Who Pays When Artificial Intelligence Makes a Mistake?
Introduction
When an artificial intelligence system denies a loan to a solvent client, wrongly labels a patient as low risk, causes an autonomous vehicle collision, or invents a legal precedent, the first question is what went wrong. For the injured party, however, the more important question is: who will pay for the damage? An algorithm has no assets, insurance, or legal personality. Behind its decision lies a chain of people and organizations: the model developer, the data supplier, the integrator, the company deploying the system, the employee who accepts the output, and sometimes the user who abuses it. Consequently, the phrase “artificial intelligence error” easily conceals the fact that this is a human-organized risk.
The core thesis of this essay is that the cost should not be automatically attributed to either the programmer or the end user. Responsibility should be tied to control over the risk, the economic benefit derived from the system, the ability to prevent harm, and access to evidence. The entity that chooses to introduce AI into a business process must generally be held accountable to the damaged party, while the ultimate burden can be distributed regressively among the manufacturer, integrator, data supplier, and other participants. Such an approach protects the victim from the technical and contractual complexity of the chain, while still leaving room for each participant to bear the share corresponding to their omission.
Why an Artificial Intelligence Error Is Not Considered as an Ordinary Error
A classic defective product typically has relatively stable properties. If a brake fails due to a poor material, an expert can inspect the part, reconstruct the production process, and link the defect to the damage. An AI system differs because the outcome can depend on the model, software version, training data, data inputted in a specific case, set threshold, subsequent learning, and the way a human interpreted the output. The same model can be dependable in one environment and unacceptably risky in another. Therefore, an error is often not a single point of failure, but rather the consequence of a combination of design and usage.
The second problem is that the damaged party generally lacks access to system logs, technical documentation, evaluations, internal warnings, or supplier contracts. Even when they obtain data, they may not be able to explain why a complex model produced a specific result. The traditional rule is that the plaintiff must prove damage, illegality or defect, and causation can thus become a formal right without actual access to compensation. If key evidence is exclusively in the hands of the defendants, the legal system must allow for its discovery and, under certain circumstances, use rebuttable presumptions.
The third problem arises because AI acts convincingly even when it is wrong. A generative system can formulate an incorrect answer in a grammatically flawless manner, while a scoring system can present a discriminatory pattern as a neutral number. The danger lies not only in the erroneous output, but also in the trust that the organization places in the automated recommendation. The “human in the loop” does not remove responsibility if the employee lacks the time, data, expertise, or authority to truly alter the decision.
The fourth characteristic is the variability of risk throughout the lifecycle. A supplier can release a new model version, a user can connect the system to an additional database, and an integrator can change the instructions without formally amending the underlying contract. A system that passed initial testing therefore does not necessarily remain the same system. Accountability must encompass the post-deployment period: performance monitoring, incident reporting, update control, and the capability for safe withdrawal. It is particularly important to separate a model error from an automation error. A model may only provide a recommendation, but a software process can automatically turn that recommendation into a request rejection, a payment cancellation, or a machine’s physical action. The shorter the path from the model’s output to the real-world consequence, the greater the organization’s duty to introduce independent checks and technical safeguards.
This dynamic demonstrates why a one-time user consent or initial certification is not sufficient. Reasonable care is an ongoing obligation that accompanies the system for as long as a participant can influence its behavior.
Who Can Be Liable to Pay
The first candidate is the manufacturer or supplier of the AI system. Their liability is strongest when the damage stems from a design defect, inadequate testing, an unsafe update, known bias, or inaccurate claims regarding the product’s capabilities. A supplier should not be absolved of liability through a generic warning that the model “may make mistakes” if they simultaneously market the system as fit for medical, financial, or other sensitive decision-making. A warning is meaningful only if it is specific, understandable, and accompanied by limitations that correspond to the actual risk.
The second candidate is the organization deploying and using the system i.e., deployer. It selects the business purpose, determines the input data, sets thresholds, decides whether the output will serve as a recommendation or a final decision, and organizes human oversight. To the client, patient, employee, or consumer, this organization is the visible face of the service.
