AI Glossary

A conceptual division of an AI system's existence into phases, from design and data collection through development and deployment to operation and decommissioning, used to assign risks and responsibilities.

An intergovernmental set of value-based principles for trustworthy AI that has become a reference point for many national and supranational regulations and unifies the basic terminology.

A tool that allows AI systems to be described and compared along dimensions such as data, the role of humans, the deployment context, and potential impact on society.

Running a model on an organization's own infrastructure instead of the public cloud, chosen in particular for reasons of data confidentiality and control.

A model whose weights or code are publicly available, allowing it to be self-deployed, modified, and fine-tuned.

A license allowing software to be freely used, distributed and modified under specified conditions, also applied to some AI models and tools.

A model whose trained weights are publicly available for download and self-deployment, even though its training data or code need not be published.

OpenAI's programming interface, which allows developers to integrate its models into their own applications and services.

The ability of a rightholder to reserve that its works may not be used for text and data mining; the reservation must be expressed in a machine-readable manner.

A situation in which a model adapts too closely to the training data and its ability to work with new data deteriorates.

The risk that an advanced system trained to optimize a certain goal begins to circumvent control mechanisms or human oversight in order to achieve the goal; particularly monitored in agentic systems.