Why Probabilistic Artificial Intelligence Cannot Be Audit Grade

Artificial intelligence systems are now used in areas where decisions may later be reviewed by regulators, courts, or auditors. In these settings, it must be possible to go back and examine a specific decision: to reproduce it, verify how it was made, and challenge it using only the information that existed at the time of the decision. Although courts do not use the term “audit grade,” these expectations rely on a simple underlying requirement: once the relevant evidence is fixed, the outcome must also be fixed. We call this requirement record-determinism. In other words, the preserved record must uniquely determine the decision result. Starting from this requirement, we show that any system whose outcome remains probabilistic after the record is fixed cannot meet replay-based audit expectations. If the same record can lead to different outcomes, the decision cannot be reliably reconstructed or independently verified. We formalize this idea by showing that auditability requires zero uncertainty in the outcome once the record is known. We also show that probabilistic techniques can be used in audit-ready systems only if every source of randomness is fully captured in the record. When this happens, the system’s behavior becomes deterministic relative to that record. This result does not depend on any specific technology and clarifies the difference between replay-based auditability and statistical or population-level audits commonly used in science and finance.