product-liabilitysoftwareliabilityai-regulationfine-tuning

Product liability for software and AI in the Netherlands

November 11, 202613 min readPixel Management

This article is also available in Dutch

The date that matters in Dutch bill 36 906 isn't the day the act commences. It's 9 December 2026.

Product liability for software is a manufacturer's no-fault liability for injury, property damage and loss of private data suffered by natural persons because of defective software. Directive (EU) 2024/2853 makes software and AI systems products, including when they're delivered as SaaS, and the Dutch bill writes that into Book 6 of the Dutch Civil Code (BW). It's not a general remedy for businesses: a corporate customer that loses revenue over a wrong chatbot answer gets nothing from this regime.

What is your role under Dutch product liability for software and AI?

Your role depends on what you do with the software, not on who trained the model. The table shows how we read the text. No case law exists under the new rules yet and we aren't a law firm, so have a lawyer check how this applies to your product.

SituationRoleExposure
Buying SaaS for internal useUser, not an economic operatorNo product liability. Contract and tort still apply
Integrating a model through an API into your own productManufacturer of the end productFull, including for the embedded model. Recourse against the supplier afterwards
Fine-tuning a model and selling it under your nameManufacturerFull
Building custom software for a clientBuilder is manufacturer, client possibly tooBuilder in full. Client if it's their name or their design
Building an internal tool yourselfManufacturer for own useOnly if a natural person suffers injury or damage to private property
Substantially modifying someone else's product without that manufacturerTreated as manufacturerFor the part the modification touches

The date rule: the product counts, not the damage or the commencement date

The Dutch implementation act passed the Tweede Kamer as a formality item (hamerstuk) on 15 October 2026. The Eerste Kamer, the Dutch senate, had it on the agenda for 10 November 2026, also as a hamerstuk. The commencement date will be set by royal decree (Article IV). When we wrote this on 6 November, no date had been published.

If you ship software, Article III matters more. It inserts Article 189aa into the transitional act for the Civil Code: the changes do not apply to products placed on the market or put into service before 9 December 2026 ("vóór 9 december 2026"). That cut-off is a fixed date in the act itself, so it doesn't depend on the royal decree.

  • The product's date is what counts. Software you release on 20 November 2026 that causes injury in 2029 stays under the old law.
  • A later royal decree doesn't move the line. On the text, a product released on 15 December falls under the new rules once the act applies, even if that's in 2027.
  • The wording splits on 9 December itself. Article 2 of the directive covers products placed "after 9 December 2026", while Article 21 keeps the old directive alive for products placed "before that date". The Dutch explanatory memorandum reads this as "on or after" 9 December, and the bill follows it. Treat a launch on that day as new regime.

Older software isn't safe for good. In its reply to parliament of 3 September 2026, the Dutch government says an update or upgrade does not create a new product, with one exception: a substantial modification. Recital 39 of the directive calls a substantially modified product a new product, and recital 40 fixes the moment at the day the change is actually made. So in our reading, a SaaS platform from 2024 that's substantially modified in 2027 falls under the new rules from then on. Article 189aa doesn't say so in terms, which makes this a question for your lawyer.

We'd advise against bringing a launch forward just to beat 9 December. You trade testing time for protection that can vanish at your next major change, and contract and tort liability don't move at all.

When does a company that fine-tunes or wraps AI become a manufacturer?

A company that embeds or fine-tunes AI can become a manufacturer by three routes. The best-known one, substantial modification, is the least common.

Route 1: you release it as your product

A manufacturer is anyone who develops a product or puts their name or trademark on it (Article 4(10) of the directive). Recital 13 adds that AI system providers within the meaning of the AI Act are treated as manufacturers, and Article 3 of that act defines a provider as whoever develops an AI system, or has one developed, and places it on the market under their own name. A thin layer over an API, sold under your brand, already qualifies. You don't need to fine-tune anything.

The Dutch government's reply explains how the chain then works. An AI system can be a component of the product. The maker of a component is liable if that component caused the defect, while the end-product manufacturer is in principle liable for all defects in the product. So towards an injured person, "the model did it" is no defence. It's an argument for your recourse claim against the supplier.

