Civil Law And AI-Controlled Industrial System Accident Liability In Europe

1. Introduction

AI-controlled industrial systems include AI-powered robots, automated production lines, autonomous forklifts, smart cranes, industrial control systems, predictive-maintenance systems, AI safety systems, and machine-learning software controlling industrial equipment.

An accident may occur because:

• the AI makes an incorrect decision;
• sensors provide defective information;
• the machine-learning model misclassifies a situation;
• software contains a design or coding defect;
• an update changes the system's behaviour;
• the safety-control system fails;
• the manufacturer failed to anticipate reasonably foreseeable misuse;
• the operator relies excessively on an AI recommendation;
• maintenance or cybersecurity failures interfere with the system.

European law does not yet have one single “AI industrial accident liability” regime. Instead, liability is constructed from EU product-safety/product-liability legislation together with national civil, tort, contract and occupational-safety law.

The new EU framework is particularly important because the Machinery Regulation expressly addresses increasingly autonomous machinery, including systems with learning, adaptability and real-time information processing.

2. Meaning of AI-Controlled Industrial System Accident Liability

The concept concerns civil liability where an AI-controlled or AI-assisted industrial system causes death, personal injury, property damage or other legally compensable loss.

Example

Suppose an AI-controlled robotic arm in a factory is programmed to identify workers and stop when a person enters its operating zone.

The AI incorrectly classifies a worker as an object that can be ignored.

The robotic arm moves and seriously injures the worker.

Potential questions include:

1. Was the AI system defective?
2. Was the physical machine defective?
3. Were the sensors defective?
4. Was the AI model improperly trained?
5. Was the safety system inadequate?
6. Was the manufacturer negligent?
7. Was the operator negligent?
8. Was the software updated properly?
9. Did the employer fail to maintain the machine?
10. Who should bear the loss?

This produces a multi-layer liability problem rather than a simple manufacturer-versus-worker dispute.

3. European Legal Framework

A. EU Machinery Regulation

The central regulatory instrument is Regulation (EU) 2023/1230 on machinery.

It establishes health and safety requirements for machinery and related products. Importantly, its definition of machinery can include systems requiring only the uploading of application-specific software, while a safety component can itself be software.

The Regulation specifically recognises modern machinery involving:

• real-time information processing;
• problem solving;
• sensors;
• learning;
• adaptability;
• autonomy;
• operation in less structured environments.

Manufacturers must conduct risk assessment and design machinery to eliminate or reduce risks, including reasonably foreseeable misuse.

For AI-controlled machinery, particularly important requirements concern:

• safety of control systems;
• software intervention;
• software modifications;
• self-evolving behaviour;
• autonomous decision-making;
• recording of safety-related decision-making;
• ability to correct the machinery.

The current consolidated text requires particular controls for machinery with fully or partially self-evolving behaviour or logic operating with varying levels of autonomy.

4. AI Act

The EU AI Act, Regulation (EU) 2024/1689, is also relevant where AI forms part of an industrial system.

It is principally a regulatory and safety framework, rather than a comprehensive civil-compensation statute.

For AI incorporated into machinery, the interaction between the AI Act and machinery legislation becomes important, particularly where an AI system constitutes a safety component.

Therefore:

AI Act compliance does not automatically eliminate civil liability.

A system can satisfy regulatory requirements and still potentially create liability under applicable civil/product-liability rules if the relevant elements of a claim are established.

5. New EU Product Liability Directive

A major development is Directive (EU) 2024/2853 on liability for defective products.

The new regime expressly treats software, including AI systems, as products for no-fault product liability purposes. This applies whether software is integrated into hardware, supplied through networks or cloud systems, or provided through a software-as-a-service model.

This is extremely important for industrial AI.

For example:

AI software → controls robot → robot malfunctions → worker injured

The software component cannot simply be dismissed as something legally separate from the physical machine.

The Directive also addresses digital manufacturing files, including files capable of automatically controlling machinery such as drills, milling machines and 3D printers.

The new Directive therefore substantially strengthens the legal basis for claims involving AI-controlled industrial equipment.

6. Main Types of AI Industrial Accidents

1. Robotic-arm accidents

An AI-controlled robot may incorrectly identify a worker or fail to activate a safety stop.

