New Product Liability Directive: software and AI from December 2026.

1. What changes, and from when

The old Product Liability Directive dates from 1985 and knows neither software nor updates nor machines that keep learning after delivery. Its successor, Directive (EU) 2024/2853, was adopted on 23 October 2024, published in the Official Journal on 18 November 2024 and entered into force on 8 December 2024. Member states must transpose it by 9 December 2026. Germany is nearly there: the Bundestag passed the Act on the Modernisation of Product Liability Law on 8 October 2026, the Bundesrat declined to call the mediation committee on 16 October 2026, and promulgation in the Federal Law Gazette was still outstanding on 23 October. The new Product Liability Act takes effect on 9 December 2026 and replaces the 1989 statute.

What matters is the cut-off date, not the date of the damage: the new rules apply to products placed on the market or put into service from 9 December 2026. Everything delivered before stays under the old law - as long as the machine is not substantially modified later, because then it counts as a new product (section 3). A machine that leaves the factory in November 2026 is liable under the old law; the same machine in December under the new one.

2. Software, AI and related services are now products

The central sentence of the directive is Article 4(1): “product” includes “electricity, digital manufacturing files, raw materials and software”. Recital 13 spells out what is meant: “operating systems, firmware, computer programs, applications or AI systems”, whether stored on a device, accessed through the cloud or supplied as software-as-a-service. Providers of AI systems within the meaning of the AI Act are to be regarded as manufacturers. The German Act carries this into section 2: software is a product, with a single exception.

That exception is narrower than it sounds. Free and open-source software is excluded only where it is “developed or supplied outside the course of a commercial activity” (Art. 2(2)). Whoever integrates an open-source library into the machine controller and sells the machine is liable for that library like for any other component - the volunteer developer is not, the machinery manufacturer is (recital 15).

Three further definitions matter for mechanical and plant engineering. A related service (Art. 4(3)) is a digital service without which the product could not perform one of its functions - the cloud connection of a plant whose process parameters come from a server qualifies. Digital manufacturing files (Art. 4(2)) are files that enable the automated control of machinery or tools; recital 16 names files for drilling, turning and milling machines and 3D printers. And whoever develops or manufactures a product for their own use also counts as a manufacturer (Art. 4(10)). The special-purpose machine built in-house for your own production is covered.

3. What “defective” will mean: learning, connecting, not updating

A product is defective where it does not provide the safety that a person is entitled to expect or that is required under Union or national law (Art. 7(1)). What is new is the list of circumstances a court must take into account. Three of them aim directly at connected and learning machines:

1. Learning after delivery. Courts must consider “the effect on the product of any ability to continue to learn or acquire new features after it is placed on the market or put into service” (Art. 7(2)(c)). Recital 32 is blunt: a manufacturer that develops a product with the ability to develop unexpected behaviour remains liable for behaviour that causes damage.
2. Cybersecurity. The yardstick also includes safety-relevant cybersecurity requirements (Art. 7(2)(f)). An exploitable vulnerability can make a product defective. Which requirements those are for connected machinery is laid down by the Cyber Resilience Act - product liability is the civil-law lever behind it.
3. Control after placing on the market. Until now the state at delivery counted. Now, as long as the manufacturer retains control, the relevant moment is when the product left the manufacturer’s control (Art. 7(2)(e)). Control lies with whoever can supply updates or upgrades, or have a third party supply them (Art. 4(5)). Whoever offers remote maintenance has control.

From the third point follows the rule that many summaries call “liability for failing to update”. Strictly, it is a defence that has been removed: a manufacturer can normally escape liability if the defect probably came into being only after the product was placed on the market. That no longer works where the defect - under the manufacturer’s control - is due to a related service, to software including updates, or to “the absence of software updates or upgrades necessary to maintain safety” (Art. 11(2)). The directive explicitly creates no obligation to supply updates (recital 51); a manufacturer that supplies an update the customer does not install is exempt. The Cyber Resilience Act requires a support period of at least five years for connected products anyway - the two acts interlock here.

The second consequence concerns the installed base. Whoever substantially modifies a product outside the manufacturer’s control and then makes it available or puts it into service counts as its manufacturer (Art. 8(2)). A substantial modification can also result from a software upgrade or from the continuous learning of an AI system (recital 40) - and it restarts the ten-year period (Art. 17(1)). A machine from 2019 that receives a new control system with a different intended purpose in 2027 is, for liability purposes, a product of 2027. What counts as a substantial modification is determined by product safety law, for machinery from 20 January 2027 the Machinery Regulation.

4. Burden of proof: disclosure and presumptions

Formally nothing changes: the claimant must prove “the defectiveness of the product, the damage suffered and the causal link” between the two (Art. 10(1)). In practice two instruments change the position fundamentally.

Disclosure of evidence (Art. 9): where the claimant presents facts and evidence sufficient to support the plausibility of the claim, the manufacturer must disclose the relevant evidence at its disposal - necessary, proportionate, with protection for trade secrets, and on the court’s order in an easily accessible and easily understandable form. Recital 42 adds that this includes documents the defendant has to create by compiling evidence it already holds. Where a defendant fails to disclose, defectiveness is presumed (Art. 10(2)(a)).

Presumptions (Art. 10): defectiveness is also presumed where the product breaches mandatory product-safety requirements or shows an obvious malfunction during reasonably foreseeable use; the causal link is presumed where the damage is of a kind typically consistent with the defect. The rule that matters for AI is paragraph 4: where it is “excessively difficult” for the claimant to prove defect or causation, “in particular due to technical or scientific complexity”, it is enough to show that either is “likely”. Recital 48 names machine learning as an example of such complexity, and the case where the claimant would have to explain the inner workings of an AI system. All presumptions are rebuttable (Art. 10(5)) - but the manufacturer has to rebut them, with the records it has.

