> Bron: https://neuralex.nl/en/blog/aansprakelijkheid-ai-antwoord > Who bears responsibility when an AI system gives a wrong answer? Learn how liability can be divided across the AI supply chain and which contractual and technical measures can reduce risk. [Back to Insights](/en/blog) AI Act ·6 October 2026 ·6 min read # Who is liable when an AI answer is wrong? Who bears responsibility when an AI system gives a wrong answer? Learn how liability can be divided across the AI supply chain and which contractual and technical measures can reduce risk. There is no single answer to who is liable when an AI-generated answer turns out to be wrong. Liability depends on who did what, what the parties agreed contractually, what kind of damage occurred and which legal rules apply. The model provider may play a role, but so may the software vendor, integrator, organisation deploying the system and, in some situations, the employee acting on its output. For a customer or another affected party, “the AI did it” will generally not be a sufficient explanation. An organisation chooses to introduce an AI system into its processes and remains responsible for how that system is used. That does not automatically mean that all liability rests with the organisation. Behind the scenes, responsibility may be allocated among several parties in the supply chain. ## Who is involved in the AI supply chain? An AI service often involves several parties. The **model provider** may supply a language model or API. A **software vendor** builds an application around it. An **integrator** connects that application to company data, processes and systems. Finally, an **organisation** deploys the solution for employees, customers or other users. These roles can overlap. An organisation may act as its own integrator, while a software vendor may operate its own model. That is why responsibility should not be assessed by looking only at the underlying AI model. You also need to consider the complete application: which data it uses, what instructions it receives, which sources it can access and which actions it is allowed to perform. Systems that fall within specific European AI rules may also be subject to requirements relating to risk management, transparency and oversight. For a practical introduction, see [AI Act for SMEs](/en/blog/ai-act-voor-het-mkb). ## Why “the AI did it” is not a defence If a chatbot gives incorrect information to a customer, that customer will usually be dealing with your organisation rather than with the underlying model provider. If you have allowed the chatbot to communicate on behalf of your business, the problem may therefore arrive at your organisation first. This is particularly relevant when the AI performs a task that forms part of your normal service. Examples include providing product information, giving customer guidance, performing calculations, issuing instructions or supporting decisions within a business process. The more serious the consequences of an error can be, the more important appropriate controls become. A disclaimer does not automatically change this. Informing users that “AI can make mistakes” may be useful, but it is not a general exemption from responsibility. An organisation that knows a system is unreliable for a particular task cannot necessarily transfer all consequences to the user. ## When can the organisation itself be responsible? An error may be more likely to be attributed to the deploying organisation when an AI system is used for an unsuitable purpose, has not been tested adequately or is allowed to make decisions without appropriate oversight. The same applies where known limitations are ignored or employees routinely rely on AI output without verification. System architecture also matters. A general-purpose language model that is free to answer from its internal model knowledge presents different risks from a system that grounds answers in controlled company sources. Retrieval-augmented generation can help tie answers to your own documents, but it does not eliminate errors altogether. See [RAG without hallucinations](/en/blog/rag-zonder-hallucinatie) for more detail. For actions with financial, legal or other significant consequences, human approval is often an important safeguard. [When should a human approve?](/en/blog/wanneer-moet-een-mens-goedkeuren) explains how to decide which actions can be automated and which require intervention. ## What can you regulate through contracts? Contracts between parties in the AI supply chain are important because they can define who is responsible for particular elements of the service. Relevant provisions may cover allocation of responsibilities, warranties, limitations of liability, indemnities and service levels. A provider may specify the purposes for which its service is intended, the level of service it aims to provide and the support available when incidents occur. An integrator may accept responsibilities relating to configuration, security and testing. The deploying organisation may remain responsible for human review and for the final use of the system within its business processes. Contractual limitations are not unlimited. Not every form of liability can necessarily be excluded in every situation. Agreements between a supplier and its customer also cannot automatically determine the rights of consumers or other third parties. European product liability rules are also evolving and may, in certain circumstances, extend to software. The exact rules and their applicability should be checked for the specific product, service and situation involved. ## How can you reduce the risk in practice? Liability risk cannot be managed through contracts alone. The technical and organisational design of the AI system is at least as important. - **Provide sources:** where possible, let the system show what information an answer is based on so that important claims can be verified. - **Introduce human review:** require risky outputs to be checked before they are communicated externally or trigger an action. - **Use honest disclaimers:** explain what the system can and cannot do without suggesting that a disclaimer removes all responsibility. - **Keep relevant logs:** record the inputs, sources, model versions and actions involved where doing so is legally and technically appropriate. - **Measure quality:** regularly test whether answers remain accurate, relevant and sufficiently grounded. Monitoring uptime alone is not enough. See also [AI monitoring](/en/blog/ai-monitoring). - **Create an escalation route:** decide when an employee, specialist or lawyer must take over. These safeguards do not guarantee that errors will disappear. They do, however, help demonstrate that risks have been assessed systematically and that the organisation does not rely blindly on model output. ## Which questions should you discuss with your supplier and lawyer? - Which party is responsible for the model, application, integration and final decision? - For which use cases does the supplier state that the system is intended? - What warranties are provided regarding functionality, availability, security and support? - What limitations of liability and indemnities are included? - What happens if a model error causes damage to a customer or another third party? - Who investigates an incident and which logs or other evidence are available? - Which AI outputs must be reviewed by a human? - Which risks explicitly remain with your organisation? - Which legal or sector-specific requirements apply to the intended use? ## Have your situation reviewed by a lawyer Neuralex does not provide legal advice. Liability involving AI depends on the specific application, contractual relationships and rules that apply to your organisation. A qualified lawyer should determine what applies in your situation, particularly where AI can affect financial, legal, safety-related or other significant interests. AI Act & compliance ## Do you know who is responsible when your AI assistant gets something wrong? We build with source citations, human review and logging so errors can be prevented and traced. For the legal assessment you work with a lawyer. Curious what that means for your situation? [Ask your question](/en/contact) [Read about RAG without hallucination](/en/blog/rag-zonder-hallucinatie) --- Volledige (opgemaakte) versie: https://neuralex.nl/en/blog/aansprakelijkheid-ai-antwoord