2. Autonomous vehicle accidents

Industrial autonomous vehicles, forklifts and warehouse robots may collide with workers or equipment.

3. AI-controlled manufacturing

An incorrect AI instruction may cause:

• excessive pressure;
• excessive temperature;
• incorrect cutting;
• incorrect assembly;
• mechanical failure.

4. Predictive-maintenance failure

AI may incorrectly determine that machinery is safe when a component is actually approaching failure.

5. Sensor failure

AI may receive defective information because cameras, LiDAR, radar, temperature sensors or pressure sensors malfunction.

6. Cybersecurity incidents

A cyberattack could alter AI-controlled industrial behaviour.

7. Software-update accidents

An update may introduce a new defect or alter the behaviour of a previously safe machine.

8. Learning-system accidents

A self-learning system may change its behaviour after deployment, creating a difficult question about whether the relevant defect existed when the product was first placed on the market.

7. Who Can Be Liable?

Potential defendants can include:

The difficult question is often causal allocation between these actors.

8. Product Defect

Three broad categories are especially important.

A. Design defect

The AI-controlled system was inherently unsafe.

Example:

The system was designed to continue operation when a worker entered an area that should have triggered an emergency stop.

B. Manufacturing defect

The particular machine or software installation differs from the intended safe design.

C. Information/instruction defect

The manufacturer failed to provide adequate warnings, instructions, limitations or safety information.

For AI systems, additional questions arise:

• Was the training data appropriate?
• Were dangerous edge cases tested?
• Was the AI sufficiently robust?
• Were sensor failures considered?
• Were foreseeable human mistakes considered?
• Was the model validated after deployment?

9. Causation

Causation is likely to become one of the most difficult issues.

Consider:

AI error + defective sensor + operator mistake + inadequate maintenance → accident

The court may need to determine:

1. Which event actually caused the accident?
2. Were several causes concurrent?
3. Was the AI defect substantial?
4. Did human intervention break the causal chain?
5. Was the accident foreseeable?
6. Did maintenance failure contribute?
7. Did a software update create the relevant defect?

The claimant will generally need to establish the legally required causal connection under the applicable regime.

10. Important Case Laws

Because AI-controlled industrial accidents are a relatively new phenomenon, there are few reported European judgments directly concerning an AI-controlled factory system. The following cases therefore include both direct machinery/product-liability authorities and important CJEU analogical authorities.

Case 1 — Linddana Machine Accident

Cour de cassation, 1re chambre civile, 12 September 2018, No. 17-21.594 — Direct machinery authority

This is one of the most useful European cases for the subject.

A purchaser was seriously injured while using a wood-chipping machine manufactured by Linddana.

The machine had an emergency-stop mechanism, but the accident raised questions concerning whether the operator could actually activate that safety mechanism while being drawn into the machine.

The case also concerned an expert report and the application of French defective-product liability.

The underlying court had considered evidence that the machine's safety arrangement presented serious risks, notwithstanding its conformity with a European standard.

The Cour de cassation ultimately focused on procedural fairness and the need for proper adversarial examination of the evidence.

Importance for AI systems

This case illustrates an important principle:

Regulatory or technical-standard conformity does not necessarily end the civil-liability inquiry.

For AI machinery, the equivalent question could be:

“The AI system complied with the applicable technical standard — does that automatically mean that it was legally safe?”

The answer cannot automatically be assumed to be yes.

The case is particularly useful for AI-controlled machinery because it demonstrates the importance of:

• safety design;
• emergency systems;
• expert evidence;
• conformity standards;
• actual operation of safety mechanisms;
• causation.

Case 2 — Boston Scientific Medizintechnik

Joined Cases C-503/13 and C-504/13, Boston Scientific Medizintechnik GmbH v AOK Sachsen-Anhalt and Betriebskrankenkasse RWE, CJEU, 5 March 2015

Direct product-liability authority; analogically relevant to AI industrial products.

The case concerned pacemakers and implantable cardioverter-defibrillators.

The CJEU held that where products belonging to the same production series or group present a potential defect, an individual product may be classified as defective without proving that the particular product had already physically manifested the defect.

The Court also recognised certain costs associated with replacing the defective product as compensable personal-injury damage.

Relevance to AI industrial systems

AI systems frequently contain common:

• software versions;
• algorithms;
• model architectures;
• firmware;
• safety configurations.

