Short answer: a judgment of an Italian court in Oristano contained traces of the judge’s conversation with a chatbot, and the Court of Appeal in Cagliari held that this alone does not prove the judge handed the assessment of evidence, the interpretation of the law or the decision over to a machine. Under Italian law (Law 132/2025, Art. 15), these three things belong exclusively to the judge, and AI may at most support them. An important caveat: the Cagliari ruling was made at the stage of an application to suspend enforcement, so it is not a final resolution of the dispute, and some Polish reports oversimplify this.
Below we separate what the sources confirm from our own analysis and forecasts, which we clearly mark. If you’re interested in the general rule “AI makes things up, so check”, see our guide AI as a search engine. Here we ask something more serious: what if such a tool is in the hands of the person who delivers the judgment.
What happened: the facts and discrepancies in reports
Interia (citing the daily “Il Messaggero”) described a civil case from Sardinia in which the reasoning of a judgment contained a fragment of the judge’s dialogue with a chatbot. Italian commentaries indicate that the assistant’s reply began with the word “Perfetto” and announced that it would prepare a paragraph of reasoning in the style of a judgment. The losing company, according to those commentaries in a dispute over the return of a deposit and damages, argued on appeal that the judgment was non-existent or void because AI had been used.
| Element | Interia / WP | Italian legal commentary |
|---|---|---|
| Court of first instance | Justice of the peace in Oristano | Salvis Juribus: judgment of the Tribunale di Oristano no. 178/2026. Discrepancy: we don’t decide which version is correct |
| Court of appeal | Court of Appeal in Cagliari | Corte d’appello di Cagliari, ordinanza no. 182 of July 23, 2026 |
| Nature of the ruling | “Upheld the judgment” | Cautelare phase: application to suspend enforcement (Arts. 283 and 351 of the Italian Code of Civil Procedure), without ruling on the merits. The suspension was granted only in part, for reasons other than AI |
| The AI objection | The court rejected it | The first objection was found “undoubtedly unfounded” at this stage, because a single fragment does not prove delegation of the decision |
| Next steps | Appeal to the Supreme Court | We did not find this in the Italian commentaries we read |
| Names, case numbers | None | Ruling numbers are in the commentaries; no names |
Italian law: Law 132/2025 and Art. 15
Italy became the first EU country to adopt its own comprehensive AI law after the EU AI Act: Law no. 132 of September 23, 2025 (as described in the Ministry of Justice’s reply to a parliamentary question). Its Art. 15 concerns the justice system.
| Element | What the sources say |
|---|---|
| Art. 15(1) of Law 132/2025 | The interpretation and application of the law, the assessment of facts and evidence, and the issuing of decisions are reserved to the judge. AI may only be support and assistance (Ministry of Justice, Salvis Juribus) |
| Judicial “giustizia predittiva” | According to CMS’s commentary, predicting the outcome of a dispute based on statistical models is prohibited by default |
| Role of the ministry | Defines permitted uses and may authorise trials in courts after an opinion from AgID and ACN (CMS). A Permanent Observatory on AI was established on July 10, 2024 |
| The 2026 experiment | From January 1, 2026, a thousand Copilot 365 licences for volunteer career judges, with an emphasis on training and oversight (CMS) |
| CSM recommendations (October 8, 2025) | Legality, transparency, proportionality, judicial autonomy; critical assessment of outputs, which the judge may reject; data sovereignty, human oversight, training |
| UNESCO (December 2025) | 15 non-binding principles, including a ban on automating adjudication and transparency towards the parties |
What follows: Italian law does not ban AI in courts. It sets a functional boundary: it may help, but it may not judge in the judge’s place. The author of the Salvis Juribus commentary stresses that what matters is the function the output serves in practice, not the label “draft”.
