Will artificial intelligence eventually replace judges and lawyers?

No, and that shouldn't be allowed, says the First Advocate General at the Court of Cassation. On September 1, 2026, Michel Nolet de Brauwere delivered the opening address for the judicial year, the so-called mercurial, entirely devoted to artificial intelligence and the justice system. He lists thirteen risks, ranging from fabricated rulings to a Court overwhelmed by cases, but explicitly does not advocate for a ban. We read the speech and offer a few comments.

Why the Public Prosecutor's Office Is Talking About AI This Year

Every year, the Court of Cassation opens the judicial year with a formal session at which the Attorney General or an Advocate General delivers a speech on a current legal topic. This year, the First Advocate General chose to address the topic of AI, under a title borrowed from Rabelais: “Science without conscience is but the ruin of the soul.”.

The reason is concrete. Magistrates are increasingly being presented with legal opinions drafted with—or even entirely by—AI tools. None of the country’s six attorneys general had previously addressed the subject in a legal journal. The first attorney general felt that the matter could no longer be put off, even though he admits that his text was already outdated before the ink was dry.

His perspective is nuanced. AI can make the administration of justice faster, more accessible, and more efficient—in document review, in making case law accessible, in translation, and in recurring disputes. But the final decision must remain in human hands. According to him, a judge’s legitimacy stems from his humanity.

Thirteen risks, and two cases that capture the imagination

The speech lists thirteen risks: the illusion that AI can replace judges and attorneys, uncontrollable systems, algorithmic jurisprudence, lack of transparency, bias and manipulation, errors and hallucinations, dependence on foreign suppliers, cognitive offloading, de-responsibilization, violation of attorney-client privilege, overburdening of the courts, loss of independence, and a purely quantitative assessment of the justice system.

Two specific cases illustrate why the public prosecutor’s office is concerned. In a recent cassation case, the plaintiff’s brief contained eight references to rulings by the Court itself. Two of them did not include the legal principle cited. The remaining six were simply incorrect. The attorney who had signed the brief blamed the errors on a former staff member and withdrew the unverifiable citations, without stating whether an AI tool had generated them. For the First Advocate General, this constitutes a breach of the duties of competence, integrity, and loyalty, and he calls on the disciplinary authorities to send a strong message. We previously wrote in our blog about what such a fabricated source could already cost a lawyer in Belgium today: Fictitious AI-generated legal rulings in conclusions.

The second case concerns volume. A brief in a criminal case was nearly 100 pages long and contained more than 50 grounds of appeal. As a rule, the Court must respond to each ground of appeal. The First Advocate General strongly suspects that AI generated the document and cites the looming flood of cases overwhelming the Court in the near future as the greatest danger. Anyone who can churn out dozens of grounds for appeal at low cost will paralyze a court that must respond in detail to each one.

In addition, the speech highlights the judicial system's dependence on digital technology. The judicial system's IT environment runs on software from non-European vendors, and the American Cloud Act allows the U.S. government to request data from those providers, even if it is stored in Europe. Judges at the International Criminal Court have already experienced what the unilateral U.S. sanctions As of February 6, 2025, this means: blocked email accounts and payment methods. On June 26, 2026, the presidents of the Constitutional Court, the Court of Cassation, and the Council of State, therefore asked the federal government in a joint letter—quoted in the article—to conduct a data protection impact assessment for the use of Copilot within the government.

What remedies does the First Advocate General propose?

The opinion concludes with lines of reasoning, not detailed proposals. The judge could instruct the parties to limit the length of their briefs, following the example of the Dutch Supreme Court. Arguments based on legal doctrine or case law without precise and verifiable citations could be declared inadmissible. The Court’s obligation to respond could be limited through a simplified selection procedure, such as those used by the European Court of Human Rights and the French Court of Cassation. Extending the bar’s monopoly at the Court of Cassation to all matters is also under consideration.

The existing legal arsenal is also addressed. Anyone who uses the legal system for purposes that are clearly intended to delay proceedings or are unlawful already risks a fine of 15 to 2,500 euros under Article 780bis of the Judicial Code, although that provision does not apply in criminal or disciplinary cases. In addition, the disciplinary authorities of the bar and the attorney’s civil professional liability remain in effect.

Meanwhile, the regulatory framework is expanding. The Council of Europe Framework Convention on Artificial Intelligence took effect on April 21, 2026, and requires the States Parties to ensure that AI systems do not undermine the independence of the judiciary or access to justice. The Artificial Intelligence Regulation (AI Act) classifies AI systems that assist judicial authorities in investigating and applying the law as high-risk. And on January 31, 2025, the OVB and avocats.be adopted their guidelines for the use of artificial intelligence To: The attorney must review the AI's output, including its sources, but has no general obligation to report the use of AI.

