Key terms in plain words.

Jump to: ABCDEFGHIMOPRSTVW

Accountability

Having a named owner for every AI system and a clear path for escalation when something goes wrong.

AI agent

An AI system that can take actions on its own, such as sending messages or changing records, not just produce text.

AI literacy EU Not settled

The knowledge people need to use and oversee AI sensibly, including its limits and risks. Under the EU AI Act as amended in 2026, organizations must take measures to support it.

Algorithmic discrimination

Unfair treatment of people that results from how an AI system was built, trained or used.

Annex III (EU) EU

The list in the EU AI Act of use areas, such as employment, credit and education, where AI systems are treated as high-risk.

Automated decision-making

A decision made by technology without meaningful human involvement.

Bias

When a system treats groups of people unequally because of patterns in its data or design.

Compliance drift

A slow, step-by-step loss of control effectiveness over time, for example log retention quietly resetting to a shorter default after a migration. Unlike a one-off failure, it affects the whole period.

Conformity assessment

A check that a high-risk AI system meets the legal requirements before it goes on the market.

Consequential decision (US) US

A term used in some US state laws for decisions with significant effects on people, such as jobs, housing, credit or healthcare. Definitions vary by law.

Data fiduciary (India) IN

Under the DPDP Act, the organization that decides why and how personal data is processed.

Data minimization

Collecting and keeping only the data you need for the stated purpose.

Data principal (India) IN

Under the DPDP Act, the person the personal data is about.

Data protection impact assessment (DPIA) EU

An assessment of privacy risks required in some cases under the GDPR and UK GDPR before high-risk processing.

Deepfake

Realistic AI-generated or altered audio, images or video that show people saying or doing things they did not.

Deployer

An organization that uses an AI system under its own authority.

Drift

A change in how well a system performs because real-world data or conditions have changed since it was built.

Explainability

How well people can understand why an AI system produced a particular result.

Fine-tuning

Further training of an existing model on new data to adapt it to a task.

Foundation model

A large model trained on broad data that can be adapted to many tasks.

Frontier model

A term used in some laws for the most capable and largest AI models. Each law defines its own thresholds.

Fundamental-rights impact assessment (EU) EU

An assessment some deployers of high-risk systems must carry out under the EU AI Act before use.

General-purpose AI model

A model trained on broad data that can serve many different tasks, such as a large language model.

Governance charter

The founding document of an AI governance program: its purpose, authority, scope, targets, meeting rhythm, quorum and escalation routes. It answers what and why; the operating model answers how.

Guardrails

Technical and process controls that limit what an AI system can do or say.

Hallucination

When an AI system produces confident output that is false or made up.

High-risk AI system

A system in a use area the law treats as high risk, such as hiring, credit or safety components of regulated products.

Human oversight

People who can review, override or stop an AI system's decisions.

Impact assessment

A structured review of how a system could affect people, and how to reduce harm.

Intermediary (India) IN

Under the IT Act, a platform that stores, transmits or hosts content for others.

Market surveillance authority (EU) EU

A national body that checks products, including AI systems, meet EU rules.

Model card

A short document describing what a model is for, how it was tested and where it falls short.

Operating model

How an AI governance program actually runs day to day: workflows, team structures, decision levels and tooling. It can be adjusted more easily than the charter.

Post-market monitoring

Watching how a system performs after release and acting on problems.

Prompt injection

Input crafted to make an AI system ignore its instructions or reveal information.

Provenance metadata

Information attached to content that records where it came from and whether AI made or changed it.

Provider

An organization that develops an AI system, or has one developed, and puts it on the market under its own name.

RACI

A chart that says who is Responsible (does the work), Accountable (answers for it; exactly one person), Consulted (gives input) and Informed (is told) for each activity.

Red-teaming

Deliberately trying to make a system fail or misbehave before release, to find weaknesses.

Risk tier

The level of risk assigned to an AI system, which sets how much governance it needs.

Robustness

How well a system keeps working correctly when conditions change or someone tries to trick it.

Serious incident (EU) EU

Under the EU AI Act, a defined category of incident involving an AI system that must be reported. See the source for the exact definition.

Shadow AI

Use of AI tools at work without the organization's knowledge or approval.

Significant Data Fiduciary (India) IN

A data fiduciary designated by the government under the DPDP Act, with extra duties.

Synthetic content

Audio, images, video or text produced or changed by AI.

Synthetically generated information (India) IN

The term in India's 2026 IT Rules amendment for AI-generated or altered content that appears authentic.

Training data

The data used to build or adapt a model.

Transparency

Being open about when AI is used and how it reaches its outputs.

Value chain

Everyone involved in bringing an AI system to use, such as model developers, providers, importers, distributors and deployers.

Watermelon metric

A measure that looks green on a dashboard but hides a red reality underneath, for example 100% policy adoption alongside a growing backlog of exception requests.