used in browsers and elsewhere.

Running model computations on the user’s own device.

Reduction of numerical precision in model parameters or computations.

Generative model family trained through denoising processes.

Ability to process multiple kinds of signals, such as text and images.

Content generation informed by extra tools or information.

Interpretation of meaning beyond raw representation.

JSON format for representing linked data.

Structure of entities connected by defined relationships.

Practices to improve site accessibility and interpretation by search engines.

Editorial term for practices aimed at visibility in answer engines.

Research and practice concerning visibility in generative responses.

Contract allowing software components to communicate.

Framework for delegated authorization to resources.

Vocabulary for describing and validating JSON structures.

Syndication format for distributing updates.

Syndication and publishing format.

Research manuscript shared before or independently of peer review.

Measurement of system quality against defined criteria.

Task and metric set used to compare systems.

Adaptation of a pretrained model using additional data.

Mechanism to constrain specific behaviors or risks of a system.

Markup language that structures Web documents.

Design that enables access for people with diverse abilities.

Structured description of A2A agent capabilities.

Proposed API for JavaScript tools exposed to agents.

Proposed convention, not required by Google Search.

Discovery optimization without ranking guarantees.

Crawler for ChatGPT search.

Crawler token for training collection.

Google preference token for certain AI uses.

Verifiable link between an answer and evidence.

Acknowledging missing evidence instead of fabricating.

Separation between instructions and external content.

Hostile instruction inside retrieved content.

Explicit oversight for sensitive actions.

Permissions scoped to actual needs.

Running the model on the device.

API for accelerated neural computation.

GPU computation and rendering API.

Browser-managed AI capabilities.

Linked data using a JSON context.

Signed assertions on media history.

Technical history distinct from factual truth.

Human review of sources, rights and wording.

EU transparency requirements for certain content.

Clear and predictable interaction design.

Alternative when an API or model is unavailable.

Solution proportionate to task and resources.