A verdict without law: how a machine's output becomes evidence

Three Russian courts in six months punished the parties for citing acts that do not exist and acts that were never in the case. In France a novel dropped out of the first Goncourt selection after a detector attributed a substantial part of the text to a machine — and plagiarism allegations surfaced there too. In Brussels the deadlines for the requirements were pushed to 2027–2028. In Washington the rules are being proposed by a company. Five stories, and they share one nerve: a machine's verdict becomes evidence before a norm appears to say who answers for it.

No one has asked a detector for a long time whether the instrument lies. What they ask is what to do with the verdict. Below — five stories about how that looks in court, in a literary prize, in a report on an attack and in the publication register.

The issue window is 15–30 October 2026. Not every date fits inside it: the literary story began on 25 September, the court story on 14 May, and the law in Brussels was adopted in June and July, while in Washington there are only drafts so far. Every event in the text carries its own date.

A verdict without law A verdict without law
A verdict without law

1. A detector attributed the novel to AI

The novel by the Haitian-Canadian writer Thélyson Orélien, “C'était ça ou mourir”, came out in France at the end of the summer, and by the end of September a scandal was already running around it: the AI-text detector Pangram attributed a significant part of the book to a machine. 25 September the Académie Goncourt took the novel off the first selection. The official wording, according to franceinfo, is not only suspicion of the use of AI but plagiarism allegations too. The same day the novel left the Femina selection.

The figure that went round the press turns out, on closer inspection, to be a different one. The “100%” figure was what anonymous accounts on X were citing, showing selective excerpts. Pangram's founder Max Spero ran the whole book and got 95% — 20 Minutes reported it citing Le Monde on 25 September. The difference between “an excerpt” and “the whole book” is no small thing here: it shows how a percentage torn out of its protocol takes on a life of its own.

21 October the publisher Boréal announced that it was ending its work with the author. The ground — traces of AI metadata in the documents, not in the text. In the same statement the publisher admits that to what extent AI was used, it still does not know. That matters more than the break itself: the decision was taken on the metadata, while the admission of uncertainty is on the substance.

And on 27 October the president of the Académie, Philippe Claudel, said on air that he had asked for sixteen books of the selection to be run through Pangram 4. We do not pass the request off as a completed run: what we have are his words, not a protocol. But the direction is visible — the detector does not stand against the literary institution, it is already inside it, as one more voice on the jury.

A similar story — with a caveat, and it matters. A novel by the Peruvian writer David Silva Rojas, “Donde el agua trabaja”, took the Almadía Ventosa-Arrufat prize (€20,000), and El País reported a score of 99%. The publisher cancelled the presentations, and the author denied using AI in a published statement — but the prize was left in place. In March 2024 the Mozilla Foundation published an analysis showing that detectors systematically over-score people for whom English is not a first language. The caveat is obligatory: in both stories the detector version is not named, so the percentages from there are reported, not measured by us.

Our own case comes with a protocol. 26 October 2026, Pangram 4.1 and the Russian service Textovod, three Russian texts, one day:

  • our letter to the appraiser — 185 words, submitted in full: Pangram gave 93% with the note that the ending is written by a human, Textovod gave 0% and “human”;
  • the appraiser's own letter — 99 words, pasted from a chat and signed by a human, the same typo twice: Pangram — “possibly human, 100%”;
  • the control letter was written by a model from start to finish, about 180 words, but what reached the interface was an excerpt of 73 words — Pangram gave “possibly AI, 100%” and added its own note that it classifies short text less reliably.

Hence the formulation we hold to: a percentage does not reproduce without the model version, the language and the length of the text. Our note on this came out on 26 October, and on 31 October we issued version 2 with a Correction section: the first version carried the length of the whole control letter, while what reached the interface was an excerpt. We published the correction rather than making it quietly.

2. The court looks at the signature

The Russian line is the fastest: three fines in six months. Two of them were for acts that do not exist at all, the third for twelve citations without the authorities themselves. These cases fall outside the issue window — this is a story from May to October — but it is exactly the one that sets the frame.

