AI Glossary
L
A software framework for developing applications based on language models, connecting them with data and tools.
A framework for building agentic applications modeled as graphs of states and transitions, enabling complex flow control.
A model trained on a large volume of text that can understand language and generate text; the basis of modern conversational and generative AI systems.
An internal multidimensional space in which a model represents the meaning of inputs; points close to each other correspond to content-similar data.
An agent focused on legal tasks, e.g. contract analysis, legal research or document preparation.
The legal ground that permits the processing of personal data.
The question of which legal ground (e.g., legitimate interest) can support the processing of personal data contained in training data, and under what conditions such processing is permissible.
A means by which one can seek review or amendment of a decision, including decisions of supervisory authorities in the field of AI.
One of the legal bases for processing personal data, frequently relied upon for data processing in the development and operation of AI, provided it overrides the interests and rights of the data subjects.
A gradation of an agent's degree of autonomy according to the role a human holds toward it, from a mere observer through an approver and consultant to a collaborator and a fully autonomous executor of tasks.
The question of who bears responsibility for content created by AI and for its potential legal consequences, for example the infringement of third-party rights or the dissemination of false information.
The obligation to compensate for harm caused to another; with AI, it is crucial to determine which entity in the value chain is liable for the damage.
An authorization to use a copyrighted work under specified conditions.
A contract by which a rights holder grants authorization to use a work.
A trend where rights holders (e.g. publishers, music and image agencies) reach an agreement with AI providers on a paid license to use content for training or operating models, instead of going to court.
Rights protecting a person's likeness and identity against unauthorized commercial use.
Protection against the unauthorized use of a person's appearance or identity, including in the context of digital replicas and AI-generated content.
A family of Meta models released with open weights, widely used for self-hosted deployment and fine-tuning; it featured in the Kadrey v. Meta case.
The distinction between explaining one specific output (local) and explaining a model's overall behavior across many cases (global).
Automatically created records of an AI system's activity that enable retrospective checking, auditing and incident handling; they are mandatory for high-risk systems.