Suppose 100 industrial robots use the same AI safety model and testing reveals a systematic safety defect.

Boston Scientific provides an important analogy:

A systemic defect affecting a product group may have legal significance even where every individual machine has not yet suffered an accident.

This becomes particularly important for recall and preventive safety measures.

Case 3 — Henning Veedfald v Århus Amtskommune

Case C-203/99, CJEU, 10 May 2001

Direct EU product-liability authority; analogical industrial relevance.

Mr Veedfald was harmed by a defective product used in connection with medical treatment.

The CJEU considered when the defective-product regime applies and addressed the statutory exemptions from producer liability.

The Court's judgment is an important authority concerning the scope of EU strict product liability.

Relevance to AI industrial systems

The case demonstrates that the product-liability regime can apply in circumstances where a product is used as part of a broader service.

This is significant for:

AI + machine + industrial service

For example, an AI-controlled production system may be operated as part of a manufacturing service rather than simply being sold as an isolated machine.

The court must therefore examine:

• who produced the relevant product;
• whether it was supplied;
• how it was used;
• whether the statutory conditions for product liability are satisfied.

Case 4 — Moteurs Leroy Somer v Dalkia France

Case C-285/08, CJEU, 4 June 2009

Highly relevant industrial-equipment authority.

An alternator manufactured by Moteurs Leroy Somer overheated and caused a fire in a generator system installed at a French hospital.

The case concerned the scope of EU product-liability law concerning damage to property intended for professional use.

Legal significance

The CJEU held that the Product Liability Directive did not harmonise liability for damage to property intended for professional use and used for professional purposes in the circumstances considered.

Consequently, national law may remain relevant for such losses.

Importance for AI industrial accidents

This case is particularly valuable because it concerns:

defective industrial equipment → malfunction → fire → damage to another industrial system

That is extremely similar to potential AI-industrial accidents.

For example:

AI-controlled furnace malfunction → overheating → factory fire → destruction of neighbouring machinery

Different parts of the loss may fall under different legal regimes.

This means the claimant must carefully distinguish:

• personal injury;
• damage to other property;
• damage to the defective machine itself;
• business interruption;
• consequential economic loss.

Case 5 — W and Others v Sanofi Pasteur MSD

Case C-621/15, CJEU, 21 June 2017

Analogical product-liability authority concerning causation and scientific uncertainty.

The case involved alleged damage from a vaccine and the evidentiary question of proving defect and causation where there was no established scientific consensus.

The CJEU considered whether serious, specific and consistent evidence could, subject to national procedural requirements, support proof of defect and causal connection.

Relevance to AI

AI systems create similar evidentiary problems.

Suppose an AI-controlled machine causes an accident, but:

• the algorithm has changed;
• the model has been retrained;
• logs are incomplete;
• the precise decision cannot easily be reconstructed;
• the developer claims that the accident was statistically unforeseeable.

The Sanofi reasoning illustrates the importance of evidence and causation under scientific/technical uncertainty.

However, the case should not be read as creating a general rule that statistical evidence automatically proves an AI defect.

Case 6 — Commission v United Kingdom

Case C-300/95, Commission of the European Communities v United Kingdom, CJEU, 29 May 1997

Direct product-liability authority concerning the state-of-scientific-knowledge defence.

The case examined the defence under Article 7(e) of the former Product Liability Directive concerning whether the state of scientific and technical knowledge at the relevant time allowed the producer to discover the defect.

The CJEU recognised that the producer may rely on the development-risk defence where the statutory conditions are satisfied.

AI relevance

This is particularly important for cutting-edge AI.

Imagine that:

• an industrial AI system is deployed in 2027;
• a previously unknown failure mode emerges in 2029;
• no reasonable manufacturer could have identified the failure mode when the system entered the market.

The legal question becomes whether the applicable product-liability regime permits a state-of-scientific-and-technical-knowledge defence.

The new Product Liability Directive retains carefully structured defences, so the exact temporal and statutory requirements must be examined.

Case 7 — Skov Æg v Bilka

Case C-402/03, Skov Æg v Bilka Lavprisvarehus, CJEU, 10 January 2006

Analogical authority concerning allocation of responsibility within the supply chain.

The CJEU considered liability of a supplier under the EU defective-product regime.

