Synthetic voice · AI-generated (text-to-speech).
By mid-August 2026 the Arbeitsgericht Berlin (Berlin labour court) had recorded some 16,000 cases — as many as in all of 2024. The full-year figure for 2025 was 18,600, and 2026 will reach it considerably earlier. Reporting has given the phenomenon a name: agentic flooding — statements of claim produced en masse with chatbots.
More interesting than the number is the mechanism: what changed is not the law but the price of invoking it — and the cost of the resulting volume lands on the court and on the opposing party, not on the party producing it. This is not cultural criticism but a change in price with a predictable consequence, and it hits companies twice: as recipients of such letters, and as clients of a law firm whose opponent works this way.
At a glance
- What is documented: Arbeitsgericht Berlin some 16,000 cases up to mid-August 2026 (2025: 18,600); Germany’s Sozialgerichte (social courts) in 2025: 263,500 actions (+10 %) and 40,000 Eilverfahren (urgent interim relief proceedings, +47 %).
- What follows: The effort does not disappear, it moves — from the party producing it to the party checking it. For Austria, the additional need is estimated at roughly 250 extra judicial posts.
- The catch: Total caseloads are not AI figures. The filings identified as AI-generated in Berlin number 51 — a projection to the end of September 2026.
The numbers, as far as they go
The labour and social courts are at the centre of this, because proceedings there can be conducted without a lawyer.
| Area | Figure | As of |
|---|---|---|
| Arbeitsgericht Berlin, all cases | some 16,000 (2025: 18,600) | mid-August 2026 |
| Arbeitsgericht Berlin, identified as AI-generated | 51 filings — projection | to end of September 2026 |
| Sozialgerichte nationwide, actions | 263,500 (+10 % on 2024) | year 2025 |
| Social courts nationwide, Eilverfahren | 40,000 (+47 % on 26,995) | year 2025 |
| Bavarian social courts, first instance | +26 %; Bürgergeld and Grundsicherung (basic income support) +60 %, urgent interim relief more than doubled | September 2026 |
| Versicherungsombudsmann (insurance ombudsman), complaints | around +50 % | January to May 2026 |
The most important qualification belongs up front, not in the small print: the 16,000 and the 18,600 are total caseloads, not AI figures. In Berlin the bulk are unfair-dismissal cases, and AI as a contributing cause is a plausible but unquantified assumption voiced by courts and press. What has been identified as AI-generated is the smaller subset — the 51 mentioned above.
What does come solidly from the courts is the description of quality. Christine Fuchsloch, President of the Bundessozialgericht (Federal Social Court): “The pleadings are often too long, lose themselves in remote considerations and cite fabricated, unfitting or long-superseded court decisions.” Edith Mente, President of the Bavarian Regional Social Court: “The AI evidently used by many claimants does not really know social law well and therefore often recommends nonsensical applications.”
Why this is mechanics, not cultural criticism
First: what fell is not the court fee but the cost of drafting. For many people the barrier was never the law but the wording — twenty pages in an unfamiliar technical language. That effort has dropped from hours to minutes; with any other good, one would expect a volume response.
Second: the effort moves, it does not disappear. Checking is systematically more expensive than asserting: a pleading with 30 references is written in minutes and verified in hours. That holds for the court, for the opposing party, and for that party’s law firm.
Third: part of the volume is not merely volume but wrong. If a model recommends nonsensical applications in social law, proceedings arise that cannot succeed from the outset — more cases at a falling hit rate.
The counter-position belongs here: some of these proceedings are justified. Anyone who gave up an existing claim because of language or uncertainty can now assert it — that is access to justice, not abuse. The load on the courts is therefore not a moral argument against claimants but a capacity problem — with the side effect Fuchsloch names: such pleadings also “create false expectations.”
Too much material in front of unchanged capacity — two sheets made it through the slot. And one in the stack is the one whose citations do not exist.
The second strand: citations that do not exist
Alongside volume sits a quality problem that is easier to count. Damien Charlotin maintains an international database of AI citation errors in court proceedings; as reported on 16 June 2026 it held around 1,600 documented cases from 35 countries, 1,116 of them from the United States — with an eightfold annual increase reported. No separate figure for Germany is listed there.
For the German-speaking region the individual cases carry the argument. In Austria, an application for legal aid grew into a pleading of roughly 60 pages containing 30 references to decisions of the Oberster Gerichtshof (Austrian Supreme Court) — all of them fabricated; another case, from Salzburg, ran to 400 pages of AI-generated text. The Supreme Court dismissed a Nichtigkeitsbeschwerde (plea of nullity) containing fabricated references on 7 October 2025 (14 Os 95/25i), pointing to counsel’s duty of loyalty under § 9 RAO, the Austrian Lawyers’ Act.