Case law and examples: Italy and the world
| Case | What happened | Takeaway |
|---|---|---|
| Corte d’appello Cagliari, ord. 182/2026 | Trace of an AI conversation in the Oristano judgment; delegation objection rejected at the cautelare stage | A trace of AI doesn’t prove delegation of the decision; the burden of proof is on the party |
| Corte di Giustizia Tributaria, Palermo, judgment 981/2026 | The reasoning describes editorial tasks: removing repetitions, organising connections, final editing (EC News commentary of 7.10.2026) | The line between editorial help and shaping the reasoning is not obvious |
| Corte d’appello Torino (2024 judgment), Cassazione 2025 | Citation of non-existent rulings; the Cassazione set aside the judgment in that part. The minister announced a review and, in case of gross negligence, disciplinary action (reply to a parliamentary question) | The judge is responsible for citations |
| Cassazione penale, sez. III, no. 14631/2024 | Defence counsel relied on a ChatGPT answer; the court mentioned it and rejected the argument | A chatbot’s answer is not evidence |
| Tribunale di Firenze, ordinanza of March 14, 2025 | Non-existent rulings from ChatGPT in a filing; the court did not apply Art. 96 of the Code of Civil Procedure, as no damage or bad faith was shown, but found the lack of verification reprehensible | A lawyer is responsible for the work of associates and tools |
| Colombia, judgment T-323/2024 | A judge used ChatGPT 3.5 to expand the argumentation; the Court held that AI cannot replace a judge’s reasoning and ordered guidelines | Arguments should precede the decision, not justify it after the fact |
| USA, judges Wingate and Neals (2025) | Rulings with errors were withdrawn because assistants used AI (Perplexity, ChatGPT) and drafts entered the record before verification | A failure of the verification procedure, not just of the tool |
| United Kingdom, guidance of April 14, 2025 | The judge is responsible for everything issued in their name; AI for summaries and administrative tasks, not for legal analysis with public chatbots | AI allowed, but with control and confidentiality |
Legal assessment: is this a violation?
Below is our analysis, based on the sources described above. We are not a law firm.
- Is the mere use of AI prohibited? Under Art. 15(1), no. What is prohibited is delegating interpretation, the assessment of evidence and the decision. Editing, proofreading or suggesting wording can be support.
- Who has to prove delegation? According to the commentary on the ruling, the party relying on it. A trace alone is not enough; additional elements are needed: instructions given to the system, drafts, discrepancies between the case file and the reasoning. The court does not shift the burden of proof onto the other side, nor does it presume that the judge failed to do the work.
- Does a judge have to disclose the use of AI? The sources we read contain no such general obligation. The UK guidance says explicitly that a judge doesn’t have to describe the research preparation for a judgment. That’s a gap: the party has no way to check how the reasoning was produced.
- AI Act: systems used by judicial authorities to assist in researching and interpreting facts and law are high-risk according to Recital 61, and “final decision-making must remain a human-driven activity”. Purely ancillary tasks, such as anonymisation or administrative tasks, are excluded. The Salvis Juribus commentary notes that the Cagliari ordinanza does not address the AI Act or data protection.
- The duty to give reasons. Our note, beyond the sources discussed: the Italian Constitution (Art. 111) requires judgments to be reasoned, so the question of who authored the reasoning is a question about the integrity of the reasoning. The same applies in Poland, where guarantees include Art. 45 (right to a fair trial) and Art. 178 of the Polish Constitution (judicial independence), and in most European legal systems.
Political dimension: trust, efficiency and sovereignty
This is our interpretation, based on facts from the cited sources.
- Efficiency pressure. Courts in many countries are drowning in cases, and AI tempts with the promise of faster reasoning. The bigger the backlog, the greater the temptation of a “shortcut”, especially when nobody checks how the text was produced.
- Trust in courts. The Oristano dispute shows that even the appearance of delegation undermines the authority of a judgment, even if a higher court doesn’t set it aside. For the legitimacy of courts, what matters is not only that the judge decided, but also that the parties can believe it.
