AI Governance

AI governance is the set of policies, roles, processes, and records an organization uses to decide which artificial intelligence systems it builds or buys, who is accountable for each one, what evidence is required before release, and how systems are monitored, changed, and retired. AI ethics defines the values an organization wants its AI to respect. Governance turns those values into decision rights and documented checks that apply to every system across its lifecycle, from a large language model assistant bought from a vendor to an in-house computer vision model on a production line.

Frameworks and Regulation#

Most governance programs build on a small number of reference frameworks:

• NIST AI Risk Management Framework: A voluntary US framework released in January 2023 and organized into four functions. Govern sets policies and accountability, Map establishes each system's context and intended use, Measure assesses risks with tests and metrics, and Manage prioritizes and treats those risks.
• ISO/IEC 42001: Published in December 2023, the first international standard for an AI management system. Organizations can be audited and certified against it, much as they are for information security under ISO/IEC 27001.
• OECD AI Principles: Adopted in 2019 and updated in 2024, they sit underneath many national policies and give the other frameworks a shared vocabulary.

NIST's Generative AI Profile, NIST AI 600-1 (July 2024), applies the risk management framework to risks such as hallucinations and prompt injection.

The EU AI Act is a risk-based regulation that entered into force in August 2024 and applies in phases. Bans on practices such as social scoring and emotion recognition in workplaces and schools, together with an AI literacy duty, have applied since 2 February 2025, obligations for providers of general-purpose AI models since 2 August 2025, and transparency duties such as disclosing deepfakes since 2 August 2026. The Digital Omnibus on AI, in force since 27 July 2026, moved the requirements for high-risk systems listed in Annex III, such as hiring and credit scoring, to 2 December 2027, and those for AI in regulated products such as medical devices and machinery to 2 August 2028. Those requirements cover risk management, data governance, logging, human oversight, and conformity assessment.

Core Controls#

An inventory comes first, since an organization cannot govern systems it does not know about. A complete inventory records each system's owner, purpose, data sources, model versions, and vendors, and it surfaces shadow AI: tools that teams adopted without review. For models built in-house, a model registry often serves as that inventory.

Risk classification then sorts systems into tiers by purpose, autonomy, affected people, and the cost of errors. A model that ranks internal documents needs lighter review than one that screens job applicants or controls machinery, and higher tiers get stricter testing, sign-off, and human oversight.

Documentation makes decisions reviewable. Model cards record intended use, evaluation results, and known limitations, and data provenance records capture where training data came from and under what license or consent. Reproducibility belongs in the same record: the exact dataset version, base model, code version, and settings behind each candidate model, so a reviewer can trace any deployed model back to its inputs. For vision models trained on Ultralytics Platform, each run records its parent model, Ultralytics version, host environment, and Git commit when reported, and a saved dataset version pins the exact images and labels it trained on.

Testing supplies the evidence for approval. Acceptance thresholds are set for each use case before results come in, and tests cover accuracy across relevant subgroups to detect bias in AI, robustness to shifted inputs such as low light or a new camera angle, and security, often through AI red teaming. The UK NCSC and partner agencies publish guidelines for secure AI system development covering design, development, deployment, and operation.

After release, model monitoring tracks performance and data drift, an incident process defines who can pause or roll back a system, and a retirement process removes models and data that are no longer needed. Decision rights tie the controls together: for every system, governance names who approves deployment, who can stop it, and who accepts the risk that remains after testing.

How Governance Differs From Related Practices#

• Regulation: Externally imposed law. Governance is the internal system an organization runs, and it is also how the organization shows it complies.
• Risk management: Identifying, measuring, and treating specific risks. Governance decides who performs that work, who accepts the remaining risk, and how decisions are recorded.
• MLOps: The pipelines that train, deploy, and maintain models. Governance sets the approval gates and records those pipelines must satisfy.
• AI safety: The technical study of making AI systems behave as intended. Governance decides how safety findings shape what an organization is allowed to ship.