The EU Machinery Regulation and machine learning: what applies from 20 January 2027

1The dates, corrected

The Regulation was published in OJ L 165 on 29 June 2023 and entered into force on 19 July 2023. Its general date of application is 20 January 2027 (Article 54). The originally published text gave every date as the 13th or 14th of the month; a corrigendum [3] moved them to the 19th or 20th, and corrected the date from which the penalties article applies, misprinted as 14 October 2023, to 20 October 2026. Secondary sources still quoting 14 January 2027 are working from the uncorrected text.

There is no dual regime. Products placed on the market in conformity with Directive 2006/42/EC before 20 January 2027 may continue to be made available (Article 52(1)), and EC type-examination certificates issued under the Directive remain valid until they expire (Article 52(2)); everything placed on the market from that date must conform to the Regulation [4].

2Where machine learning enters the text

Annex I lists the categories that need one of the more demanding conformity-assessment procedures. A category goes into Part A when it “presents a serious inherent potential risk” and, among other criteria, there is a lack of harmonised standards or “a degree of uncertainty in the existing risk assessment methods” (Article 6(7)). Part A has six entries. The last two are new:

Annex I, Part A, point 5

Safety components with fully or partially self-evolving behaviour using machine learning approaches ensuring safety functions.

Annex I, Part A, point 6

Machinery that has embedded systems with fully or partially self-evolving behaviour using machine learning approaches ensuring safety functions that have not been placed independently on the market, in respect only of those systems.

Two details matter for scope. First, the trigger is a machine-learning approach ensuring a safety function: a learned model used for productivity, path optimisation or quality inspection is not, on this wording, in Part A, while a learned person detector that commands a stop is. Second, neither “self-evolving behaviour” nor “machine learning approaches” is defined in the Regulation [5]. Whether a model that is trained offline and frozen at release counts as “self-evolving” is therefore a question of interpretation; the recitals speak of systems whose behaviour evolves, and point 6 explicitly covers partially self-evolving behaviour. Teams should expect notified bodies to take a conservative view until guidance or case practice settles it. The essential requirements in Annex III use a related phrase, “fully or partially self-evolving behaviour or logic”, and do not use the words “machine learning” at all.

Part B lists categories such as protective devices designed to detect persons and logic units ensuring safety functions; a learned person-detection component may sit in both lists in practice, which is one more reason to plan for Part A treatment.

3Conformity assessment: no self-certification for learned safety functions

Under the Directive, many high-risk machines could be self-certified if the manufacturer applied the full set of harmonised standards. Under Article 25(2) of the Regulation, Part A categories must use one of three procedures, all involving a notified body: EU type-examination (module B, Annex VII) followed by conformity to type based on internal production control (module C, Annex VIII); full quality assurance (module H, Annex IX); or unit verification (module G, Annex X). Internal production control alone (module A) is not on the list, whatever standards are applied. Part B keeps module A, but only where harmonised standards or common specifications specific to the category cover all relevant essential requirements (Article 25(3)); everything else may use module A (Article 25(4)).

An EU type-examination certificate is valid for at most five years (Annex VII, point 6.1). For a product whose perception model is retrained every few weeks, the certificate is not the end of the story: what is examined is a type, and the question becomes which later changes are still covered by it (§5).

4The new essential requirements

4.1Risk assessment over the product's evolution

The general principles of Annex III require the risk assessment to cover hazards “that might arise during the lifecycle of the machinery or related product that are foreseeable at the time of placing on the market as an intended evolution of its fully or partially self-evolving behaviour or logic, or as a result of the machinery or related product designed to operate with varying levels of autonomy.” The risk assessment is therefore expected to anticipate how a learning system will change, not only how it behaves on the day it ships.

