On 1 September 2026, Michel Nolet de Brauwere, first advocate general at Belgium's Court of Cassation, delivered the court's traditional opening-of-the-judicial-year address, the Mercuriale. He devoted it entirely to artificial intelligence. The 32-page text is titled "Science sans conscience n'est que ruine de l'âme". It opens with an admission that is itself instructive: the author writes that he resisted the temptation to have an AI tool draft the speech, and that no prosecutor general among the six in the Kingdom had previously devoted a Mercuriale to the subject.
As reported in the hook for this piece, the address catalogues thirteen risks. They include hallucinated rulings, algorithmic jurisprudence, opacity, bias, cognitive offloading, erosion of judicial independence, and dependence on non-European vendors exposed to the US CLOUD Act. It does not call for a ban. It insists that the final decision stay with a human judge. Separately, the presidents of Belgium's Constitutional Court, Court of Cassation and Council of State have asked for a data protection impact assessment on the use of Microsoft Copilot within government.
The strongest case for caution
The sceptics' argument deserves a fair statement. Courts do not behave like ordinary software customers. A judgment is an exercise of public power over a person's liberty, property or family, and it must be reasoned in a way the losing party can contest. A system that produces fluent but fabricated case law, or that nudges a tired judge toward the statistical median, can damage legitimacy in ways that are invisible until an appeal. The EU legislature reached the same conclusion in 2024. Annex III point 8(a) of the AI Act classes as high-risk any AI system "intended to be used by a judicial authority or on their behalf to assist a judicial authority in researching and interpreting facts and the law and in applying the law to a concrete set of facts", including in alternative dispute resolution. On that view, thirteen risks justify the strictest possible regime.
Why a list of risks is not a case for prohibition
We read the Mercuriale differently, and so, apparently, does its author. Look at what the thirteen risks have in common. Most of them are failures of process and procurement, not of technology as such: unverified output, unexplained reasoning, unmanaged vendor dependence, unmeasured bias. Those are fixable with disclosure rules, verification duties, audit rights and contract terms. A prohibition would not fix any of them. It would only push use underground, into personal accounts and consumer chatbots that no court has assessed.
The speech's own opening shows this. A senior prosecutor tells his audience he could not guarantee that machine-drafted text would have been less pertinent than his own. Judges and clerks face large caseloads, and the benefits of summarisation, translation and research support are real. The principled line is the one the speech draws: the tool may assist, but the judge decides and answers for the decision. That is proportionate regulation in its simplest form. It targets the locus of accountability and leaves the technology available.
The sovereignty risk is the most actionable one
Of the thirteen items, reliance on vendors subject to the US CLOUD Act is the one that law can address most concretely. Case files hold sensitive personal data, and a court that cannot say where its data is processed and who can compel access to it has a data protection problem before any question about model accuracy arises. The request from Belgium's three highest courts for an impact assessment on Copilot is the right instinct. It treats a productivity tool as a processing operation requiring documented analysis, not as a licence renewal. Openness cuts both ways here, though. Courts should demand portability, auditability and model-agnostic contracts. They should not respond by funding closed national alternatives that are no more transparent than the ones they replace.
The regulatory gap is now explicit
The speech lands in a timing gap. The Digital Omnibus on AI, approved by the Council on 29 June 2026, moves the obligations for stand-alone Annex III high-risk systems from 2 August 2026 to 2 December 2027, according to Cuatrecasas. Law-firm analyses attribute the delay mainly to harmonised standards not being ready. A Gibson Dunn summary of the deal also notes that national authorities remain competent for judicial authorities.
The delay is defensible. Imposing conformity obligations before the technical standards exist would give courts and vendors a rulebook they could not follow. But it leaves roughly sixteen months in which judicial AI use is governed by little more than data protection law, professional ethics and whatever internal guidance courts adopt. The Mercuriale effectively fills that vacuum with a normative statement from within the judiciary.
What courts should do before December 2027
Four steps would turn the speech into practice without waiting for Brussels.
- Publish an internal use policy. It should say which tools judges and clerks may use, for which tasks, and with what verification. Citing an unchecked machine-generated authority should be a professional-conduct issue.
- Require disclosure. If AI materially assisted a draft, the file should say so, so that parties can challenge it.
- Run impact assessments before deployment. Belgium's courts have started this with Copilot. It should become the default, with results shared across the Member States' judiciaries.
- Draft contracts for exit. Data location, audit rights and the ability to switch vendors matter more than any single model's benchmark score.
None of this requires a ban, and none of it needs new primary legislation. The danger the Mercuriale identifies is not that judges will use AI. It is that they will use it unevenly, without records, and on infrastructure they do not control, while the formal regime waits for standards. Courts that adopt clear policies now will be better placed to meet the AI Act's requirements when they arrive, and the public will be better able to trust their decisions in the meantime.