Self-evolving and AI safety functions: the new high-risk machinery

This is general information about Regulation (EU) 2023/1230, verified as at 12 July 2026, not legal advice. Confirm the exact category wording against Annex I.

Why this is new

The Machinery Directive predates the current wave of machine learning in industrial equipment, and its Annex IV had no concept of a safety function that changes its own behaviour after the machine is placed on the market. The Regulation closes that gap by naming self-evolving, machine-learning safety functions as high-risk and putting them in the strictest tier, Part A. Placing them in Part A rather than Part B is deliberate: the risk is judged serious enough that harmonised standards alone should not let a manufacturer self-declare.

The two categories

• Safety components with self-evolving behaviour. A safety component that is placed on the market in its own right and whose behaviour, ensuring a safety function, is fully or partially self-evolving using a machine-learning approach.
• Embedded self-evolving systems. Machinery with an embedded system that has fully or partially self-evolving behaviour using machine learning and that ensures safety functions, where that system is not placed on the market independently. The assessment focuses on the safety-related self-evolving behaviour.

The common thread is a safety function whose logic can change through learning after the machine leaves the factory. A fixed, pre-trained model that does not continue to learn in the field is a different case, and whether it falls in these categories depends on the exact wording in Annex I, so read it there.

What it forces

Because these categories are in Part A, a notified body is always involved under Article 25(2). You cannot reach a CE mark by self-assessment, and applying harmonised standards does not change that. This is the same rule that governs the other Part A categories; see the notified-body guide.

Related duties: cybersecurity and evolving behaviour

Alongside the classification, the Regulation adds essential requirements addressing protection against corruption and against attempts to create a hazardous situation, and covers software updates and machinery with evolving or learning behaviour. In practice a machine with a learning safety function faces both the Part A conformity route and these essential requirements, so the safety case has to show that learning cannot degrade the safety function and that the function resists tampering. The specific Annex III clauses should be cited from the final text when you build your technical file.