Maximum Credible Harm
Plain-language explanation
The most severe reasonably foreseeable harm an AI system could cause or materially contribute to, used as a starting point for reasoning backward through the conditions, capabilities, failures, and controls that shape the risk.
Why It Matters
Starting with harm can prevent an assessment from being constrained by the controls an organization already has and can make severe risk pathways easier to test.
In Practice
Maximum credible harm is an AuditDIFF practitioner concept, not a statutory definition and not a term defined by the NIST AI Risk Management Framework.
The analysis begins by asking what severe harm is reasonably foreseeable in the system’s actual context. Practitioners can then work backward through the scenario, required AI capabilities, deployment conditions, control failures, evidence, and residual risk.
The concept should not be used to imply that a scenario is likely or that it meets a legal definition of catastrophic risk.