The third candidate is the professional who relies on AI. A lawyer, physician, engineer, auditor, or financial advisor owes an independent duty of professional care. Using a tool may be reasonable, but delegating professional judgment to a tool generally is not. Courts are increasingly sanctioning court filings containing fabricated sources because the signatory guarantees that they have verified the content. Consequently, the professional will often be liable to the client, while any claim against the supplier depends on the contract, product claims, and evidence of a defect.
The fourth candidate is the user who intentionally circumvents safeguards, inputs illegal data, or uses the system outside its clearly designated purpose. Their action can sever or mitigate the causal link with respect to other participants. However, the concept of misuse must not become a universal defense for manufacturers. If a certain behavior was reasonably foreseeable, the design and instructions should account for it. Finally, insurers and compensation funds can distribute the residual risk, especially in autonomous transport, medicine, and critical infrastructure, where a single error can cause catastrophic damage.
The European Union Between Regulation and Liability for Damages
The European Union has constructed the most developed horizontal regulatory framework through the Artificial Intelligence Act. As of August 2, 2026, the European Commission and national authorities have begun applying a new set of rules, including transparency obligations for certain interactive systems and the labeling of artificially generated or modified content. The Act distributes obligations according to role and risk level: providers of high-risk systems must manage risks, data quality, documentation, logs, human oversight, accuracy, and cybersecurity, while deployers of such systems have their own operational duties.
However, a regulatory fine and compensation for damages are not the same thing. The AI Act provides supervisory measures and heavy monetary fines, but it does not establish a uniform procedure in which every victim of an AI error automatically receives compensation. A breach of a regulatory obligation can serve as important evidence of negligence or defectiveness, but the injured party must still find an appropriate cause of action under national contract or tort law, data protection law, anti-discrimination rules, or product liability.
A significant change is brought by EU Directive 2024/2853 on liability for defective products. It explicitly treats software, including AI systems, as a product, and its developer as a producer for the purpose of strict liability. The assessment of defectiveness takes into account the system’s ability to continue learning after being placed on the market, security updates, and cybersecurity vulnerabilities. The Directive allows courts to order the disclosure of relevant evidence and establishes presumptions when the defendant fails to comply or when proving fault is excessively difficult due to technical complexity. Member States are required to implement its rules by December 9, 2026, and the new regime applies to products placed on the market after that date.
Nevertheless, this regime does not cover every type of damage. Its focus is on harm to physical integrity, private property, and certain damage to individuals’ data; pure economic loss, discrimination, or violations of privacy require other legal bases. Additionally, the Commission withdrew its proposal for a specific AI Liability Directive. As a result, the intended horizontal mechanism that would have facilitated proving fault and causation in cases outside product liability was not realized.
The United States and Liability Through Existing Law
The American approach is less centralized. Federal policy in 2026 emphasizes innovation, sectoral action by existing authorities, and limiting the fragmentation of state regulations. White House recommendations call for avoiding the creation of a new federal regulator solely for AI and for states to retain the application of general laws on fraud, consumer protection, and child protection, while simultaneously considering restrictions on state-specific rules for AI development. This does not mean an absence of liability. Lawsuits continue to be grounded in contract, negligence, product liability, discrimination, privacy, copyright, and sector-specific rules.
The weakness of this approach is its inconsistency. The same error can yield different consequences depending on the state, sector, and contractual clause. The advantage lies in the courts’ flexibility to adapt existing principles to the facts. The American experience thus shows that AI does not need to be explicitly mentioned in the law for liability to exist, but also that without rules on evidence, an injured party may find it difficult to obtain the models, logs, and internal tests required to succeed in litigation.
