AI Glossary

An agent specializing in breaking a goal down into sub-steps and assembling a procedure for carrying them out.

Methods that seek to clarify after the fact why a model arrived at a particular output, without themselves changing its internal functioning.

A documented procedure by which the provider systematically monitors the functioning of a high-risk system after deployment and evaluates newly identified risks.

Ongoing monitoring of an AI system's functioning after it has been placed on the market.

A documented procedure by which the provider systematically monitors the functioning of a high-risk system after deployment and evaluates newly identified risks.

The use of AI to predict the likelihood that a specific person will commit a crime; in certain forms prohibited as incompatible with the presumption of innocence.

A court order requiring an AI company to preserve and segregate records (logs) of a system's outputs as potential evidence. In NYT v. OpenAI, the court ordered ChatGPT logs to be preserved and later allowed routine deletion to resume.

A tool that makes it easier for an injured party to prove the relationship between a failure of an AI system and the resulting harm where it is difficult to explain the internal workings of the system.

A legal effect whereby a system meeting a harmonized standard is deemed to comply with the corresponding requirements of the regulation until the contrary is proven.

A legal presumption that an entity meets certain requirements if it complies with harmonized standards.

The initial phase of training a model on a large general corpus of data, followed by fine-tuning for specific tasks or behaviors.

Any operation performed on personal data, such as collection, storage, sorting, or deletion.

A special liability regime for harm caused by a defective product.

A process in which an already supplied AI system is withdrawn from users due to identified risks or non-conformity.

Liability arising from the pursuit of a professional or business activity.

The automated processing of personal data to evaluate certain aspects of a person.

Practices that are entirely prohibited by the AI Act, such as subliminal manipulation, certain forms of social scoring, and prohibited biometric categorization.

An instruction given by a user to an artificial intelligence system in order to produce an output.

The storing of repeatedly used parts of an input in order to speed up and reduce the cost of subsequent queries to the model.

A term for the legal questions surrounding the copyright protection of prompts.

Methods of creating and optimizing instructions that lead to higher-quality AI outputs.

A security attack in which an attacker inserts hidden instructions into the input in order to bypass the model's intended behavior or its restrictions.

The contested question of whether and when the outputs of a model themselves can enjoy copyright protection, and to whom any rights belong.

We examine how courts have approached the question in our articles A Czech court has ruled that AI-generated graphics are not a copyrighted work and Thaler v. Perlmutter.

Indirect discrimination in which a model uses a seemingly neutral data point that is nonetheless closely related to a protected characteristic (e.g. a ZIP code as a substitute for ethnicity).

The modification of data so that a specific individual cannot be identified without additional information.

Data freely accessible on the internet whose use for AI training may nevertheless be subject to copyright or personal data protection rules.

A principle under which personal data may be processed only for compatible purposes; using previously collected data to train AI raises the question of compatibility.

The commencement of use of an AI system for its intended purpose; the moment from which the relevant obligations apply to its operation.

A technique for reducing the size of AI models in order to lower their computational demands.

An architecture that combines a language model with an external knowledge base from which the AI retrieves relevant information before responding.