The EU Machinery Regulation, Regulation (EU) 2023/1230, applies from January 14, 2027. It is the first EU machinery law that deals directly with machine learning. Any safety component or embedded system that uses machine learning to change its own behavior while performing a safety function must be checked by an independent notified body before it can be sold, and manufacturers can no longer certify it themselves. For factories buying cobots, autonomous mobile robots or AI-based safety systems, the rule sets new requirements for design, logging and documentation about 14 months from now.

What the regulation replaces and when it applies

The regulation was adopted on June 14, 2023 and published in the Official Journal on June 29, 2023. It replaces the Machinery Directive 2006/42/EC, which has governed machine safety in the EU since 2006. Because it is a regulation rather than a directive, it applies directly in every member state, without national laws to transpose it.

Most of the text applies from January 14, 2027. A few parts started earlier: the rules on notified bodies (Articles 26 to 42) have applied since January 14, 2024, so that test bodies can be designated and ready in time. Machines placed on the EU market from the 2027 date must meet the new essential health and safety requirements in Annex III.

The lawmakers explained why a rewrite was needed. The recitals note that newer machines are less dependent on human operators, can learn new actions and can work in unstructured places such as construction sites. They cite a 2020 Commission report that found gaps in product-safety law for AI, the internet of things and robotics.

Which AI-enabled machines are covered

The regulation does not regulate “AI” in general. It focuses on one narrow but important case: software that learns and that keeps people safe. Annex I, Part A, the list of the highest-risk product categories, now includes two new entries:

  • safety components with fully or partially self-evolving behavior that use machine learning to ensure safety functions (point 5);
  • machinery with embedded systems of that kind that were not sold separately, in respect of those systems only (point 6).
An industrial robot arm polishing an acoustic guitar body inside a factory
Industrial robots fall under the Machinery Regulation; the new machine-learning categories apply only when a learning system performs a safety function. Photo: Henrysz / Wikimedia Commons, CC BY 4.0

The recitals spell out the limit. The third-party rules for safety software apply only to systems whose behavior evolves through machine learning. They do not apply to fixed software that “is incapable of learning or evolving” and only runs programmed automated functions. A conventional programmable logic controller running a fixed safety routine is therefore not caught by points 5 and 6, even if it is complex. A vision-based person-detection system that keeps adapting its model after it is installed would be.

How conformity assessment changes

For products in Annex I, Part A, the manufacturer cannot use the internal production control route (module A), which is self-certification. It must choose one of three procedures involving a notified body:

  1. EU type-examination (module B), followed by conformity to type based on internal production control (module C);
  2. full quality assurance (module H);
  3. unit verification (module G), in which each unit is checked.

There is one practical relief for integrators. If a machine embeds a safety system that was already assessed by a third party when it was sold on its own, the finished machine does not need to be re-certified by a third party simply because it contains that system. In practice this pushes certification toward component suppliers: a robot maker that buys a certified learning safety module avoids a second assessment for that part.

New design rules for learning machines

Annex III adds requirements that read like an engineering specification for safe machine learning on the factory floor:

  • Risk assessment over time. It must cover hazards that may arise from the intended evolution of the machine’s self-evolving behavior, not only its behavior on the day it ships.
  • Stay inside the box. Control systems of machines with self-evolving behavior must not make the machine act beyond its defined task and movement space.
  • Always correctable. It must be possible at all times to correct the machine to keep it inherently safe.
  • No unsafe self-modification. Changes to settings or rules made by the machine or by operators, including during a learning phase, must be prevented where they could lead to hazardous situations.
  • Decision records. Data on the safety-related decision-making of software-based safety systems must be recorded and kept for one year after collection, to be shown to a national authority on a reasoned request.
  • Version tracing. A tracing log of interventions and of safety-software versions uploaded after sale must be kept for five years.
  • Protection against corruption. Software and data critical to safety compliance must be identified and protected against accidental or intentional corruption, a requirement that brings cybersecurity into machine safety.
  • Talking to people. Where relevant, autonomous machines should communicate their planned actions to operators in an understandable way.

How it fits with the EU AI Act

The EU AI Act lists EU machinery legislation among the product laws in its Annex I. An AI system that is a safety component of such a product, and that needs third-party assessment under that product law, counts as high-risk under Article 6(1) of the AI Act. After the Digital Omnibus on AI amended the timetable, the Commission’s AI Act Service Desk lists August 2, 2028 as the date when rules for high-risk AI embedded in regulated products apply.

The result is two deadlines. Machinery safety rules for learning safety functions apply from January 14, 2027, and the AI Act’s additional obligations for the same kind of system (data governance, logging, human oversight, accuracy and robustness) follow in August 2028. Manufacturers that design their logs and risk files once, to satisfy both, will save rework.

What it means for factories and machine builders

Most AI on the factory floor today, such as the visual inspection and predictive maintenance systems described in our guide to how AI is used in manufacturing, monitors quality or machine health and does not perform a safety function. Those uses are not swept into Annex I, Part A by the machine-learning entries. The rules bite where learning software decides whether a person is in danger: speed and separation monitoring on cobots, perception for autonomous mobile robots, adaptive light curtains and similar systems.

A practical checklist for the next 14 months:

  1. List every safety function in your machines and mark which ones use models that keep learning after installation.
  2. Ask suppliers whether their learning safety components will carry third-party certification under the new regulation.
  3. Book notified-body capacity early, since all affected products need an assessment before January 2027.
  4. Build the one-year decision-data and five-year version logs into the control architecture now.
  5. Freeze learning in the field where you can: a model that is fixed after validation is simpler to certify than one that keeps adapting.

What to watch next

  • Harmonized standards and common specifications that show how to meet the machine-learning requirements in practice.
  • How many notified bodies are designated for Annex I, Part A machine-learning categories before January 2027.
  • Commission guidance on where “self-evolving behavior” begins, for example for models updated by over-the-air releases.
  • Coordination guidance between the Machinery Regulation and the AI Act ahead of August 2028.

Key takeaways

  • Regulation (EU) 2023/1230 replaces the Machinery Directive and applies from January 14, 2027.
  • Machine-learning systems that perform safety functions and change their own behavior need third-party conformity assessment.
  • Fixed, non-learning safety software is not covered by the new machine-learning categories.
  • New requirements include staying within a defined task space, one-year decision records, five-year version logs and protection against corruption.
  • AI Act obligations for AI embedded in regulated products follow on August 2, 2028.

Frequently asked questions

When does the EU Machinery Regulation apply?

Regulation (EU) 2023/1230 applies from January 14, 2027, replacing the Machinery Directive 2006/42/EC. The rules on notified bodies have applied since January 14, 2024.

Does the Machinery Regulation cover all AI in machines?

No. The strictest new categories cover safety components and embedded systems whose behavior evolves through machine learning while they ensure safety functions. Fixed software that does not learn is not in those categories.

Can manufacturers self-certify machine-learning safety components?

No. Products in Annex I, Part A must go through EU type-examination, full quality assurance or unit verification with a notified body, rather than internal production control.