4.2Safety and reliability of control systems (1.2.1)

The control-system requirement is where most of the new obligations sit. Among them:

• Limits that learning cannot cross. “The limits of the safety functions are to be established as part of the risk assessment performed by the manufacturer and no modifications are allowed to the settings or rules generated by the machinery or related product or by operators, including during the machinery or related product learning phase, where such modifications could lead to hazardous situations.”
• A tracing log of software versions. The tracing log of the data generated in relation to an intervention, and of the versions of safety software uploaded after the product was placed on the market or put into service, must be enabled for five years after each upload, exclusively to demonstrate conformity following a reasoned request from a competent national authority.
• Bounded action. Control systems of machinery with self-evolving behaviour or logic designed to operate with varying levels of autonomy must not cause it “to perform actions beyond its defined task and movement space”.
• Recorded decisions. For such systems, recording of “data on the safety related decision-making process for software based safety systems ensuring safety function” must be enabled after placing on the market, and the data retained for one year after collection, again only to demonstrate conformity on a reasoned request.
• Correctability. “It shall be possible at all times to correct the machinery or related product in order to maintain its inherent safety.”
• No uncontrolled drift. Particular attention must be given to parameters that “shall not change in an uncontrolled way, where such change could lead to hazardous situations”, and to preventing hazardous modifications of settings or rules, including during a learning phase.

4.3Protection against corruption (1.1.9)

Software and data critical for compliance must be identified as such and protected against accidental or intentional corruption; the machine must identify the safety-relevant software installed on it and provide that information at all times in an easily accessible form; and it must “collect evidence of a legitimate or illegitimate intervention in the software or a modification of the software installed on the machinery or related product or its configuration.” Products certified under an applicable Cybersecurity Act scheme are presumed to conform to 1.1.9 and 1.2.1 insofar as the certificate covers them (Article 20(9)).

4.4Human interaction and autonomy

The ergonomics requirement (1.1.6) now includes, where relevant, adapting machinery with self-evolving behaviour designed to operate with varying levels of autonomy so that it responds to people adequately and appropriately and communicates its planned actions to operators in a comprehensible manner. Protection against moving parts (1.3.7) must be adapted to human-machine coexistence in a shared space and to direct interaction, including the psychological stress such interaction can cause.

Section 3 introduces autonomous mobile machinery: mobile machinery with an autonomous mode “in which all the essential safety functions of the mobile machinery are ensured in its travel and working operations area without permanent interaction of an operator”. Where relevant it must have a supervisory function that lets a supervisor receive information remotely and only stop, start, or move the machine to a safe position and state, and “if the supervisory function is not active, the machinery shall not be able to operate” (3.2.4). Its control system must perform the safety functions by itself even when actions are ordered through the supervisory function (3.3). Depending on the risk assessment it must operate in an enclosed zone with peripheral protection, or be equipped with devices that detect humans, domestic animals or other obstacles in its vicinity (3.3.3), and a steering-system failure must not affect safety (3.3.5). Those detection devices are, increasingly, learned perception, which brings the system back to Annex I.

5Software updates and substantial modification

A substantial modification is a modification “by physical or digital means” after placing on the market or putting into service, “which is not foreseen or planned by the manufacturer”, which affects safety by creating a new hazard or increasing an existing risk, and which requires additional guards or protective devices that necessitate modifying the safety control system, or additional measures for stability or mechanical strength (Article 3(16)). Whoever carries out a substantial modification is treated as the manufacturer and must apply the relevant conformity-assessment procedure (Article 18).

Three consequences follow for a retraining cadence. The definition includes digital changes, so a software update is not exempt by nature. Changes the manufacturer foresaw or planned are excluded, which ties back to the requirement that the risk assessment anticipate the intended evolution of self-evolving behaviour: a manufacturer that documents in advance how its models will change, within which limits, and how each change is verified, is in a far better position than one that treats every retrain as unforeseen. And the Regulation has no equivalent of the AI Act's pre-determined change mechanism (AI Act Article 43(4)) [6]: the boundary between a planned update and a substantial modification is drawn by the manufacturer's own documentation, and tested by the notified body and the market-surveillance authority.

6Technical documentation and access

Annex IV Part A sets out the technical documentation. Two points are new and specific to software. Point (m) covers “the source code or programming logic of the safety related software”, to be provided to demonstrate conformity on a reasoned request from a competent national authority, where necessary for it to check the essential requirements. Point (n) requires, for sensor-fed, remotely driven or autonomous machinery whose safety-related operations are controlled by sensor data, “a description, where appropriate, of the general characteristics, capabilities and limitations of the system, data, development, testing and validation processes used”. Point (g) continues to require reports and results of tests, inspections and examinations.

