AI Glossary
I
The obligation to report serious incidents and malfunctions of an AI system to the competent authorities within set deadlines, so that risks can be assessed and mitigated.
A set of procedures for reporting serious incidents related to AI.
An arrangement by which one party assumes the risk of third-party claims (e.g., for copyright infringement) related to the use of an AI tool.
A variant of the attack in which malicious instructions are not entered directly by the user but are hidden in content that the model processes (a website, document, or email), and the model incorrectly interprets them as a command.
The process in which an already trained model generates a response based on the user's input.
The expenses associated with each query to a deployed model; a significant factor in product design and in choosing between larger and smaller models.
A principle seeking to ensure that AI does not undermine the reliability of the information environment, for example by spreading manipulative or deceptive content, while respecting freedom of expression.
A control layer that checks what enters the model and what comes out of it, and blocks content that violates the established rules.
A specific type of model tuning based on training responses to users' prompts and instructions.
The purpose for which an AI system is intended and used according to the information provided by its provider.
An appeal against a partial decision before the dispute concludes; in the Ross Intelligence case the court permitted such an appeal given the difficulty of the legal questions.
Mechanisms used to verify that an organization's activities comply with legal and internal rules.
An effort to align the regulatory approaches of different jurisdictions (EU, USA, OECD, Korea) so that they do not conflict and so as to reduce the burden of cross-border operations.
A periodically updated document summarizing the expert consensus on the capabilities and risks of advanced AI, serving as a basis for policymakers.
A trend in which content companies enter into equity or licensing relationships with AI developers (for example, an agreement to use characters in a video generator), turning disputes into commercial cooperation.
An international standard providing guidance on AI risk management that links general risk-management principles with considerations specific to artificial intelligence.
The first international standard for an artificial intelligence management system (AIMS), published at the end of 2023. It enables organizations to obtain certification from an independent third party as evidence of responsible AI governance.
A companion standard to ISO/IEC 42001 focused on AI system impact assessment, providing guidance for evaluating their effect on individuals and society.
J
A technique for bypassing a model's safety and content restrictions in order to make it generate prohibited output.
An attempt to bypass a model's safety restrictions using specially crafted prompts so that it produces content it would otherwise refuse.
K
A dispute brought by thirteen authors (including R. Kadrey) against Meta over training the Llama models on books from so-called shadow libraries. In June 2025, Judge Vince Chhabria found the training to be fair use but emphasized that the plaintiffs could have succeeded had they demonstrated market dilution caused by flooding the market with competing outputs.
A body of practical knowledge, experience, and procedures usable in business.
A data structure that links information through relationships between individual entities.
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.