5. The liability limits go

The German Product Liability Act of 1989 has a cap for personal injury, a deductible for property damage and an expiry of claims after ten years. The directive deletes the first two without replacement, extends the third for latent personal injury and forbids member states to set new caps (recital 56):

Then there is the liability chain: alongside the manufacturer, the manufacturer of a defective component is liable, for non-EU manufacturers the importer and the authorised representative, failing those the fulfilment service provider, and a distributor that does not name an EU-based liable party within one month of being asked (Art. 8). All are jointly and severally liable (Art. 12); recourse is possible, and a manufacturer that integrates software from a micro or small enterprise may waive it contractually - never towards the injured person.

6. What is really new for machinery manufacturers - and what is not

Anyone who reads the directive as a liability avalanche for B2B business overlooks Article 6. Compensable are death and personal injury including “medically recognised damage to psychological health”, property damage and the destruction of data not used professionally. Excluded remain the defective product itself, property “used exclusively for professional purposes”, professionally used data and pure economic loss (recital 24). Only natural persons can claim. The customer’s production downtime, the damaged neighbouring line, the lost process data: all of that stays where it is today - in the contract and in section 823 of the Civil Code.

The new risk lies with the people at the machine. The operator, the maintenance fitter, your own service technician are natural persons; their injury by a machine whose control software was defective, whose safety update never came or whose learning controller produced an unexpected state falls fully under the new law - without a cap, with a duty to disclose and with the presumptions from section 4. Add products used by private individuals or for mixed purposes, such as tools and small machines in the trades.

And a point that is often misrepresented: there is no separate AI liability. The Commission formally withdrew its proposal for an AI Liability Directive in October 2025. The Product Liability Directive mentions AI systems only in the recitals; the articles say “software” and “ability to continue to learn”. Which obligations the AI Act imposes on machinery with AI is a different question - and since the Digital Omnibus a deferred one: the AI requirements for machinery apply only from 2 August 2028. Product liability does not wait; from 9 December 2026 it says who pays when something happens. The German Chambers of Commerce and Industry made exactly this point in their statement on the bill: updates and self-learning functions lead to a permanent responsibility of manufacturers over the entire product life cycle.

7. Six things a machinery manufacturer should settle before 9 December 2026

1. Product inventory: which software, firmware, cloud services and AI functions are in the machines that ship from December - including purchased and open-source components? You are liable to the injured person for every one of them.
2. Control and updates: for which products can you supply updates, and for how long? Whoever offers remote maintenance has control and loses the “defect arose later” defence. Define the support period and document every safety update you supply.
3. Evidential readiness: could you show within a few weeks, in an easily understandable form, which software version, which training state and which parameters a particular machine had on a particular day? That is disclosure under Article 9 - operating logs, version records and release documents are your defence.
4. Learning systems: which functions change their behaviour in operation? Bound the learning space, log model states and define from when a change is a substantial modification - it turns the machine into a new product.
5. Supply chain: do your contracts with software and component suppliers contain recourse provisions, documentation duties and update commitments? Waiving recourse against small enterprises is permitted, but it is a deliberate decision.
6. Insurance: check with your insurer whether your product liability cover reflects software, related services and the end of the cap - before the first machine under the new law leaves the factory.

8. When this is not urgent for you

Three situations in which the new directive changes little:

1. No software, no connectivity, no learning. A purely mechanical or hard-wired machine without remote access is a product under the new law just as under the old. The limits go for it too, but the new defect criteria do not bite.
2. Exclusively stationary B2B plant without risk to people. Where a defect can only lead to production downtime and damage to professionally used plant, the dispute stays in contract law. The human-machine interface decides.
3. Installed base left untouched. The installed base stays under the old law as long as nobody substantially modifies it. A retrofit with a new controller or a new intended purpose is exactly such an intervention - then section 3 is worth a look.

9. Frequently asked questions

10. Sources

The German parliamentary documents are the actual evidence for the transposition and are linked as such.

• Directive (EU) 2024/2853 on liability for defective products, EUR-Lex (OJ L of 18 November 2024): Art. 2, 4, 6-12, 15-17, 21-23; recitals 13-16, 24, 32, 40, 42, 48, 51, 56.
• Corrigendum to Directive (EU) 2024/2853, EUR-Lex (OJ L of 7 May 2026): Art. 2(1), “after 8 December 2026”.
• Directive 85/374/EEC, consolidated version: deductible (Art. 9) and optional liability cap (Art. 16) of the old regime.
• Produkthaftungsgesetz of 15 December 1989, gesetze-im-internet.de: sections 1, 3, 10, 11, 12, 13, 14. In German.
• Gesetz zur Modernisierung des Produkthaftungsrechts, text adopted by the Bundestag, BR-Drs. 584/26 (9 October 2026) and decision of the Bundesrat (16 October 2026): sections 2, 4, 5, 7, 9, 16-20, 23. In German.
• Government bill, BT-Drs. 21/4297 (25 February 2026): the Federal Government’s reply on the cap and the deductible; committee recommendation BT-Drs. 21/8429 (7 October 2026). In German.
• Federal Ministry of Justice and Consumer Protection, press release on the cabinet decision (17 December 2025): principle of one-to-one transposition. In German.
• German Chambers of Commerce and Industry (DIHK), statement on the draft bill (8 October 2025, in the ministry’s archive of statements): permanent responsibility over the product life cycle; quoted in our translation. In German.
• European Commission, Withdrawal of Commission proposals (OJ C of 6 October 2025): withdrawal of the proposal for an AI Liability Directive, COM(2022) 496.