HIRAF key terms

Computer systems that perform tasks normally requiring human intelligence, such as reasoning, learning, perception, or language use. The term covers both narrow, task-specific systems and theoretical general systems.

AI designed for a specific, bounded task such as translation, recommendation, or image classification. It cannot transfer its competence to unrelated problems, unlike hypothetical general-purpose AI.

A branch of AI in which algorithms improve performance on a task by learning patterns from data rather than following hand-coded rules. Learned behavior is captured in model parameters.

A training approach that uses labeled examples pairing inputs with correct outputs, so the model learns to predict labels for new inputs. Common tasks include classification and regression.

A training approach that finds structure in unlabeled data, such as clusters, associations, or lower-dimensional representations. No target answer is supplied during learning.

A training approach where an agent learns behavior by taking actions in an environment and receiving rewards or penalties, gradually favoring actions that increase cumulative reward.

A mathematical structure with learned or configured parameters that maps inputs to outputs for a task. It is the artifact produced by training and used for inference.

The dataset used to fit a model's parameters. Its quality, representativeness, and licensing strongly shape model behavior and later risk.

Data records that include the correct or desired output alongside the input, enabling supervised learning. Label accuracy and consistency directly affect model quality.

The human or automated process of attaching labels, tags, or judgments to raw data so it can be used for training or evaluation. Requires clear guidelines and quality checks.

A failure mode where a model fits idiosyncrasies and noise in training data so closely that it performs poorly on new, unseen data. Usually detected by comparing training and validation performance.

A model's ability to perform well on data it did not see during training. It is the central goal of machine learning and is estimated through held-out validation and test sets.

The process of assessing a model against requirements and held-out data before deployment, checking accuracy, robustness, fairness, and fitness for the intended use.

A large model pretrained on broad data that can be adapted to many downstream tasks through fine-tuning or prompting. Examples include large language and image-generation models.

Further training of a pretrained model on a smaller, task- or domain-specific dataset to adapt its behavior. It can improve relevance but may narrow or distort original capabilities.

The input text or instruction given to a generative model to elicit an output. Prompt design strongly influences response quality, safety, and consistency.

The act of running a trained model on new inputs to produce predictions or generated content. It is the production-time counterpart to training.

Systematic error or unfair skew in data, model outputs, or decisions that disadvantages particular groups or distorts results. It can arise at data collection, labeling, design, or deployment stages.

The principle that AI systems should avoid unjustified disparate treatment or outcomes across groups. It is operationalized through metrics, testing, and mitigation choices that involve trade-offs.

The degree to which the reasons for an AI system's output can be understood by relevant people. It supports accountability, contestability, and effective human oversight.

The practice of disclosing sufficient information about an AI system, such as its purpose, data, limitations, and logic, so stakeholders can assess and challenge it.

The assignment of clear responsibility for an AI system's design, operation, and outcomes to identifiable people or organizations, including remedies when harm occurs.

Arrangements that keep people meaningfully in control of an AI system, such as review, approval, intervention, or shutdown authority. It is a safeguard against harmful or erroneous outputs.

A design pattern in which a person reviews, confirms, or overrides system outputs before consequential action is taken. Effective use requires time, competence, and authority to intervene.

The tendency of people to over-trust automated outputs and under-weight their own judgment or contradictory evidence. It can weaken the protection that human review is meant to provide.

The structured process of identifying, assessing, treating, and monitoring risks that AI systems pose to people, organizations, and society across the system lifecycle.

The systematic evaluation of potential harms, considering likelihood, severity, affected parties, and existing controls. It informs prioritization and decisions about whether to proceed.

The selection and implementation of responses to assessed risk, such as avoiding, reducing, transferring, or accepting it. Chosen controls should be documented and monitored for effectiveness.

The potential for harm or loss from incorrect, misused, or poorly understood model outputs. It covers model error, inappropriate application, and failures in governance or monitoring.

Degradation of model performance over time as real-world data or behavior diverges from what the model learned. Monitoring and retraining are typical responses.

Continuous observation of a deployed model's inputs, outputs, and performance indicators to detect drift, errors, misuse, or emerging harm, and to trigger corrective action.

The policies, roles, structures, and processes an organization uses to direct and control AI development and use in line with its objectives and obligations.

Legally binding rules imposed by public authorities that govern AI development, deployment, or use, often through obligations, prohibitions, and enforcement mechanisms.

An organization's or government's stated positions, principles, and requirements guiding decisions about AI. Policy sets direction but may not itself carry legal force.

A regulatory or management strategy that calibrates obligations and controls to the level of risk a use case presents, imposing stricter requirements on higher-risk applications.

An AI system whose use can significantly affect safety, fundamental rights, or critical interests, thereby attracting stricter legal or governance obligations. Classification depends on context and applicable rules.

The process of demonstrating that a system meets specified requirements before it is placed on the market or put into service, often through documentation, testing, and independent review.

The set of roles, policies, and controls that manage data availability, usability, integrity, and security. It establishes who may access and use data, and under what conditions.

The fitness of data for its intended use, judged by attributes such as accuracy, completeness, consistency, and timeliness. Poor quality propagates into model errors and unfairness.

The documented origin and history of a dataset, including sources, collection methods, and transformations. It supports lawful use, traceability, and reproducibility.

The protection of personal information throughout collection, use, sharing, and retention, in line with legal and ethical expectations. It includes minimizing data and limiting purposes.

Information relating to an identified or identifiable individual. Its processing is subject to legal duties such as lawful basis, notice, and rights of the individual.

Structured records describing a model's purpose, data, design, performance, limitations, and intended use. It supports review, audit, and safe reuse by others.

Practices and tooling for reliably building, deploying, monitoring, and maintaining machine learning systems in production, integrating data, model, and software engineering workflows.