Responsible AI.

DefinitionResponsible AI names a practice.

It describes the discipline through which an organization decides what a system may be used for, what evidence supports that decision, who is accountable for it, and under what conditions it should be stopped. Responsibility attaches to the organization, not to the system.

This distinguishes the term from adjacent ones it is often used interchangeably with. Ethics supplies the values; governance supplies the decision rights and controls; assurance supplies the independent test. Responsible AI is the practice that connects them across the lifecycle of a system.

HistoryThe terminology is recent; the underlying questions are not.

Concern about delegating consequential judgment to computation predates machine learning by several decades. The change after 2015 followed the deployment of systems making decisions at scale in employment, credit, policing, health and welfare, and the consequent shift from philosophical caution to institutional codification.

ConvergenceFive principles recur across the guidance; their interpretation does not.

A systematic review published in Nature Machine Intelligence in 2019 mapped eighty-four sets of AI ethics principles issued by companies, research institutions and public bodies. It found convergence on five principles — transparency, justice and fairness, non-maleficence, responsibility, and privacy — alongside substantive divergence in how each is interpreted, why it is held to matter, whom it applies to, and how it should be implemented.

The distinction is material. Agreement at the level of principle is genuine and close to universal, but it establishes little in itself: two organizations may hold identical published commitments while operating at materially different levels of control.

Core principlesThe working set, and what each one requires in practice.

The formulations below are the recurring ones across the OECD Principles, the UNESCO Recommendation, the NIST characteristics of trustworthy AI and the EU AI Act’s obligations for high-risk systems. Naming is not standardized; the substance largely is.

TerminologyTerms frequently used interchangeably, and the distinctions between them.

Much of the apparent disagreement in governance discussions is terminological rather than substantive. The distinctions below are those that materially change the meaning of a claim.

InstrumentsWhat is guidance, what is standard, and what is law.

These instruments differ in legal force. Treating a voluntary framework as binding, or a binding obligation as advisory, is a common and consequential error.

PracticePrinciple becomes governance only when it resolves into evidence.

Every principle above resolves, or fails to resolve, into the same four questions: which control implements it, who owns that control, what artifact demonstrates the control operating, and what happens when it fails. An organization that can answer those four for a named system has a governance position. One that can only cite the principle has a statement of intent.

This is why the useful unit of assessment is a system rather than an organization. Principles are held at company level; evidence exists at system level, for a particular version, population, workflow and period.

How this connectsWhere AIGX applies Responsible AI.

AIGX Research addresses the final question in that sequence: what evidence supports a governance claim, and how that evidence can be examined. Our sector frameworks translate these principles into controls with named obligations and required artifacts; the methodology sets out how an assessment is scoped, evidenced and reviewed.

This page is reference material. It presents the consensus of the field and does not state an AIGX position.