The new Machinery Regulation and machine learning: for manufacturers using AI in safety functions

Short answer: The EU's new Machinery Regulation (Regulation 2023/1230) and Türkiye's Machinery Safety Regulation (2023/1230/AB), prepared on its basis, both apply from 20 January 2027. Safety components that use machine learning to ensure safety functions and have fully or partially self-evolving behaviour are now in the category that requires third-party conformity assessment. A model on the line that only performs quality inspection, and no safety function, does not fall into that category.

Türkiye's regulation was published in the Official Gazette on 16 October 2026 and repeals the Machinery Safety Regulation (2006/42/AT) in force since 2009. So the topic concerns not only manufacturers exporting to the EU, but anyone placing machinery on the market in Türkiye. This post focuses on the engineering side of the regulation for machinery and automation manufacturers that use AI. It is not legal advice; assess your own product with your notified body and legal counsel.

What changes, and when?

	Old regime	New regime
EU text	Directive 2006/42/EC	Regulation (EU) 2023/1230, directly applicable
Turkish text	Machinery Safety Regulation (2006/42/AT), Official Gazette 3 March 2009	Machinery Safety Regulation (2023/1230/AB), Official Gazette 16 October 2026, No. 33402
Application	Until 20 January 2027	From 20 January 2027

Under the transitional provision of the Turkish regulation, products placed on the market under the old rules before 20 January 2027 may remain on the market, and old EC type-examination certificates stay valid until they expire. The EU text as first published gave the application date as 14 January 2027; the date of 20 January 2027 came with a later corrigendum.

Which components are in scope?

Part A of Annex I of the regulation lists the products subject to the strictest conformity assessment. Two items are critical for AI:

• Item 5: Safety components with fully or partially self-evolving behaviour using machine learning approaches ensuring safety functions.
• Item 6: Machinery that has such systems embedded (in respect of those systems only).

The recitals of the regulation exclude software incapable of learning or evolving from this scope. "Self-evolving behaviour" itself is not separately defined in the text. A model that keeps learning in the field is clearly in scope. How a model that has finished training, is frozen and does not change in the field should be assessed is a matter of interpretation. As of November 2026 the Commission's application guide has not yet been published in final form.

The real distinction is the safety function. If a vision model on the line only sorts out defective products, it is making a decision about product quality; it is not a safety component. If the same camera detects an operator entering a danger zone and stops the machine, that is a safety function. This distinction should be made from the start by writing down which component does which job; we covered the engineering questions of quality inspection in our post on machine vision quality control.

Conformity assessment: your own declaration is not enough

For products in Part A of Annex I, the regulation allows only the following routes (Art. 25(2) of the EU text, Art. 26 of the Turkish regulation):

• EU type-examination followed by conformity to type (Module B + C)
• Full quality assurance (Module H)
• Unit verification (Module G)

Internal production control (Module A), the route based on the manufacturer's own declaration, is not available in this category. A notified body is involved in all three routes. That means the time needed to work with a notified body has to be added to the project schedule.

Essential requirements for AI

The essential health and safety requirements in Annex III of the regulation address self-evolving behaviour directly for the first time. From an engineering perspective, the main points are:

Defined task and movement space (1.2.1). Control systems with self-evolving behaviour must not cause the machine to perform actions beyond its defined task and movement space. This calls for a safety layer, independent of the model, that limits the model's output.

Recording decision data (1.2.1). In software-based systems ensuring safety functions, data on the safety-related decision-making process must be recorded and kept for one year after collection. The record is used to demonstrate conformity following a reasoned request from a competent authority.

Correctable at all times (1.2.1). It must be possible at all times to correct the machine in order to maintain its inherent safety.

Software version and intervention log (1.2.1(f)). Versions of safety software uploaded after placing on the market, and interventions, must be traceable; this log must remain available for five years after the upload.

Protection against corruption (1.1.9). Software and data critical for compliance must be identified and protected against accidental or intentional corruption. The machine must be able to identify the software it needs to operate safely and to collect evidence of legitimate or illegitimate interventions in its software or configuration. This item applies to all software, not only machine learning.

Risk assessment. The risk assessment must also cover hazards that may arise from the evolution foreseeable at the time of placing on the market.

Is a model update a substantial modification?

The regulation defines a "substantial modification" as a change made by physical or digital means after placing on the market, not foreseen or planned by the manufacturer, that affects safety by creating a new hazard or increasing an existing risk. Whoever makes a substantial modification is considered to be the manufacturer, with the corresponding obligations.

Two practical consequences follow. Not every model update is automatically a substantial modification. But for an update to count as "foreseen and planned", the update process must be defined in the technical file from the start: with which data, within which limits, and passing which tests.

Preparation on the engineering side

Most requirements overlap with what a well-built AI project should produce anyway. For a safety-related model, we set up the following from the start:

• A safety boundary separate from the model: A deterministic, separately verifiable layer that prevents the machine from leaving its defined space whatever the model decides.
• A versioned and frozen model: The identity, training data and test results of every model in the field are traceable. If there is learning in the field, its limits are written down.
• Decision records: The input, output and model version of every safety-related decision are recorded in line with the retention period.
• A fixed test set and acceptance criterion: Every new version is measured on the same test set; the acceptance criterion is known before the update.
• Change management: How updates are made is defined in the technical file; a change outside that definition requires a new assessment.

We describe the technical side of this infrastructure on the MLOps and AI Security pages.

Standards and the relationship with the AI Act

As of November 2026 no harmonised standards have been published under the new regulation; the Commission's list still belongs to the old directive. The most concrete source for machine learning and functional safety is the technical report ISO/IEC TR 5469, published in 2024; but it is guidance and does not give a presumption of conformity.

On the EU AI Act side, the Omnibus of July 2026 decided that requirements for AI in machinery will come through provisions to be added to the Machinery Regulation; these are expected to apply by 2 August 2028. We compiled the full timeline in our post on the EU AI Act.

Frequently asked questions

We only sell in Türkiye. Does this regulation concern us? Yes. Türkiye's Machinery Safety Regulation (2023/1230/AB) also applies to machinery placed on the market in Türkiye from 20 January 2027.

Is the visual quality inspection system on our line within Annex I? Generally not, if it only makes decisions about product quality and performs no safety function. If the same system takes on a safety function, the assessment changes.

Does a model that has finished training and does not change in the field count as "self-evolving"? The text does not define the term. A model that keeps learning in the field is clearly in scope; the status of a frozen model is a matter of interpretation and should be discussed with the notified body early.

Does every model update require recertification? Not necessarily. Changes foreseen and planned by the manufacturer fall outside the definition of a substantial modification; for that, the update process must be defined in the technical file from the start.

Sources

• Regulation (EU) 2023/1230, the Machinery Regulation and the corrigendum setting the application date to 20 January 2027
• Machinery Safety Regulation (2023/1230/AB), Official Gazette 16 October 2026, No. 33402 (Turkish)
• Regulation (EU) 2026/1744, the Digital Omnibus on AI
• ISO/IEC TR 5469:2024, Artificial intelligence and functional safety