14 May 2026. The Commercial Court of the West Siberian Circuit, case А27-7831/2025. A claim in a dispute over payments — a debt of 50,000 ₽ and a penalty of 7,645 ₽ — cited the case law of the Supreme Court and of the Supreme Commercial Court. Some of the authorities came from other people's disputes, and some had never existed at all, quotations included. The company changed its representative; the new one asked the court to disregard the fact that his predecessor had “fabricated the source materials” — it did not help. A fine of 50,000 ₽ under part 5 of Article 119 of the Commercial Procedure Code. The debt and the fine coincide here by accident: they are different figures from different lines of the document.

27 August 2026. The Intellectual Property Rights Court, case С01-1025/2026 (at first instance А71-11377/2025): twelve citations without the authorities themselves. The ruling says that such conduct “can have no justification whatsoever”, and that preparing a claim with the help of AI is not among the justifications. The 5,000 ₽ fine is the maximum for an individual under part 1 of Article 119 of the Commercial Procedure Code; for organisations the ceiling is 100,000 ₽.

1 October 2026. The Commercial Court of the Vologda Region, case А13-14310/2025: four rulings, with genuine numbers, but no such acts exist. Behind the numbers in the case database sit a settlement agreement, membership fees and a sublease. A fine of 30,000 ₽; the part of Article 119 is not named in the ruling.

The formula is the same in all three: the one who signed pays — not the model and not the developer. The Supreme Court said as much back in 2018 in the Olkon case: the company answered for its representative's sharp words. And separately: on 21 May 2026 the Plenum of the Supreme Court adopted ruling No. 15 “On the preparation of civil cases for trial”, where paragraph 4 of clause 42 says: parties to a case who have submitted information about facts obtained with the use of artificial intelligence technologies must tell the court about it. The norm is procedural, civil; our three cases are commercial, and carrying it over there can be done only through practice. There is no sanction for silence in that clause — only an obligation.

Checking is cheap here: in the Vologda case it took minutes — four numbers, four queries in the case database. What costs is not the check but its absence.

3. The price of an attack: a hundred agents for $210

8 September 2026. Shrivu Shankar, vice-president for AI strategy at Abnormal AI, described an experiment: he launched a hundred autonomous agents against his own pet projects. Five hours, five successful breaches — three through vulnerabilities, two through weak passwords, sixteen social engineering attempts, not a single zero-day found. The run cost $210, a single account $40.

Caveats, without which this reads as a scare story rather than a fact. The models ran with their built-in restrictions lifted and were deployed by the author on his own infrastructure; the breaches ran against his own projects, not against somebody else's software; the prompt was deliberately primitive. In the popular Russian translation on Habr (article 1097872) there are two shifts against the original: “fairly AI-generated” became “clearly”, “$40” became “about $40”, and the headline about “five times” was added by the translator.

What these figures do not mean. They do not imply that attacking is cheaper than defending: there is no figure for defence in this issue, and there is nothing to compare it with. What they do imply is something else — a check has a price, and it is time, not budget. When the cost of the error is higher than the cost of the check, the protocol decides, not intuition.

4. Russia: secondary legislation for Law No. 243-FZ — zero adopted

26 July 2026 Law No. 243-FZ was adopted. Part of it came into force on 1 September 2026, but the key provisions — items 3–5 of part 2 of Article 5, parts 2–5 of Article 6, Articles 8–10 — work only from 1 March 2027, and the transition period runs until 1 September 2032.

Those provisions need secondary legislation. We looked at the official publication register for 15–30 October: the “Government” section carried 208 acts — 76 rulings and 132 orders. Not one of them with artificial intelligence in the title. The top entry in the section as of 31 October is ruling No. 1430 of 30 October; nothing has been published for 31 October.

9 October Deputy Minister of Digital Development Iosif Lebedev said at a forum: “There is an approved schedule for it… six rulings.” We found neither an official list nor the schedule itself. A draft amendment to ruling No. 418, which assigns the role of AI regulator to the Ministry of Digital Development (about ₽3bn), has been hanging in consultation in its fourth version since 14 September; the “from 1 October” announcement comes from the press, there is no such date in the norm. FSTEC has no order with AI requirements, and none is scheduled in the plan of regulations for 2026. FSTEC order No. 117 and its amendments by order No. 137 are not acts under Law No. 243-FZ — they rest on a different basis.

