Poland: the Supreme Administrative Court calls unverified AI output gross negligence

The three decisions

• 27 Aug 2026Supreme Administrative Court, I OSK 935/26, order of Judge Marek Stojanowski. A cassation appeal signed by an advocate cited a Civil Code article for a principle it does not contain, and a statute that does not exist in Polish law. The court found it more probable than not that the appeal was prepared with AI tools, and that its errors were “the result of so-called AI hallucination”. Appeal dismissed.
• 31 Aug 2026Supreme Administrative Court, I OSK 590/26, judgment of three judges. Much of the appeal rested on seven Supreme Administrative Court judgments: two were never registered at the court, and five carry the wrong date and concern unrelated subjects. The court said the appeal “was obviously prepared using AI tools” and left those citations out of account. The appeal succeeded in part, on other grounds.
• 3 Sep 2026Regional Administrative Court in Gorzów Wielkopolski, I SA/Go 210/26, judgment of three judges in a VAT case. The complaint, signed by an advocate, attributed to a Supreme Administrative Court judgment a passage that appears in no ruling in the administrative courts’ central database, and gave the judgment the wrong date. The court called it “a classic manifestation” of language-model confabulation. Complaint dismissed.

What the courts said

The Supreme Administrative Court accepted that AI tools are slowly becoming standard, probably unavoidably, in administrative court practice too. Its objection was to unreflective use. On the advocate who signed the appeal:

“The conduct of the author of the cassation appeal, who as the party’s professional representative did not take the trouble even cursorily to verify the content generated by artificial intelligence, and so allowed the pleading he signed to contain such evidently untrue sentences, must be assessed in terms of gross negligence.” „Postępowanie autora skargi kasacyjnej, który jako profesjonalny pełnomocnik strony nie zadał sobie trudu by choćby pobieżnie zweryfikować treści wytworzone przez sztuczną inteligencję i dopuścił w ten sposób aby pismo procesowe pod którym złożył podpis zawierało tak ewidentnie nieprawdziwe zdania należy natomiast oceniać w kategoriach rażącego niedbalstwa.”

It adopted, in full, a point the same court had made on 23 June 2026 in case I FZ 104/26: a professional representative is paid, and the client is entitled to expect the highest professionalism. Where the representative drafts with AI tools that are equally open to non-lawyers and does not even check the result,

“it becomes fair to ask what exactly the representative was paid for.” „zasadne staje się wręcz pytanie, za co dokładnie pełnomocnik otrzymał wynagrodzenie”

Four days later, a three-judge bench of the same court:

“a professional representative may not use AI tools unreflectively, omitting any verification of the content so produced.” „z narzędzi AI profesjonalny pełnomocnik nie możne korzystać w sposób bezrefleksyjny, z pominięciem jakiejkolwiek weryfikacji wytworzonych w ten sposób treści”

And the regional court in Gorzów Wielkopolski, on a complaint signed by an advocate who is paid for the work:

“Introducing false assertions and fictitious rulings into pleadings is not only a denial of professionalism and of the duty of due care, but also offends the dignity of the court and strikes at trust in a profession of public trust. Uncritical use of digital tools cannot replace real intellectual work and sound legal knowledge.” „Wprowadzenie do pism procesowych twierdzeń fałszywych oraz fikcyjnych orzeczeń stanowi nie tylko zaprzeczenie profesjonalizmu i powinności zachowania należytej staranności, lecz także uchybia powadze sądu i godzi w zaufanie do zawodu zaufania publicznego. Bezkrytyczne posługiwanie się narzędziami cyfrowymi nie może zastąpić rzeczywistej pracy intelektualnej i rzetelnej wiedzy prawniczej.”

Why Poland matters

There is no Polish practice direction to point to, no declaration to file and no fine. What there is, is a national supreme court putting the words “gross negligence” on a signature over unchecked AI output, twice in five days, following its own June decision to the same effect, in judgments anyone can read. The question it asks is the one every court on this site is converging on: who checked the work, and can they show it?

Sources

• Naczelny Sąd Administracyjny, I OSK 935/26, order of 27 August 2026, read in full
• Naczelny Sąd Administracyjny, I OSK 590/26, judgment of 31 August 2026, read in full
• Wojewódzki Sąd Administracyjny w Gorzowie Wielkopolskim, I SA/Go 210/26, judgment of 3 September 2026, read in full
• Each is a print of the decision as published in the administrative courts’ case-law database, from the AI Hallucination Cases document store. I FZ 104/26 of 23 June 2026 is quoted as the Supreme Administrative Court adopted it in I OSK 935/26.

Related

• AI rules for lawyers, country by country · Europe · England and Wales
• Verification is becoming a duty. Evidence of verification is the missing layer.
• The sanctions record

Read in full at the three decisions. Translations are ours. Information, not legal advice, and not advice on Polish law.

Common questions

Does Poland have rules on lawyers using AI?

Not as a court rule or practice direction. The guidance comes from decided cases: in August and September 2026 the Supreme Administrative Court and a regional administrative court found pleadings prepared with AI tools and not checked, and criticised the professional representatives who signed them.

What did the Supreme Administrative Court say about unverified AI output?

That a professional representative who did not even cursorily verify the content generated by AI, and signed a pleading containing evidently untrue sentences, must be assessed in terms of gross negligence (I OSK 935/26, 27 August 2026), and that a professional representative may not use AI tools unreflectively, without verifying what they produce (I OSK 590/26, 31 August 2026).

Were the lawyers fined?

No. None of the three decisions imposed a fine or a costs order for the AI errors. The courts set the false citations aside, decided the cases on the remaining arguments, and recorded their criticism in published judgments.