The documentation and the EU declaration of conformity must be kept for at least ten years after placing on the market (Article 10(3)). Access to source code is a right of competent national authorities on a reasoned request; the Regulation does not give notified bodies a general right to source code, and the EU type-examination annex says nothing about it. In practice a notified body examining a learned safety component will still ask how it was trained, tested and validated, which point (n) makes part of the file.

7The AI Act, before and after the Digital Omnibus

As adopted, the AI Act [6] classifies an AI system as high-risk where it is a safety component of a product covered by the legislation in its Annex I and that product must undergo third-party conformity assessment (Article 6(1)). Annex I, Section A listed the Machinery Directive; for Section A products the AI Act's requirements were to be checked inside the sectoral conformity assessment (Article 43(3)), with obligations applying from 2 August 2027.

Regulation (EU) 2026/1744 [2], in force since 27 July 2026, changed this for machinery. It deletes machinery from Annex I, Section A and adds the Machinery Regulation as point 21 of Section B, adopting what its recitals call a sectoral approach. For high-risk AI tied to Section B legislation, only a short list of AI Act articles applies directly. Instead, the Omnibus amends Article 8 of the Machinery Regulation so that the Commission adopts delegated acts adding requirements for high-risk AI safety components to Annex III, reflecting the AI Act's requirements for high-risk systems and its provisions on quality management, logging, post-market monitoring and serious-incident reporting; those delegated acts are to apply by 2 August 2028. Until Machinery-specific standards exist, AI systems complying with harmonised standards or common specifications under the AI Act are presumed to conform to those requirements (new Article 20(10)). The Omnibus also clarifies that AI used solely for non-safety purposes such as performance optimisation or convenience is not a safety component, while AI whose failure would endanger health and safety is.

8Standards

Presumption of conformity comes from harmonised standards cited in the Official Journal under the Regulation (Article 20(1)), or from common specifications the Commission may adopt where no suitable standard is expected within a reasonable period (Article 20(3)). The Commission issued its standardisation request to CEN and CENELEC on 20 January 2025 [7]; it explicitly asks for standards covering safety functions with self-evolving behaviour, protection against corruption, human-machine interaction and autonomous mobile machinery, coherent with the AI Act and the Cyber Resilience Act. As of October 2026 no list of harmonised standards had been cited under the Regulation; the Commission states that a first list is being prepared [8]. Standards cited under the Directive, such as EN ISO 12100, EN ISO 13849-1, EN IEC 62061, EN ISO 10218 and EN ISO 3691-4 [9] [10], do not by themselves give a presumption of conformity under the Regulation. An international technical specification on functional safety and AI systems (ISO/IEC TS 22440) is in development and was not published at the time of writing.

9Enforcement

Member States set penalties that must be effective, proportionate and dissuasive and may include criminal penalties for serious infringements, applying from 20 October 2026 (Article 50). Unlike the AI Act, the Regulation sets no EU-level fine amounts. Market-surveillance authorities can act against products that present a risk under Chapter VI, which has applied to Directive-era products since 19 July 2023.

10Commonly misreported

• The dates. 20 January 2027 is the date of application; 14 January 2027 comes from the uncorrected text.
• Self-certification with standards. That route does not exist for Annex I Part A, including learned safety functions.
• “Notified bodies get the source code.” The source-code provision is for competent national authorities, on a reasoned request.
• “Machine learning” versus “self-evolving”. Annex I uses the first; Annex III uses “self-evolving behaviour or logic”. Neither is defined.
• The AI Act timeline for machinery. Many 2025 analyses give 2 August 2027 and single AI Act conformity assessment for machinery. Since July 2026, machinery is in Section B and AI requirements arrive through Machinery Regulation delegated acts applying by 2 August 2028.
• Retention. One year for decision-making data, five years for the software-version tracing log, ten years for the technical documentation.

11From requirements to artefacts

The table maps the provisions above onto evidence an engineering team can produce for every model release. It is a reading of the text, not a statement of what any notified body will accept.

This article quotes the Regulation as published in the Official Journal and corrected by the corrigendum, and Regulation (EU) 2026/1744 as published on 24 July 2026. Where it quotes, it quotes the text; where it summarises, check the provision before relying on it. It is not legal advice.