The most-cited German reference case is older than the debate. In an order of 2 July 2025 (case no. 312 F 130/25) the Amtsgericht Köln (Cologne local court) dealt with a lawyer’s pleading whose references were fabricated throughout from page 8 onwards — a commentary passage attributed to the wrong author, journal articles that do not exist. The court found a possible breach of § 43a(3) BRAO and hinted at attempted procedural fraud; no sanction outcome is documented.
For a litigator’s duty of care, less of this is new than the fuss suggests, and more of it is practical. The duty is not new: whoever cites vouches for the reference existing and supporting the point. What is new is a tool that has perfect command of the form of a correct citation and no information about its content — a fabricated case number looks exactly like a real one. Asking the model does not help; it answers in the same mode that produced the error. The only thing that works is a control layer outside the model: look up every reference in the primary source.
What this means for your company
Treat serial cases as a series. When uniform, machine-drafted claims arrive in waves, handling each one individually is the most expensive possible answer. What works: work through the pattern once on the merits, build a legally reviewed standard response, examine the rest only for deviations — and automate the triage, not the assessment. The mistake to avoid is confusing form with substance.
Look up every reference the other side cites. This is currently the cheapest defensive measure there is — a pleading whose references stop existing from a certain page loses its credibility as a whole. It is not a strategy, though: the other side’s error is a gift, not a plan. For law firms on this side of the table, the tooling question is covered in legal tech and AI for law firms.
Sort out your own side first. If AI helps draft your correspondence, you need one non-negotiable rule: whoever cites, checks — and documents that they checked. That documentation is what you can put on the table if it is ever disputed.
Conclusion
The figures describe no new legal position but a new volume position — documented several times over, only partly measurable in its cause. A tool has lowered an access barrier that was too high for some, and in doing so shifted a load for which nobody budgeted capacity. The consequence is uncomfortable but simple: checking remains more expensive than asserting, so checking has to get cheaper — through standardisation on the volume side and a hard rule on the citation side.
If you have a series of near-identical claims on your desk, or need a robust checking routine, let us talk. I read cases like these as a commercial lawyer and build the systems meant to help with them myself.
FAQ
Were all 16,000 cases at the Arbeitsgericht Berlin generated by AI?
No. The roughly 16,000 cases up to mid-August 2026 are the court’s total caseload (2025: 18,600). Attributing the rise to AI is a suspected contributing cause, not a measured quantity. The filings specifically identified as AI-generated in Berlin number 51 — and that figure is a projection to the end of September 2026.
What should a company do when it receives a series of near-identical AI-drafted claims?
Treat them as a series: work through the pattern once on the merits, build a legally reviewed standard response, and examine each individual case only for where it deviates. Uniform form does not mean uniformly unfounded, however — dismiss them wholesale and you lose exactly the cases with substance. And your own reply must not carry the same source of error.
May a lawyer use AI to draft pleadings?
Using it is not the problem; adopting citations without checking them is. In an order of 2 July 2025 (case no. 312 F 130/25) the Amtsgericht Köln found that a pleading whose references were fabricated throughout from page 8 onwards may breach § 43a(3) BRAO, the German Federal Lawyers’ Act, and even hinted at attempted procedural fraud. What is new is not the duty of care, but a tool that produces formally flawless citations for sources that do not exist.
How do I spot a hallucinated citation?
Not by its format, because the format is usually correct. The only reliable route is looking up every reference in the primary source — not asking the model that produced it. The tell is clustering: in one Austrian case all 30 references to Supreme Court decisions were fabricated.
Sources — as of 16/09/2026
- ad-hoc-news.de (18/08/2026) — https://www.ad-hoc-news.de/wirtschaft/ki-flutung-der-justiz-arbeitsgericht-berlin-verzeichnet-16-000-verfahren/69964602
- kosmo.at (19/08/2026) — https://www.kosmo.at/oesterreichs-justiz-am-limit-immer-mehr-lassen-schriftsaetze-von-ki-erstellen/
- taz.de (20/08/2026) — https://taz.de/Kuenstliche-Intelligenz-vor-Gericht/!6205827/
- anwaltspraxis-magazin.de (12/09/2026) — https://anwaltspraxis-magazin.de/schlagzeilen/2026/09/12/rasanter-anstieg-der-fallzahlen-bei-arbeits-und-sozialgerichten/
- borncity.com (16/06/2026), database maintained by Damien Charlotin — https://borncity.com/news/gefaelschte-urteile-in-klageschriften-1-600-ki-fehler-in-35-laendern-dokumentiert/
- legal-tech-verzeichnis.de, Amtsgericht Köln, case no. 312 F 130/25 — https://legal-tech-verzeichnis.de/legal-tech-nachrichten/anwalt-blamiert-sich-mit-halluziniertem-ki-schriftsatz-gericht-spricht-von-verstoss-gegen-brao/
This article is general information and not legal advice in an individual case. As of 16 September 2026; the court figures keep moving.
Leon Lotz
Leon Lotz is a business lawyer and founder of MusketierSoftware. He combines legal depth with real software craft.