The Court held that the harmonised regime prevented Member States from imposing additional no-fault liability on suppliers beyond the circumstances specified by the Directive.

AI-industrial relevance

An AI-controlled industrial system can involve:

AI developer → software supplier → sensor supplier → machine manufacturer → integrator → distributor → employer

The case demonstrates why the court must identify the legal role of each actor rather than simply imposing statutory product liability on every participant in the supply chain.

Case 8 — P... v Mafroco

Cour de cassation, 1re chambre civile, 9 December 2020, No. 19-21.390

French product-liability authority concerning defective machinery and economic loss.

The case involved a machine that allegedly suffered repeated failures.

The claimant sought, among other things, compensation for operating losses and the absence of a replacement machine.

The Cour de cassation held that the defective-product regime did not cover damage to the defective product itself or economic losses consequential upon damage to that product.

Relevance to AI industrial systems

This distinction is extremely important.

Suppose an AI-controlled industrial machine malfunctions.

The factory suffers:

1. injury to a worker;
2. damage to another machine;
3. destruction of the AI-controlled machine itself;
4. two weeks of production interruption.

These losses may not all be treated identically under product-liability legislation.

The claimant may therefore need to rely on:

• product liability;
• contractual liability;
• negligence;
• warranty;
• insurance;
• national commercial law.

11. Direct vs Analogical Case Law

Important: There is presently no substantial body of reported CJEU or national appellate case law specifically deciding accidents caused by a fully AI-controlled industrial system. The closest authorities are machinery-safety and defective-product cases. The AI-specific analysis therefore comes principally from applying the newer EU machinery/product framework to established liability principles.

12. Manufacturer Liability

A manufacturer can potentially face liability where:

Design

The AI-controlled system was designed in an unsafe way.

Software

The software contained a defect.

Integration

The AI was improperly integrated with physical machinery.

Safety system

The emergency-stop or protective system was inadequate.

Instructions

The manufacturer failed to provide adequate warnings.

Updates

A software update introduced or failed to correct a safety problem.

The Machinery Regulation is particularly significant because its safety architecture expressly addresses software intervention, safety-related decision-making and autonomous/self-evolving machinery.

13. AI Developer Liability

The AI developer could potentially face liability where the relevant legal regime treats the AI software as a product or where national law establishes an independent duty of care.

Potential allegations include:

• defective algorithm;
• inadequate training;
• insufficient testing;
• inadequate validation;
• failure to detect dangerous edge cases;
• unsafe update;
• defective safety model;
• inadequate documentation;
• failure to warn about limitations.

The new Product Liability Directive is especially significant because it expressly treats AI systems and software as products for product-liability purposes.

14. Employer/Operator Liability

The employer may also be responsible.

Examples:

• failing to train workers;
• bypassing safety mechanisms;
• ignoring warnings;
• using machinery outside its intended conditions;
• inadequate maintenance;
• failing to install software updates;
• allowing unauthorised modifications.

Thus:

AI manufacturer liability and employer liability can coexist.

The existence of an AI defect does not automatically eliminate the relevance of human conduct.

15. Human Oversight

One of the central principles is:

Automation does not automatically eliminate human responsibility.

A factory may have an AI system capable of autonomous decisions, but human operators may still have responsibilities concerning:

• supervision;
• emergency intervention;
• maintenance;
• configuration;
• inspection;
• system shutdown.

At the same time, manufacturers cannot necessarily shift all responsibility to the operator simply by inserting a warning into a manual.

The court will examine the actual design and foreseeable use of the system.

16. AI Learning and Changing Behaviour

Traditional product liability generally asks whether the product was defective in the legally relevant sense.

AI creates an additional question:

What if the system changes after deployment?

For example:

Day 1: AI correctly identifies workers.

Month 6: Model update changes classification behaviour.

Month 8: Worker is incorrectly classified.

Month 9: Accident occurs.

Potential questions:

• Was the original system defective?
• Was the update defective?
• Who authorised the update?
• Was monitoring adequate?
• Was the manufacturer responsible for the update?
• Was the change foreseeable?
• Were safety controls capable of preventing the changed behaviour?

The Machinery Regulation's provisions concerning software intervention, safety-related decision-making and self-evolving systems are particularly relevant to this problem.

