Hallucinated Rulings, Invalid Briefs: What Lawyers Need to Know About Using AI Now

When AI Invents What Doesn't Exist – and the Court Notices

A signed pleading, submitted via the secure electronic lawyer's mailbox, complete with marginal numbers, case references, and ostensibly well-sourced arguments: on the surface, the appeal against the refusal of leave before the Federal Labour Court (Bundesarbeitsgericht, BAG) looked like an ordinary legal document. But closer scrutiny by the panel revealed the full extent of the problem. Four of the cited decisions simply did not exist. Three others were misrepresented in their content. The court drew the inevitable conclusion: the appeal was inadmissible – because without genuine review by counsel, no effective legal representation had taken place.

With the BAG's ruling of 30 September 2026 (case no. 2 AZN 339/26), a landmark decision has emerged that is attracting considerable attention in the legal community – and well beyond employment law. The ruling makes it unequivocally clear: AI-generated content incorporated into court submissions without independent verification can render the entire appeal inadmissible. This is a procedural risk whose full severity many practitioners may not previously have appreciated.

What the BAG Actually Decided – and Why

The legal significance of the BAG ruling becomes apparent through the mandatory representation requirement under § 11 para. 4 ArbGG. Before the BAG, parties must be represented by an authorised legal representative with rights of audience. The court has consistently clarified in its case law what this entails: any submission filed must reflect the intellectual work of the authorised representative. If a draft – regardless of who or what produced it – is adopted, the lawyer must independently review it and assume full responsibility for its content.

In the present case, the BAG inferred from the external circumstances that precisely this review had not taken place. Beyond the absence of standard law firm letterhead and the use of unusual phrasing, it was above all the hallucinated citations that proved decisive. The court found it "entirely inconceivable" that a lawyer who had actually read and reviewed the grounds of appeal would cite judgments that do not exist or that state the opposite of what was argued.

Decisions that do not exist, or that are misrepresented, speak against any review by counsel – so the BAG in its reasoning.

The authorised representative's attempt to remedy the lack of review after the fact by means of a declaration failed because it was submitted only after the deadline for filing the grounds of appeal had expired. Reinstatement to the previous position was therefore excluded.

Not an Isolated Case: The Growing Body of Case Law

The BAG does not stand alone in its ruling. In recent months, a discernible line has emerged in German case law that sharply sanctions the unreviewed use of AI in legal submissions – albeit with varying procedural consequences.

• Kammergericht Berlin (20.11.2025, Az. 17 WF 144/25): In an urgent family court proceeding, the KG found that the brief was obviously the product of a "hallucinating" AI and had been adopted without verification. The court indicated a violation of professional conduct rules – but stopped short of declaring the party unrepresented.
• Kammergericht Berlin (25.08.2025, Az. 3 ORbs 164/25): Criticism of unverified automated litigation containing false statements about the course of proceedings in a traffic law matter.
• Amtsgericht Köln (02.07.2025, Az. 312 F 130/25): The court ordered that submissions based on obviously AI-generated and fabricated sources must not be filed in the future – with reference to § 43a para. 3 BRAO.
• Landgericht Frankfurt a.M. (25.09.2025, Az. 2-13 S 56/24): Clarification that an attorney must be aware of the unreliability of AI-generated legal statements – and that the administration of justice suffers serious harm when courts can no longer trust citations.

These decisions paint a consistent picture: the problem is not the use of AI per se. The problem is the lack of verification of AI outputs.

The actual duty to verify – and what it means in practice

Dr. Maik Bunzel, founder and managing director of mabucon.eu, follows this development with particular interest, as it reveals a fundamental misunderstanding of what AI-driven automation can and cannot deliver. "AI systems are excellent research and drafting assistants. But they operate probabilistically, not factually. Anyone who ignores this and transfers outputs directly into binding documents is overlooking the critical human oversight step," is the assessment from the practice of AI implementation consulting.

From a legal standpoint, the duty to verify is already grounded in existing professional conduct rules: § 43 sentence 1 BRAO requires conscientious professional practice. The Federal Bar Association (BRAK) specifies in its guidance on the use of AI (as of December 2024) that all AI outputs must be reviewed on one's own responsibility. The German Bar Association (DAV) also makes clear in its initiative statement (No. 32/2025) that all cited sources must be verified in full – regardless of whether there are specific indications of hallucinations.

A structured verification workflow for citations should include at minimum the following steps:

• Identity check: Verify the court, type of decision, date, case number, and ECLI in the full text.
• Full-text access: No verification based solely on hit lists or AI-generated summaries.
• Source in context: Is the alleged statement actually contained therein? Does it reflect the court's view, a party's submission, or an obiter dictum?
• Statutory check: Section, subsection, sentence – and the applicable version at the relevant point in time.
• Documentation: Record the source, retrieval date, passage, and any open questions.
• No result found: Do not rely on a single database – consult additional sources, state portals, or contact the court directly.

Data sources such as Open Legal Data (with over 420,000 decisions and MCP access for AI applications) or the public federal API NeuRIS can be valuable tools in this regard. They enable existence and format checks – but do not replace the substantive review of the cited statement.

What AI automation in the legal sector can achieve – and what it cannot

The debate surrounding hallucinated court rulings tempts many to dismiss AI wholesale as an unsuitable tool for legal work. That would be a mistake. When deployed correctly – with clearly defined review processes and human sign-off at the critical junctures – AI can significantly enhance the quality of legal work while simultaneously reducing the effort required for research and drafting.

Dr. Maik Bunzel, founder and managing director of mabucon.eu, puts it succinctly: "The goal of modern workflow automation in the legal sector is not to replace the lawyer with the machine, but to deploy the machine for everything it can do reliably – and to keep the lawyer where judgment, responsibility, and liability matter." This delineation is not merely an ethical postulate but – as the BAG has now made clear – a procedural necessity.

In concrete terms, this means for law firm organisation: AI can handle research tasks, produce initial drafts, structure templates, and locate references. The final review – the proofreading, source verification, and substantive plausibility check – must remain with the responsible lawyer in a documented approval step. Anyone who skips this step due to time pressure, missing processes, or a lack of understanding of AI risks not only professional conduct violations but – as the BAG ruling demonstrates – the inadmissibility of the entire legal remedy.

Open questions: insurance coverage and liability

Beyond the procedural risks, the increasing use of AI in the legal profession raises another, as yet unresolved question: insurance coverage. The BRAK explicitly points out that beyond a certain degree of automation, it may become questionable whether what is being performed still constitutes legal activity in the insurance law sense. Should an insurer argue in a liability case that a pleading was not the result of legal work but of an autonomous AI system, coverage claims could lapse.

Binding regulations on the permissible degree of automation do not yet exist. This makes it all the more important to develop internal law firm policies now that draw clear boundaries between AI-assisted and AI-replaced work – and to document these internally.

Conclusion: AI as assistant, lawyer as the responsible party

The BAG ruling is a wake-up call – but not a prohibition. It clearly marks where the boundary between permissible AI support and impermissible abdication of responsibility lies: in the independent, documented review of every cited source. Law firms that invest now in structured AI governance – with clear workflows, defined review steps, and traceable documentation – will be able to harness AI as a genuine lever for quality. All others expose themselves to a risk that, as the BAG ruling shows, can lead directly to the inadmissibility of proceedings.

For companies with their own legal department, the same principle applies: AI-assisted legal documents forwarded without human oversight are not efficiency gains – they are liability traps. Implementing appropriate control processes is therefore not merely a matter of comfort, but of professional risk management.