Bosnia and Herzegovina and Liability Without a Specific AI Law
Bosnia and Herzegovina has no specific law comprehensively regulating civil liability for damage caused by AI systems. This does not create a legal vacuum. General rules of contract and tort law allow for compensation for damages based on fault, breach of contract, and, when the conditions are met, hazardous objects or activities. Regulations on consumer protection, anti-discrimination, labor, medical services, and professional liability can be applied according to the nature of the relationship. A court would therefore first have to identify the human or organizational decision that led to the harm: the choice of an inappropriate system, poor data, a lack of verification, inadequate warnings, or a failure to respond to a known incident.
The new Law on Personal Data Protection of Bosnia and Herzegovina, published in the Official Gazette of BiH No. 12/25, is particularly significant. Article 24 grants individuals the right not to be subject to a decision based solely on automated processing that produces legal or similarly significant effects, subject to prescribed exceptions and safeguards. Article 112 recognizes compensation for both material and non-material damage, regulates the liability of controllers and processors, and permits recourse claims among participants after the damaged party has been paid in full. The law thus already provides a model for distributing liability within a complex digital chain.
The greatest practical obstacle will be standard proof. Domestic procedural rules were not drafted for models whose logs, documentation, and data may reside with a foreign cloud provider. Therefore, a contract between a domestic user and a supplier should ensure log retention, access to technical information, cooperation during incidents, audit rights, notification of model updates, and recourse for damage stemming from the supplier’s fault. Without such clauses, a domestic company may be held liable to its client or employee, while subsequently lacking the evidence needed to transfer the burden onto the party that actually caused the harm.
A functional approach would be useful for judicial practice. An AI system should not be preemptively classified as a hazardous object, but autonomy, the potential scale of damage, and the impossibility of effective human intervention can justify a stricter standard of care—and, in particularly high-risk applications, statutory strict liability. Until then, courts will need to carefully apply rules regarding the burden of proof, document production, and expert testimony so that technological opacity does not become a reward for the party controlling the information.
Proving Causation and Allocation of the Burden
Lawsuits concerning AI-induced damage will rarely be won by debating whether an algorithm “thinks.” The decisive factors will be technical and organizational trials: which model version was used, what inputs were provided, what limitations were known, who approved the use, whether a warning existed, whether a human could alter the outcome, and what would have occurred without the disputed decision. For this reason, record-keeping is a matter of legal liability, not mere technical administration.
A fair process requires an asymmetrical burden of proof. The injured party should demonstrate a plausible connection between the system and the harm, alongside facts pointing to an omission. Subsequently, the entity controlling the documentation should be required to explain the system and produce relevant logs. If it refuses, destroys data, or fails to maintain records it was required to keep, the court should be permitted to presume that the evidence would have favored the damaged party. Such an approach does not introduce automatic liability for every bad outcome; it prevents opacity from being used as a procedural shield.
Causation must remain concrete. An inaccurate output does not generate liability if no one reasonably relied upon it or if the damage would have occurred regardless. Conversely, human confirmation does not sever the causal link if it was purely formal and predictably dependent on an automated recommendation. The court must analyze the actual decision-making power, rather than a job title or the button by which an employee “approves” the result.
Conclusion
When AI makes a mistake, the machine will not pay. The cost will be borne by humans, companies, insurers, and, if the law fails, the victim themselves. The legal task is not to invent consciousness or culpability for an algorithm, but rather to distribute responsibility among the entities that created, introduced, and controlled the risk. Primary protection for the damaged party requires holding the entity that placed the system into a specific relationship with them liable, while allowing for recourse against the manufacturer, integrator, or user whose oversight or failure caused the harm.
The European Union combines the regulatory obligations of the AI Act with renewed liability for digital products; however, the withdrawal of the specific AI Liability Directive leaves gaps regarding pure economic loss, discrimination, and proving fault. The United States relies more heavily on existing private and consumer law, whereas Bosnia and Herzegovina must utilize general obligations rules and its new data protection regime. In all three systems, the ultimate answer will depend on the same principle: liability should follow actual control, benefit, and evidence. Only then will technological complexity fail to become a justification for leaving damage uncompensated.
Author: Aleksandar Sajic