17. Cybersecurity and AI Industrial Accidents

Cybersecurity can create another causal pathway.

Example:

Cyberattack → altered AI model → unsafe machine command → worker injury

Potential defendants might include:

• AI developer;
• machine manufacturer;
• cybersecurity provider;
• system integrator;
• employer.

The court would need to determine whether the cyberattack:

• was foreseeable;
• was preventable;
• constituted a third-party intervention;
• broke the chain of causation;
• resulted from inadequate security measures.

Modern machinery rules expressly require consideration of reasonably foreseeable malicious attempts by third parties where relevant to hazardous situations.

18. Evidence in AI Accident Litigation

AI industrial litigation will often require unusually technical evidence.

Important evidence can include:

Technical evidence

• source code;
• model architecture;
• training methodology;
• validation results;
• safety testing;
• system specifications.

Operational evidence

• machine logs;
• sensor data;
• emergency-stop logs;
• operator commands;
• maintenance records.

AI-specific evidence

• model version;
• model weights where legally available;
• inference logs;
• decision traces;
• confidence scores;
• retraining records;
• update history.

Physical evidence

• damaged machinery;
• safety barriers;
• sensors;
• controllers;
• circuit boards.

Regulatory evidence

• conformity assessment;
• technical documentation;
• risk assessment;
• CE-related documentation;
• safety standards;
• accident reports.

19. Black-Box Problem

A major issue is the so-called black-box problem.

Suppose an AI system says:

“Move robotic arm to position X.”

But nobody can easily reconstruct why it generated that instruction.

The injured worker may argue:

“The system caused the accident.”

The manufacturer may respond:

“We cannot reproduce the decision.”

This creates an evidentiary challenge.

The Machinery Regulation's requirements concerning recording of safety-related decision-making for certain autonomous systems are therefore particularly significant.

20. Regulatory Compliance as a Defence

A manufacturer may argue:

“The machine complied with all applicable EU safety standards.”

This can be important evidence, but it does not necessarily answer every civil-liability question.

The French Linddana litigation is instructive because conformity with a European standard did not end the dispute concerning the actual safety of the machine and its emergency system.

Similarly, regulatory compliance should generally be distinguished from:

• absence of defect;
• absence of negligence;
• absence of causation;
• absence of damage.

21. Defences

Possible defences may include:

1. No defect

The manufacturer argues that the machine provided the safety reasonably expected.

2. No causation

The accident resulted from something other than the alleged AI defect.

3. Misuse

The operator used the machine in an unforeseeable manner.

4. Modification

The machine or AI system was altered after delivery.

5. Third-party interference

A third party hacked or modified the system.

6. State of scientific knowledge

The relevant defect could not reasonably have been discovered under the applicable statutory defence.

The CJEU's Commission v United Kingdom, C-300/95 judgment is important for understanding this type of defence under the former Product Liability Directive.

7. Fault of the claimant

The claimant's own conduct may affect liability or damages under applicable national law.

22. Damage and Compensation

Potential heads of loss include:

Personal injury

• medical expenses;
• rehabilitation;
• loss of earnings;
• permanent disability;
• pain and suffering where recognised;
• future care.

Death

Claims may include:

• dependency losses;
• funeral expenses;
• loss suffered by qualifying relatives.

Property damage

• damage to machinery;
• damage to buildings;
• destruction of inventory;
• damage to neighbouring equipment.

Business losses

These are particularly complicated because product-liability rules may distinguish damage to the defective product from damage to other property and consequential economic losses.

The French Mafroco decision demonstrates the importance of this distinction.

23. Special Problem of Industrial Property

Industrial AI accidents often involve property used for professional purposes.

The Moteurs Leroy Somer case is therefore especially important.

The case concerned a defective alternator that caused a fire in professional equipment. The CJEU clarified the limits of the harmonised EU defective-product regime concerning property intended for professional use.

Accordingly, an industrial claimant may have to combine:

EU product liability + national civil liability + contract + insurance law.

24. Multi-Cause Accident

Consider:

AI software incorrectly predicts that a worker is outside the danger zone.

At the same time:

• the sensor is defective;
• the machine's safety barrier is damaged;
• the employer failed to maintain it;
• the operator ignored a warning;
• the AI software had not received an available update.

The court could potentially identify concurrent causes.

