Glossary

What is AI?

We define Artificial Intelligence, AI Model, AI Tool, and Tool Capability. Every other term is defined inside the Framework itself, in the category it belongs to.

Explore the Framework here.

Artificial Intelligence (AI)

Computational system that produces outputs from inputs using methods not fully specified by fixed logic or rules, but instead learned or optimized from examples, feedback, or reward signals during a training or adaptation phase. This includes "traditional" machine learning and reinforcement learning based systems as well as deep learning architecture-based systems such as transformer and diffusion systems.

The Framework considers several flavors of AI including GAI. The definition above casts that wide net and is not a definition from any known organization.

AI Model

A specific instance of AI, typically deployed on one or more machines.

AI Tool

A product, service, platform, plug-in, API, or workflow environment made available for use that leverages, embeds, or is integrated with an AI Model or AI Models to provide one or more functions or features. If an AI Model is made available directly for use, this Framework treats that AI Model itself as an AI Tool. The term AI Solution is used synonymously with AI Tool in this document.

Tool Capability

A distinct feature within an AI Tool such as upscaling, denoising, inpainting, or text-to-video generation.

Why this Framework does not define "Generative AI"

There are varying industry, standards body, and government definitions of Generative AI (GAI). Some of those definitions refer to GAI as a subset of artificial intelligence that learns patterns from data and produces content based on those patterns. That framing of GAI, while may be suitable for certain needs, does not capture the nuance around what content is produced, what patterns are utilized, and what aspects of the inputs supplied during AI use are utilized for that content production. This kind of nuance is necessary for governance.

Therefore, this Framework avoids defining GAI, but provides two different ways to frame GAI.

See how to characterize a Tool Capability here.

While the Framework does not currently define the GAI term, individual studios and other organizations can of course choose to do so.

Tool, Tool Capability, AI Model

A vendor describes the product. A studio evaluates the one feature from that product it intends to use, and often calls that feature "the tool". The same word is likely to be interpreted differently.

The Framework takes the position that the broader product is the AI Tool, and the specific functions within it are Tool Capabilities. That lets a vendor document the product and all its capabilities while a studio focuses on the capability that is actually in scope.

An AI Model on one end might be the product itself (for instance, a foundational model made available by a vendor). And on the other end, one or more models might be enabling a single capability within the tool.

The Framework distinguishes the three levels, but also allows for them to collapse into one, like in the example of a speech-to-text single capability model.

  1. AI Tool What is made available for use — a product, service, platform, plug-in, API, or a model offered directly.
  2. Tool Capability A distinct feature within a tool, such as upscaling or voice cloning. One or more per tool.
  3. AI Model What enables a capability. One or more per capability.