A document and a remark appeared over the fortnight that are worth noting. 30 October at 17:30 Mikhail Mishustin approved the allocation of duties among deputy prime ministers: state policy on the introduction and use of AI technologies is assigned to Dmitry Chernyshenko, with the Ministry of Digital Development inside his remit. 29 October senator Artyom Sheykin said at the “Derzhavin Readings” that the secondary legislation “must be prepared promptly”.

A caveat, without which the figure “zero” is read wrongly. Zero is acts that have AI in the title, in the publication register. A full-text search does find the words “artificial intelligence” in other 2026 rulings — but those are programmes of experimental legal regimes under Law No. 258-FZ and amendments to general acts, not the development of Law No. 243-FZ. And one more thing: the word “register” is not in the law, part 5 of Article 6 says “the procedure for recording and assigning status” — which is why “a register of models” in retellings is imprecise.

5. The EU has moved the deadlines; the US has only drafts so far

The AI liability directive has not come back in Europe: draft COM(2022)0496, procedure 2022/0303(COD), withdrawn on 6 October 2025, and the Official Journal (OJ C/2025/5423) says so outright, by document number. No official steps have followed the withdrawal. 22 October the European Parliament held a debate on the item “strengthening global AI governance” — Commission statement 2026/2935(RSP); the directive was discussed inside it. The Socialists and Democrats and the Greens called for work on it to resume, and there was no resolution at the end. We could not confirm the position of the EPP group, so it is not in the text.

What does work in the EU is the Digital Omnibus (COM(2025)0836, procedure 2025/0359(COD)): the European Parliament passed it on 16 June 2026 by 423 votes to 57 with 174 abstentions, the Council agreed on 29 June, the act was published on 24 July and became Regulation (EU) 2026/1744, in force since 27 July. The obligations for systems in Annex III have been moved to 2 December 2027, those in Annex I to 2 August 2028. Systems placed on the market before 2 August 2026 must label synthetic content by 2 December 2026. There is no liability in that act: it is about access to the market and deadlines.

In the US the law is not written yet — but the text for it already exists. In June 2026 Anthropic published an “Advanced AI Framework”: it is company policy, not legislation. A draft bill on the same subject has been going round Congress since 19 October; officially none has been introduced. The documents are easy to mix up, and their thresholds differ: the June framework has research spending above $1bn a year and a compute threshold, the draft bill has losses above $1bn or 50 dead and seriously injured. Next to it sits the FRONTIER Act, pushed by Lori Trahan and Jay Obernolte: according to Business Insider, its substance is federal oversight of frontier models instead of a patchwork of state law. We could not confirm its number or the date it was introduced — hence they are not here.

And the backdrop that explains the story better than any threshold: on 26 October Anthropic reported filing registration form S-1 with the Securities and Exchange Commission — the company is preparing to go public. The rules for the industry are proposed by the one that will soon be reporting to the market itself.

The date of the US election is 3 November 2026. We have not checked the House and Senate calendars, so “after the election” here is about the agenda, not about fact.

6. What follows from this

Five stories come down to one distinction. The question “is the instrument right” is neither legal nor technical but a question of protocol: what was submitted, which version, what length, which date. The court looks at the signature. The prize academy asks for its own selection to be run through the detector — and says so on air. The Ministry of Digital Development speaks of six rulings, while the register holds zero. These are different answers to one question — what counts as evidence, and who answers for it.

Three conclusions we are taking into our work:

  1. Disclosing the use of AI is an obligation, not a courtesy. Paragraph 42 of Plenum ruling No. 15 requires parties to speak about information obtained with the help of AI — at least in civil procedure.
  2. A percentage always comes with conditions: the detector version, the language, the length of the text, the date of the run — and, if the figure is ours, a reference to the protocol as well.
  3. Checking takes minutes; not checking costs 30,000 ₽. Four queries in the case database — that is the whole price of the Vologda case for the court.

7. The numbers in this issue, and where to check them

Caveats for this issue. The Pangram and Textovod figures on our case are confirmed by screenshots of 26.10.2026. The results for Orélien and for Silva Rojas come from publications; in the Peruvian story the detector version is not named. We found no written communiqué from the Académie Goncourt, only retellings in the press. Zero acts devoted to AI is a negative sample by title in the publication register: it says that no such act is in the register, not that it will never be there. The filter parameters on the publication portal are not documented, so a recount requires manual selection by date.

Page updated 31 October 2026 · site build 2026.11.13 · ANO «ISAIS»