Route 2: you build it for your own use

Anyone who develops a product "for their own use" is also a manufacturer (Article 4(10)(c)). Liability then requires putting into service: first use in the course of a commercial activity. Recital 26 gives lifts, machinery and medical devices as examples, and the legislator doesn't say whether an office tool qualifies. The real limit is the damage. An internal tool that writes a poor summary doesn't touch this regime. One that controls a machine can.

Route 3: you substantially modify someone else's product

Article 8(2) of the directive makes you a manufacturer when three conditions are met together:

  1. The change happens outside the original manufacturer's control. That control exists if they make the change themselves or consent to it. Consent doesn't follow merely from making it technically possible or not prohibiting it (recital 18).
  2. The change is substantial. First check the product safety rules for that product. If there are none, the change must alter the purpose, performance or type in a way the manufacturer's risk assessment didn't foresee, and change or raise the hazard (Article 4(18)).
  3. You then place the modified product on the market or put it into service.

For AI, that produces a practical distinction. Fine-tuning through the model maker's own fine-tuning service probably sits within the maker's control, because the maker supplies that change. Fine-tuning downloaded open weights happens without the maker, and a licence that allows it isn't automatically consent. If you then point such a general model at something like medical triage, you've changed the purpose and created a new hazard. Route 3 becomes real, on top of route 1.

The Dutch government expects this to be rare, "only in exceptional cases" in its words. Its example is removing an e-scooter's speed limiter through software. If route 3 does apply, you escape liability by proving the defect sits in a part your change didn't touch (Article 11(1)(g)).

Don't confuse this with "substantial modification" in Article 25 of the AI Act. That concerns taking over the provider role for high-risk systems, with its own definition tied to the conformity assessment. A change can meet one test and not the other. Whether bare model weights are a product at all is unanswered: the directive and the reply speak of AI systems, and recital 13 excludes "the mere source code".

Who can claim, and who can't?

Only natural persons have their own right to compensation (Article 5). That means more than consumers: your client's employee, a visitor or a freelancer all qualify. A company can't claim for its own loss.

The directive covers death, injury including medically recognised psychological harm, property damage, and destruction or corruption of data not used for professional purposes (Article 6). It excludes property used exclusively for work, data with any professional use at all (recital 22) and the defective product itself. Pure economic loss, privacy infringements and discrimination don't by themselves trigger liability, says recital 24. And a data leak, recital 20 notes, is distinct from data destruction.

ExampleCovered?
Chatbot gives a business customer a wrong answer, customer loses revenueNo. Contract or tort
Planning software steers a forklift, a warehouse worker is hurtYes, for the injury. Not for the employer's forklift
An update wipes consumers' private photosYes, unless they can be retrieved at no cost (recital 20)
Hiring tool discriminatesNot by itself. Equal treatment law and the GDPR still apply
Public chatbot gives dangerous health adviceOpen question

That last row is open because two sentences in recital 13 pull in opposite directions. Software is "capable of causing damage through its execution", and information isn't a product. Whether a wrong answer is a defect in the system or just bad information is something a court will have to decide.

Companies do end up in the chain. An operator that has paid an injured person can pursue others who are liable (Article 14), and an insurer can step into the injured person's rights (Article 5(2)). No separate EU regime for AI sits alongside this: the European Commission withdrew the proposed AI Liability Directive in the Official Journal of 6 October 2025. For the wider question of who pays for AI mistakes, see AI risks and liability.

Software updates, security flaws and systems that learn

Software updates are covered as long as the manufacturer has control, and it has control simply by being able to supply updates, itself or through a third party (Article 4(5)). The usual defence that the defect only arose after launch then falls away for defects caused by software, updates, or a lack of updates "necessary to maintain safety" (Article 11(2)). With SaaS, where you roll out every release, that control is nearly always there.

The size of the update doesn't matter. The Dutch government is clear on this: even a change with limited impact can lead to liability if that change is what caused the defect.

