AG-813
Nuclear and Radiological I&C AI Governance
EU AI Act
NIST AI RMF
ISO 42001
AGS Sector Governance | Nuclear, Radiological & Reactor Safety | Version 2.2
Nuclear and Radiological I&C AI Governance governs AI agents used in or around nuclear instrumentation and control (I&C), human-factors functions, and radiological operations — requiring full-lifecycle safety qualification, defence-in-depth such that AI failure cannot compromise safety functions, and rigorous independent verification before any deployment that could affect nuclear safety.
Nuclear is the most stringently regulated safety domain; AI must not degrade the deterministic, defence-in-depth assurance the sector requires. This dimension subjects nuclear-adjacent agents to the sector's qualification and independence expectations (informed by IAEA I&C and human-factors guidance) in addition to the cross-cutting AGS controls.
In scope: safety qualification of AI in nuclear I&C and human-factors roles; defence-in-depth isolation of AI from credited safety functions; independent verification and licensing-aligned evidence; conservative fail-safe behaviour.
Out of scope: general industrial automation and non-nuclear critical infrastructure. This dimension governs *AI agents whose function could affect nuclear/radiological safety*.
A failure that compromises a nuclear safety function can be catastrophic and irreversible. The sector relies on deterministic, independently-verified, defence-in-depth safety — properties that opaque or adaptive AI can undermine if not rigorously contained and qualified. Without sector-specific governance, an agent could be inserted into a safety-relevant role with assurance calibrated to ordinary risk. This dimension ensures AI augments operations without weakening the credited safety case.
Test 6.1: Defence-in-Depth Isolation
Test 6.2: Independent V&V Evidence
Test 6.3: Change Re-Qualification
| Score | Criteria |
|---|---|
| 0 | AI in a nuclear safety-relevant role without safety classification or qualification |
| 1 | Classified with a qualification plan but incomplete defence-in-depth/independent V&V |
| 2 | Full-lifecycle qualification, defence-in-depth isolation, independent V&V, human-factors governance |
| 3 | Licensing-aligned evidence, bounded determinism, change-triggered re-qualification, regulator-available evidence |
Scenario A — Sole-Barrier AI: An AI anomaly-detector becomes the only check on a safety parameter; when it fails silently, the defence-in-depth assumption is violated. Independent credited systems would have caught the condition.
Scenario B — Unqualified Change: A model update improves accuracy but is deployed to a safety-relevant role without re-qualification, introducing an unvalidated failure mode. Change-triggered re-qualification would have prevented it.
Scenario C — Automation Bias: Operators defer to an AI recommendation during an abnormal event without independent verification, delaying correct action. Human-factors governance preserving operator authority would have mitigated it.
| Requirement | EU AI Act | NIST AI RMF | ISO 42001 |
|---|---|---|---|
| R1: Full-lifecycle safety qualification | Art. 9 — Risk management | MEASURE 2.6 — Safety evaluation | Clause 8.3 — Verification |
| R2: Defence-in-depth preserved | Art. 15 — Robustness | MAP 3.2 — Cost of errors | A.6 — AI system lifecycle |
| R3: Conservative fail-safe | Art. 15 — Fail-safe | MANAGE 2.4 — Deactivation | Clause 8.1 — Operational control |
| R4: Independent V&V | Art. 17 — Quality management | MEASURE 1.3 — Independent assessors | Clause 9.2 — Internal audit |
| R5: Human-factors governance | Art. 14 — Human oversight | MAP 3.5 — Human oversight | A.9 — Use of AI systems |
| R6: Determinism/explainability for safety case | Art. 13 — Transparency | MEASURE 2.9 — Explainability | Clause 8.3 — Verification |
| R7: Change re-qualification | Art. 9 — Risk management | MANAGE 4.1 — Post-deployment monitoring | Clause 8.3 — Verification |
| R8: Regulator-available evidence | Art. 11 — Technical documentation | GOVERN 1.1 — Legal/regulatory | Clause 7.5 — Documented information |
Article 9 (risk management proportionate to catastrophic risk) and Article 15 (robustness/fail-safe) apply with maximal stringency in nuclear contexts; AG-813 adds the sector's qualification, defence-in-depth, and independence expectations.
MEASURE 2.6 (safety evaluation) and MAP 3.2 (cost of errors/unintended functionality) frame the rigorous safety qualification nuclear AI demands.
Clause 8.3 (verification) and Annex A.6 (lifecycle) require lifecycle verification proportionate to impact — here, nuclear safety qualification.