Therefore, AI accident litigation should not be reduced to:

“The AI made a mistake, therefore the AI developer is liable.”

The legal inquiry is more sophisticated:

Defect → breach/duty → causation → damage → allocation of responsibility.

25. Contractual Liability

Where the industrial system is supplied under a commercial contract, contractual liability can be extremely important.

A contract may contain obligations concerning:

• performance;
• safety;
• maintenance;
• software updates;
• service levels;
• cybersecurity;
• system availability;
• integration;
• warranties.

A factory operator may therefore sue the system provider for breach of contract even where the facts do not fit neatly into statutory product liability.

26. Tort/Delict Liability

National civil-law systems may also provide general tort/delict claims.

Typical elements include:

1. legally relevant fault or other basis of liability;
2. unlawful conduct;
3. damage;
4. causation.

The exact rules differ between European countries.

For example:

• France uses general civil-liability principles under the Civil Code;
• Germany relies heavily on provisions such as §§ 823 ff. BGB;
• Italy has general tort liability under Article 2043 of the Civil Code;
• Spain has general extra-contractual liability principles under the Civil Code.

Therefore, “European civil law” is not a single uniform tort system.

27. Interaction with Occupational Safety

An industrial accident may generate parallel proceedings involving:

• civil compensation;
• employment law;
• occupational health and safety;
• administrative sanctions;
• criminal liability;
• insurance;
• product liability.

A worker injured by an AI-controlled machine may therefore have rights under several legal regimes simultaneously.

28. Important Legal Principle

A useful conceptual distinction is:

AI error

The AI generated an incorrect decision.

AI defect

The AI system failed to provide the legally required level of safety.

Human error

An operator made an incorrect intervention.

Machine defect

The physical machinery was defective.

Integration defect

The AI and physical machinery were incorrectly combined.

Maintenance defect

The system was not properly maintained.

These concepts should not be treated as identical.

29. Future Importance of the New Product Liability Framework

The new Product Liability Directive represents a major change because software and AI systems are expressly brought within the concept of products.

The Directive states that software can be a product whether supplied through hardware, networks, cloud technologies or SaaS.

For industrial AI, this potentially changes the traditional argument:

“The machine is a product, but the algorithm is merely software.”

The new framework makes that distinction considerably less persuasive where the software itself falls within the statutory concept of a product.

30. Key Principles

Principle 1

AI-controlled machinery can generate product liability as well as ordinary civil liability.

Principle 2

Software can be legally relevant as a product under the new EU Product Liability Directive.

Principle 3

The physical machine and AI software may need to be analysed together.

Principle 4

A safety component may itself be digital/software-based under the Machinery Regulation.

Principle 5

AI autonomy does not automatically remove human responsibility.

Principle 6

Manufacturer responsibility can coexist with employer/operator responsibility.

Principle 7

Causation is usually one of the most difficult issues.

Principle 8

Technical-standard conformity is important but does not necessarily resolve every civil-liability question.

Principle 9

AI learning and software updates create new questions about when and how a defect arises.

Principle 10

AI accident litigation will depend heavily on technical and operational evidence.

31. Case-Law Revision Table

32. Simple Liability Formula

For examination purposes, remember:

AI-controlled industrial system + safety defect/fault + accident + causation + legally recognised damage = potential civil liability.

Where several actors contributed:

AI developer + machine manufacturer + integrator + operator/employer + maintenance provider → possible concurrent or separately established liability.

33. Conclusion

AI-controlled industrial system accident liability in Europe is an emerging intersection of product liability, machinery safety, tort/delict, contract and occupational-safety law.

The most important development is the combination of the EU Machinery Regulation 2023/1230 with the EU Product Liability Directive 2024/2853. The Machinery Regulation expressly addresses increasingly autonomous and self-evolving machinery, while the new Product Liability Directive expressly recognises software and AI systems within the product-liability framework.

The central legal questions will normally be:

Was the AI or machine defective? → Was the system unsafe? → Who controlled or supplied the relevant component? → Did the defect cause the accident? → What damage occurred? → Which liability regime applies?

The existing European cases—especially Linddana, Moteurs Leroy Somer, Boston Scientific, Veedfald, Sanofi Pasteur, Commission v UK, Skov Æg and Mafroco—provide the established principles from which future AI-industrial accident cases are likely to develop.