₽50,000 isn’t a huge sum for an arbitration dispute. But that’s exactly what one unverified citation — generated by a neural network and dropped into an official court filing — ended up costing. AI in court is increasingly turning from a helper into a liability: case № А27-7831/2025 isn’t interesting because of the fine amount. It’s interesting because of the exact moment nobody stopped a filing that looked perfectly reliable.
What happened
A Kemerovo-based company, Tsentralny Spets Servis, filed a cassation appeal with the Arbitration Court of the West Siberian District, citing rulings from Russia’s Supreme Court and former Supreme Arbitration Court. On review, some of those rulings turned out not to exist at all, and others had been issued in completely unrelated disputes. In its decision of May 14, 2026, the court treated this as contempt of court and fined the company under Art. 119(5) of the Arbitration Procedure Code (APC RF).
We don’t know who actually drafted the appeal or exactly how the AI ended up in the document’s production chain. But the fact that obviously fabricated citations reached the court, unfiltered, is a familiar pattern. It happens wherever the decision to “trust the AI draft” isn’t checked as a separate step — and instead just rests with whoever wrote the document.
Where the check should have happened
Any legal document an AI helps draft passes through a few checkpoints where it can still be stopped: draft → internal review → filing. The fine in the “TsSS” case wasn’t one person’s mistake. The real issue is that the process itself never had a mandatory review.
The difference between “AI helped write the appeal” and “AI got the company fined” isn’t about the model or the prompt. What matters is whether the product has a separate check built into the process that runs automatically — or whether it exists only as someone’s goodwill.
What a Red Flag looks like here
In legal architecture terms, the rule is simple: find a citation to a court ruling in a document — check it against an open registry, like the arbitration case index. No such case in the registry, and the document doesn’t move forward until a lawyer checks the citation by hand. That’s a Red Flag — a hard stop that halts automation exactly where the cost of a mistake is high. In practice: the system checks the case number against that registry itself, and if it isn’t found, export is blocked until a human reviews it.
The hard part isn’t the idea of checking. Without architecture, this step is easy to skip: it isn’t built into the process — it just relies on someone remembering to check by hand.
It’s not just about courts
A cassation appeal isn’t the only place where an AI hallucination can cost a business dearly. The same mistake can happen anywhere:
- in a contract — if AI cites a clause that isn’t actually in the text;
- in counterparty due diligence — if AI says a check passed when it never actually ran;
- in an investor report — if the conclusions rest on facts nobody verified.
Wherever AI states something with confidence instead of admitting “I don’t know,” there’s a risk the answer turns out to be just as fabricated as the nonexistent court rulings in the TsSS case.
For a business building a product on AI, what matters isn’t whether to use a neural network. What matters is whether its architecture has Triage Logic — a system that decides in advance which AI answers can go out immediately, and which need a human check first. For example: a plain summary of a case’s facts can go out right away, but any reference to a specific detail — a case number, a date, a statute — gets automatically flagged as “needs verification” and doesn’t move forward without a human sign-off.
Checklist: 3 checks before an AI draft becomes a document
- Every reference is extracted and checked against a registry before human proofreading. Each case number, date, and statute citation is matched against the primary source in an open database — not against another AI draft’s summary of it.
- AI’s autonomy has defined limits. There’s an explicit list of statement types — any citation of case law, any reference to a statute, any counterparty status claim — that don’t get finalized without a “reviewed by counsel” stamp.
- One specific person owns the final check, not a vague “someone will look at it.” Until that role has a name, it’s most likely not happening at all.
Frequently asked questions
Is it even legal to use AI when preparing court filings? Yes — the law doesn’t ban using AI or automation tools to prepare documents. The court evaluates the content and accuracy of the document itself, not the tool used to draft it. Whoever signs and files it bears responsibility — that follows directly from Art. 9(2) of the APC RF: “parties to a case bear the risk of the consequences of taking or not taking procedural actions.” In plain terms: whoever filed it with the court is on the hook, not the tool they used.
How is this different from the general article on AI regulation in LegalTech? The article on AI law is about what already regulates AI use in LegalTech and FinTech broadly — which laws and regulators apply to these products. This article is about something different: what specifically went wrong inside the document-preparation process in the TsSS case, and how to build a process that doesn’t repeat it.
Could this kind of mistake happen outside a courtroom, in an ordinary B2B service? Yes. If a service generates answers, summarizes contracts, or produces reports without built-in Red Flags and Triage Logic, the risk of hallucinations stays high — from wrong deadlines in a contract to invented deal terms. In court, a fabricated citation can at least theoretically get caught by the opposing side; in a closed B2B product, if the client doesn’t double-check the answer themselves, the mistake can go unnoticed for much longer.
Where should we start if our company already uses AI in legal work and Red Flags and Triage Logic aren’t defined yet? Start with an audit of the existing logic — the Audit & Second Opinion format: a rapid review of what’s already running, pointing to the specific spots that need a hard stop.
Want to check whether your own process has gaps like this? See how legal architecture works — the formats, including an independent audit.