normakte / Machinery with AI
Not every piece of “AI” software turns a machine into high-risk machinery. What matters is whether a learning system ensures safety functions – then procedure and requirements change.
Two categories concern learning systems: point 5 safety components with fully or partially self-evolving behaviour using machine learning approaches ensuring safety functions; point 6 machinery with such embedded systems not placed on the market separately – only in respect of those systems. A notified body is mandatory for both (Art. 25(2)).
Where machinery embeds a system whose safety components were already assessed by a third party when placed on the market separately, the machinery need not be recertified for that reason alone (recital 54).
| Requirement | Content | Reference |
|---|---|---|
| Limits | The control system must not cause the machine to act beyond its defined task and movement space | 1.2.1, 2nd para. (a) |
| Recording | Data on the safety-related decision-making process are recorded and kept for one year for the authorities | 1.2.1, 2nd para. (b) |
| Correction | The machine must be correctable at any time to maintain its inherent safety | 1.2.1, 2nd para. (c) |
| Learning phase | Changes to settings or rules, including during the learning phase, are prevented where they could be hazardous | 1.2.1, 1st para. (d) |
| Traceability | Interventions and uploaded versions of safety software remain traceable for up to five years | 1.2.1, 1st para. (f) |
| Interaction | Adapt the interface to the foreseeable characteristics of operators; where appropriate communicate planned actions understandably | 1.1.6(f), (g) |
The assessment covers hazards that may arise during the life cycle and are foreseeable at placing on the market because the behaviour or logic evolves by design (Annex III Part B, point 1). The limits of safety functions are defined in the risk assessment (1.2.1, 1st para. (d)). On reasoned request, source code or programming logic must be made available to the authority (Art. 10(3)).
The AI Act, Regulation (EU) 2024/1689, classifies as high-risk AI systems that are safety components of products covered by the harmonisation legislation in its Annex I. With the Digital Omnibus on AI, given final approval by the Council on 29 June 2026 (Regulation (EU) 2026/1744), the Machinery Regulation is, according to available summaries, moved within that annex: AI in machinery would then be governed mainly through the Machinery Regulation, which the Commission may supplement by delegated acts; obligations for such high-risk systems would apply from 2 August 2028. Check the text published in the Official Journal before basing your planning on it.
No. Part A only applies where a system with self-evolving behaviour through machine learning ensures safety functions. A quality-inspection AI without a safety function is not covered.
Recital 55 excludes software that is neither capable of learning nor of evolving. For the specific model, this must be justified technically and documented; in case of doubt, clarify with a notified body.
The one-year recording concerns control systems with self-evolving behaviour designed for varying levels of autonomy. Five-year traceability of interventions and software versions applies to control systems in general (1.2.1).
Limits, records, software versions and procedure in one file. Trial file free until 20.01.2027.
Sources. Regulation (EU) 2023/1230, OJ L 165, 29.6.2023, including the corrigendum of 1.4.2025; Directive 2006/42/EC for the previous regime. On the Digital Omnibus on AI: summaries by specialist law firms (October 2026); only the Official Journal text is authentic.
The information has been checked against the official text, paraphrased and shortened; only the official text is authentic.
normakte is a drafting and calculation aid; it replaces neither conformity assessment, nor a notified body, nor a technical assessment by a specialist, nor legal advice.