AI Compliance Glossary

A statistical finding that a selection process, test, or AI system produces substantially different pass rates, selection rates, or outcomes for different demographic groups, typically measured using the four-fifths (80%) rule from EEOC Uniform Guidelines.

An organization or individual that integrates an AI system into its products, services, or internal processes for use in a specific application — as distinguished from the AI provider (developer) who built the underlying model or system.

The frameworks, policies, processes, and organizational structures that an organization establishes to ensure its AI systems are developed and used responsibly, in compliance with applicable laws, and aligned with ethical principles.

The knowledge, skills, and understanding needed to evaluate, use, and critically assess AI systems — including awareness of how AI works, its capabilities and limitations, and its potential social and ethical impacts. Required training for employees operating high-risk AI systems under the EU AI Act.

A controlled regulatory environment created by a government authority allowing AI developers — particularly startups and SMEs — to test innovative AI systems under real-world conditions with reduced regulatory burden and direct oversight, in exchange for transparency and data sharing with regulators.

A documented, continuous process for identifying, analyzing, evaluating, and mitigating risks associated with an AI system throughout its entire lifecycle, from design through deployment and decommissioning.

When an automated system treats people differently based on protected characteristics — such as race, gender, age, or disability — resulting in unfair or unlawful outcomes.

The use of computerized systems to make decisions about individuals with little or no human involvement, including decisions based on AI, machine learning, or rule-based algorithms that have legal or similarly significant effects.

An impartial evaluation of an AI system's outputs to detect and measure discriminatory patterns across protected demographic groups. Required annually by NYC Local Law 144 for automated employment decision tools.

A mandatory conformity marking required under the EU AI Act for high-risk AI systems before they can be placed on the EU market, indicating that the system meets all applicable EU regulatory requirements including technical, safety, and fundamental rights obligations.

A formal process to verify that an AI system meets regulatory requirements before it is placed on the market or put into service. Required for high-risk AI systems under the EU AI Act.

Under the Colorado AI Act, a decision made by a high-risk AI system that has a material, legal, or similarly significant effect on a consumer's access to or the cost of education, employment, financial services, healthcare, housing, insurance, or legal services.

The ability to describe how an AI system reached a specific output or decision in terms that are understandable to a human audience — whether technical experts, affected individuals, or regulators. Required in various forms by the EU AI Act, GDPR, and US state AI laws.

A structured evaluation required by the EU AI Act for public bodies and private entities deploying high-risk AI systems, assessing the potential impact on fundamental rights including privacy, non-discrimination, freedom of expression, and access to justice.

An AI model trained on large amounts of data at significant compute scale that can perform a wide range of distinct tasks, including text generation, translation, image creation, and code writing. Large language models like GPT-4 and Claude are leading examples.

An AI system that poses significant potential harm to individuals or society and is subject to heightened regulatory requirements. Different laws define this category differently.

A structured evaluation of the potential harms, risks, and benefits of an AI system, required by several AI regulations before deploying high-risk systems.

The ongoing process of collecting and analyzing data about an AI system's performance, behavior, and real-world impacts after deployment, to detect and address risks, drifts, biases, or failures that were not apparent during pre-deployment testing.

AI applications that are banned outright because they pose an unacceptable risk to fundamental rights, safety, or human dignity. The EU AI Act establishes the most comprehensive list of prohibited AI practices globally, effective February 2, 2025.

A structured adversarial testing process where a dedicated team attempts to find ways to make an AI system behave harmfully, unsafely, or contrary to its intended design — including generating dangerous content, bypassing safety filters, or being manipulated into misuse.

Under the EU AI Act, systemic risk refers to risks posed by general-purpose AI models that, due to their capabilities, widespread deployment, or scale of training compute (above 10²⁵ FLOPs), have the potential to cause widespread negative effects across multiple sectors of society simultaneously.

The practice of publicly disclosing information about the datasets used to train an AI system, including data sources, licensing status, known biases, and data governance practices. Required by California AB 2013 and the EU AI Act for covered AI developers.