Industrial machinery ML — what evidence will a notified body expect?

Buyers ask what a notified body will want to see for a learned component in a machine’s safety function. The public pages we could read give the frame, not a checklist.

Software evidence. The IEC catalog says IEC 61508-3:2010 “applies to any software forming part of a safety-related system or used to develop a safety-related system” and “requires that the software safety functions and software systematic capability are specified”. Its title is “Functional safety of electrical/electronic/programmable electronic safety-related systems - Part 3: Software requirements”.

Where AI sits. The catalog entry for ISO/IEC TR 5469:2024, “Artificial intelligence - Functional safety and AI systems”, describes “use of AI inside a safety related function to realize the functionality;” and “use of non-AI safety related functions to ensure safety for an AI controlled equipment;”.

EU law. The AI Act Service Desk Annex I lists “Regulation (EU) 2023/1230 of the European Parliament and of the Council of 14 June 2023 on machinery” as item 21 in Section B (the page text we read is cut off there). The machinery obligations themselves are in Regulation (EU) 2023/1230, and robot-specific safety requirements in ISO 10218-1:2025. The EUR-Lex and ISO servers did not return readable text in this review, so we quote nothing from them.

Editorial judgment: The evidence story with the most public support is the second TR 5469 area: keep the learned function outside the safety function, and give the notified body a conventional IEC 61508-3 software argument for the non-AI function that bounds it. We have not found an official sentence that tells a notified body how to accept a trained model inside the safety function itself.