There are limits on the other side. According to the government, the directive contains no product requirements: a duty to keep supplying updates comes from the Cyber Resilience Act and the consumer directives. The expiry period keeps running from the original release, and a product isn't defective just because a better product or newer version exists (Article 7(3)).

Two passages bear directly on agents. Recital 32 says a manufacturer "that designs a product with the ability to develop unexpected behaviour should remain liable for behaviour that causes harm". And Article 13(1) says a third party contributing to the damage doesn't reduce your liability; recital 55 gives the example of someone exploiting a vulnerability. An attack through prompt injection fits that pattern, in our reading.

What changes in the burden of proof?

The burden of proof in product liability stays with the claimant, and the Dutch memorandum says outright that it isn't reversed. Two new tools make life harder for a software supplier in court.

First, disclosure. A claimant with a plausible case can ask the court to order the defendant to hand over relevant evidence (Article 9), with protection for trade secrets. Second, the presumptions in Article 10. A defect is presumed if you don't disclose, if the product breaches mandatory safety rules meant to protect against that damage, or if there's an obvious malfunction during normal use. Where technical complexity makes proof excessively difficult, a court can also presume a defect or causation if the claimant shows it's likely. Recital 48 names machine learning as an example of that complexity.

All these presumptions can be rebutted, but in practice that takes records. If you can't show which model version, system prompt and settings were running on the day of the damage, rebutting gets hard. It's the same version control you need anyway for running AI agents in production.

The development risk defence survives: a defect that the state of knowledge couldn't reveal doesn't lead to liability (Article 11(1)(e)). The Dutch government deliberately kept it.

Open source: the exemption that doesn't help you

Article 2(2) excludes free and open-source software developed or supplied outside a commercial activity. That exemption protects the maker, not the company using the software. Recital 15 is explicit: if a manufacturer integrates such software into a commercial product, that manufacturer is liable and the open-source maker isn't.

So if you build an open model into your product, your recourse route may be empty. A licence that permits the use tells you nothing about this; see open-weight licences compared. Whether a commercial model maker that publishes weights for free acts outside a commercial activity is uncertain. Don't count on it.

Time limits and caps

ItemRuleSource
Limitation3 years from awareness of damage, defect and operatorArt. 16 directive
Expiry10 years from release, restarted by substantial modificationArt. 17(1) directive
Latent injury25 yearsArt. 17(2) directive
Property damage thresholdThe €500 deductible goesDutch memorandum
Contractual limit towards the injured personNot allowedArt. 15 directive
Financial ceilingNot allowedRecital 56

Sort this out before 9 December

Preparing for 9 December is mostly recording and rereading, per product and per supplier.

  1. Note the first availability of each product, and flag which planned changes could be substantial.
  2. Read your recourse clauses. Article 15 protects the injured person, not you: a liability cap in your model supplier's terms can still limit your recourse. See also the contract terms for AI software.
  3. Watch for recourse waivers. The directive expressly lets small software suppliers, with fewer than 50 staff and no more than €10 million in turnover or balance sheet total (Recommendation 2003/361/EC), agree that you won't pursue them (Article 12(2)). The Dutch memorandum adds that recourse can be settled by contract in the Netherlands anyway. Sign such a clause and that share of the damage is yours.
  4. Call your insurer. The Dutch memorandum says product liability is as a rule included in general business liability cover, and the government's reply says insurability is tailored per sector, product and company. Ask in writing whether software and embedded AI in your own products are covered.
  5. Keep the model version, system prompt, test results and release notes per release, for as long as the 10-year period runs.
  6. If you stop updating a safety-relevant product, record when, and how users find out.

If you only buy software for internal use, this regime asks nothing of you. If you build or sell software that can affect people, this is part of the build: we cover version control and recourse in custom software, and through AI consulting we map which role you have per product. The wider rulebook is in AI legislation in the Netherlands.

Service

Learn more about custom software?

View service →

Curious how much time you could save?

Request a free efficiency audit. We'll analyze your processes and show you where the gains are, no strings attached.