Glossary term
Glossary term
Governance and Compliance
The rules used to judge significance of AI risk, including likelihood, impact, severity, reversibility, affected population, legal exposure, safety implications, and stakeholder tolerance. Criteria should avoid purely financial scoring because AI risk often turns on severity of harm, scale, reversibility, vulnerability of affected groups, autonomy, and legal defensibility.
Canada's Algorithmic Impact Assessment scores risk on four levels using criteria including impact reversibility, duration of impact, and the rights of individuals affected.
The EU AI Act applies risk criteria via Annex III, classifying systems as high-risk based on intended purpose in domains such as employment, education, and law enforcement.
Anthropic's Responsible Scaling Policy uses AI Safety Level (ASL) thresholds tied to dangerous capability evaluations rather than purely financial impact metrics.