What are the risks if AI-generated legal rulings end up in your conclusions?

Lawyers and parties who use generative AI to draft legal briefs are running into a growing problem: so-called “hallucinations”—a court ruling or statutory provision that sounds convincing but does not exist. The answer to the question above is clear: the penalty can range from a reprimand regarding damages to a fine, and liability lies with whoever submits the text, not with the AI.

What exactly is an AI hallucination?

A hallucination is a response from an AI system that appears plausible but is, in reality, fabricated and therefore incorrect. The model always tries to provide a coherent answer and fills in the gaps with information it has seen elsewhere, even going so far as to fabricate a completely fictitious case number or ECLI number if necessary.

The danger lies in the combination of form and content. A fabricated Supreme Court ruling often comes with a date that appears correct, a credible case number, and a smoothly worded reasoning. Without critical scrutiny by the user, such a fabrication can easily slip into a conclusion. That this phenomenon is not a marginal occurrence is evident from the database which is maintained by French legal scholar Damien Charlotin and which has already documented nearly 1,400 court decisions worldwide in which AI hallucinations were identified.

Four Responses from the Judge

A number of court rulings show that a judge who detects an AI hallucination has a range of options available. These options vary in severity, and the financial consequences differ greatly depending on the course of action chosen.

The Reprimand

The mildest response is a simple reprimand. In France, they chose the court in Périgueux He did so and invited the petitioner and his counsel to verify from now on that their sources are not hallucinations.

The Ghent Court of Appeals ruled that a judgment dated September 18, 2025 a step further. It first compiled a list of the erroneous and non-existent references linked to the pages containing the conclusions, in order to subsequently disregard those sources. The court cited AI as a potential asset to legal practice but expressed its frustration with its ill-considered and uncontrolled use, which, according to the court, causes a significant waste of time for courts and other attorneys, and thus represents a societal cost. The court called such behavior completely unacceptable.

The Commercial Court in Ghent also handled that case a judgment dated December 15, 2025. The court had the unusual idea of asking Google’s AI mode itself about three cited rulings from the Court of Cassation and the Constitutional Court, and included the answers in its reasoning. In this way, it demonstrated in black and white that those rulings were completely fabricated.

Guidelines Issued by the Court

A second response is to issue general guidelines. The Irish Court of Appeal did so in a judgment dated March 26, 2026 and formulated a set of dos and don'ts for the responsible use of AI in the procedure.

This approach fits well within a common law system, where precedent holds a special status. In the Belgian legal landscape, which does not recognize the rule of precedent, such a binding list would be less obvious. Nevertheless, it is conceivable that a particular court might develop its own approach to AI. Although not formally binding, in practice the presiding judge’s expectations do indeed influence the way a lawyer drafts his or her arguments.

The Irish guidelines boil down to five principles: responsible use of AI that does not mislead the judge; disclosure of information to the other parties and to the judge regarding such use; each party’s responsibility for the content of its case file; an independent review of everything that is presented; and a prohibition on citing case law without verifying its authenticity and relevance. Essentially, this is the critical thinking that every lawyer is taught during their training.

Compensation to the other party

This is where things get serious financially. The Antwerp Court of Appeals ruled in a judgment dated December 4, 2025 that the persistent filing of ill-considered claims, based on baseless arguments, was manifestly reckless. That conduct caused damage to the opposing parties and was compensated with an amount of 7,500 euros per party, as determined in accordance with the principles of equity.

However, a nuance is in order here. The ruling shows that the users of the AI notary were serial litigants who had filed dozens of lawsuits in bad faith, and that they were not represented by an attorney. The irresponsible use of AI undoubtedly played a role, but does not appear to have been the sole basis for the conviction.

That is legally relevant. Anyone claiming damages must prove that they have suffered damage that is causally related to a fault, as provided in Article 6.5 of the Civil Code (BW) required. Unreasonable use of AI in and of itself will not always meet those conditions. Substantial financial compensation is therefore not automatic, even though the risk of such a ruling does indeed exist.

Civil fines

The fourth and most severe sanction is a civil fine. A judge may not, on his own initiative, award damages to another party, as this would exceed the scope of the claim. However, the judge may, ex officio, impose a fine on a party that uses the legal proceedings for manifestly dilatory or unlawful purposes, pursuant to Article 780bis of the Judicial Code (Jud. C.).. That fine ranges from 15 to 2,500 euros and goes to the Belgian government, not to the other party.

The procedural requirement is important. If no damages are claimed, the judge must first give the parties an opportunity to explain themselves before imposing the fine. In the same case, the Antwerp Court of Appeals imposed a fine of 2,500 euros. The judgment of the Commercial Court in Ghent dated December 15, 2025, is the most telling on this point: the court dismissed the claim on its merits but reopened the proceedings to allow the plaintiff to comment on a possible fine, precisely because of the careless and uncontrolled use of AI. In this case, the use of AI was the direct cause.

Can the maximum litigation reimbursement be awarded in addition to damages?

An interesting related question concerns the cumulation of claims. In the Antwerp case of December 4, 2025, the parties sought not only damages but also the maximum litigation reimbursement. The court of appeals denied this request, ruling that it would constitute a duplication of the compensation awarded for a vexatious and reckless appeal.

That reasoning is not convincing. According to Article 1022 of the Judicial Code, the litigation reimbursement is a lump-sum allowance for the costs and fees of the prevailing party. It is not intended to compensate for damages, but rather to cover expenses incurred. These expenses increase as a party’s conduct requires more research and more detailed responses.

Based on that logic, there is no fundamental obstacle to claiming the maximum litigation reimbursement in addition to damages for vexatious and reckless litigation. Anyone who incurs significant attorney’s fees due to the opposing party’s abusive use of AI in an attempt to uncover something that turns out not to exist has incurred a real cost that justifies the lump-sum payment.

That is not a theoretical construct. In a ruling dated March 25, 2026 (2025/AR/774) The Antwerp Court of Appeals awarded increased litigation costs to a party that had made the proceedings chaotic through its use of AI. The pleadings were confusingly drafted, with constantly changing facts and legal grounds, forcing the opposing party to submit extensive additional briefs. The increased litigation costs therefore represent a real additional financial risk for anyone who carelessly incorporates AI into their pleadings.

Specifically, what does this mean?

The bottom line is simple: responsibility for the content of a court filing lies with the party that submits it. AI is no excuse, and the lawyer’s familiar duty to conduct due diligence remains fully in effect.

In addition to the legal consequences, professional ethics also play a role. The Flemish Bar Association (OVBà) and the OBFG adopted a joint resolution on January 31, 2025, guidelines regarding the use of AI by attorneys. These rules establish, among other things, the duty of competence, the protection of personal data, and attorney-client privilege. A violation may result in disciplinary action. Furthermore, a financial penalty may lead to a claim for damages filed with the professional liability insurer, while a reprimand primarily damages the lawyer’s reputation, both among clients and in the courts where the lawyer practices.

So the practical lesson is not that AI is off-limits. It is a valuable tool, provided that every reference is verified before it appears in an article. Check every court ruling, every case number, and every statute at the source, just as you would review suggestions from a spell-checker or a human editor before incorporating them.

Conclusion

Belgian courts are responding with increasing severity to fabricated sources cited in court documents, and the consequences range from a reprimand to increased court costs, damages, and even a fine. Anyone who uses AI remains personally liable for what they submit, both legally and in terms of professional ethics.