What we agree with, and what we take issue with

The analysis in the essay is largely accurate. The point about digital sovereignty is compelling and has received insufficient attention in the Belgian debate. The warning about cognitive unloading is also based on genuine research: the essay refers to the MIT Study on Cognitive Debt when AI assistants are used systematically. Those who permanently outsource their legal thinking lose the ability to do it themselves. Nevertheless, three points deserve careful consideration.

First, reason relies on a European framework that, while formally in place, currently lacks teeth. The AI Act classifies judicial AI as high-risk, but the digital omnibus regulation extended the high-risk obligations to December 2, 2027, and August 2, 2028, as we analyzed in our blog post about the deferral of high-risk obligations. Anyone who builds or purchases an AI tool for the justice system today will therefore be operating for at least another fifteen months without an enforceable high-risk regime. This reinforces the case for establishing our own guidelines rather than undermining it, but it’s worth noting.

Second, this line of reasoning draws too sharp a distinction between unsafe, free “omnivorous” tools and safe, protected environments. Even AI applications that draw exclusively from a legal database can miscite or misattribute sources. The Supreme Court case file containing the eight incorrect citations does not prove that the tool was the problem, but rather that the signatory failed to verify the information. The duty to verify is the constant, regardless of which tool is used. The OVB guidelines already state this, by the way: the attorney remains ultimately responsible for the output. The court acknowledges this, but the emphasis on the type of tool risks diverting attention from the actual error—signing without verification.

Third, some remedies cut deep into the rights of the litigant. A limitation on the Court’s obligation to respond or a summary dismissal affects not only the attorney who has fifty grounds for appeal drafted, but also the party with three carefully crafted grounds that are rejected during the selection process. The problem of misconduct—abuse of the judicial process—can already be sanctioned today through Article 780bis of the Judicial Code, court costs, and disciplinary law. We believe that consistently applying that arsenal first is the better course of action than scaling back the duty to state reasons—a safeguard that reason itself defends elsewhere. The same applies to the comparison with students’ duty of transparency: an exam tests the person, while a trial tests the argument. It is therefore justifiable that the bar associations do not impose a general duty to report, although a lawyer must never lie when the judge asks.

We previously wrote about the related question—whether judges themselves may use AI and what that means for the adversarial process—in our blog post on AI in the Judicial System and the Right to a Fair Hearing. The Mercuriale confirms this analysis: an algorithmic suggestion that is not subject to debate is incompatible with a fair trial.

Specifically, what does this mean?

For lawyers. Verify every source provided by an AI tool, even when the tool is running in a secure environment. The Bar Association has announced that law firms will report suspicious references to the disciplinary authorities, and the bar association officers will follow suit. Never enter non-anonymized documents into an unsecured tool: doing so violates attorney-client privilege and the General Data Protection Regulation, which you can read more about on our page regarding privacy and data protection.

For businesses and their legal counsel. Anyone involved in litigation can expect their attorney to ensure that AI enhances quality without replacing human oversight. Ask about this during the intake. Anyone who builds or purchases AI tools for legal work should take into account the “high-risk” classification that will become enforceable starting at the end of 2027, as well as the broader legal framework surrounding artificial intelligence that already applies, from transparency rules to data protection.

Frequently asked questions (FAQ)

Does my lawyer have to tell me that he uses AI?
No. The guidelines issued by the OVB and avocats.be do not impose a general obligation to disclose—neither on the client nor on the judge. However, the attorney remains fully responsible for what he signs, and if the judge explicitly asks him whether he used AI, he must not lie.

Can a judge in Belgium have a judgment written by AI?
No. The AI Act states that AI may support a judge’s decision-making but not replace it, and Minister of Justice Annelies Verlinden confirmed in an interview in the Juristenkrant on May 13, 2026, that the administration of justice remains a human activity and that a judgment cannot simply be written by a chatbot. Purely administrative uses, such as pseudonymizing decisions, are not covered by this provision.

What are the risks for a lawyer who cites fabricated case law?
The fine of 15 to 2,500 euros for abuse of the legal process is formally imposed on the party to the proceedings, not its attorney, and furthermore does not apply in criminal or disciplinary cases. The attorney himself risks having the court file a complaint with the bar association, disciplinary sanctions that may include disbarment, and a liability claim—primarily from his own client, who may seek to recover the fine or a judgment for vexatious and reckless litigation from him.

Conclusion

The September 1, 2026, edition of *De Mercuriale* prohibits nothing and promises nothing, but sets the agenda: AI belongs in the Belgian justice system, provided that humans make the decisions, the sources are reliable, and the infrastructure is sovereign and secure. The analysis is convincing. Vigilance is one of the necessary remedies, because a justice system that arms itself against AI abuse by providing fewer answers pays for that protection with the rights of the diligent litigant. For anyone working with AI in Belgium—whether in the legal profession or elsewhere—the message is now the same as that of the bar associations: use the tool, but never sign off on it without verification.