- Own tools instead of someone else’s. The CSM recommendations emphasise data sovereignty. Pasting case files into a public chatbot means handing data to an entity outside the court system. Hence the pressure for in-house solutions.
- Political oversight. The case of the US judges prompted questions from senators, and in Italy a parliamentary question and a reply from the minister. AI in courts quickly becomes a topic in relations between the executive and the judiciary, which calls for caution, because oversight of technology must not become oversight of the content of judgments.
- Bias in models. The UK guidance warns that models trained mainly on US law may carry its perspective over into other systems.
Social dimension: parties, citizens and asymmetry
- Information asymmetry. A party cannot see the judge’s conversation with a chatbot unless it ends up in the file. In Oristano, it came to light by accident, through a trace in the text.
- Chatbots on both sides. The UK guidance notes that more and more people without a lawyer use chatbots, which increases the risk of fabricated citations. A court that itself trusts a chatbot will find it harder to catch them.
- Access to justice. On the other hand, the tools can help: simpler reasoning, translations, transcripts of hearings. The risk is only that the benefits go to those who can afford quality control.
- Accountability. A chatbot won’t be disciplined and won’t apologise. Someone has to be accountable, which is why all guidelines point to the judge.
Is such behaviour ethical?
Our answer: it depends on what the judge did with the output and whether they were fair to the parties, not on the mere fact of using AI. We propose four check questions:
| Question | Unethical when… | Acceptable when… |
|---|---|---|
| 1. Who decides? | The model proposes the decision, and the judge copies it | The judge decides alone; the model helps with editing |
| 2. Was the output checked? | Quotes and facts go into the judgment without verification | The judge verified every claim in the source |
| 3. Is the data safe? | Case files were pasted into a public tool | The tool is approved and protects confidentiality |
| 4. Do the parties know? | Use was hidden when it affected the content | Use is disclosed or limited and technical in nature |
Consistent with this, the CEPEJ European Ethical Charter (2018) lists, among others, respect for fundamental rights, transparency and user control. In the Oristano case, if the reports are complete, the failure would lie not so much in “asking a chatbot” as in a possible lack of transparency and control, and that is exactly what the Cagliari ordinanza did not decide, because it concerned another issue.
What lies ahead for law and courts? Three scenarios
This is speculation, not a forecast. We describe three possible directions based on current trends.
| Scenario | What it looks like | Risk |
|---|---|---|
| A. Controlled assistant | Approved tools, training, disclosed use, audits. AI summarises, transcribes, organises | Slow rollouts, dependence on vendors |
| B. Drift and “shadow AI” | Judges use public chatbots secretly because official tools are lacking | Leaks, errors, further scandals |
| C. Limited automation | Simple mass cases (e.g. small claims) decided semi-automatically, with the right to appeal to a human | Erosion of the right to a judge, data bias |
What might become normal in a few years?
- Probably normal: transcription of hearings, case file summaries, case law search with verification, language editing, translations, anonymisation.
- Probably disclosed: a note that an AI tool was used in preparing the reasoning (today, in the sources we saw, there is no such obligation).
- Unlikely to become normal: copying the decision from a model without one’s own assessment, using public chatbots for confidential files, hiding use.
History offers an analogy here: legal search engines, word processors and translators were once viewed with suspicion and are now everyday tools. The difference is that generative models can produce reasoning, not just find or record text. That’s why the “assistance versus delegation” line will probably remain, though it will shift.
How to read news like this? Check it yourself
Check whether it’s a ruling or a commentary
Establish what kind of act it is: a judgment, a decision, an ordinanza cautelare. The stage of proceedings changes the weight of the conclusions.
Compare the headline with a legal commentary
Media often simplify. Look for analyses on legal websites and, if possible, the text of the ruling itself.
Separate fact from interpretation
“A trace of AI was found” is a fact. “The judge handed the decision to a machine” is an interpretation that has to be proven.
Check the court and case number
The court’s name and case number let you find the source. Discrepancies (e.g. justice of the peace versus tribunal) are a signal to be careful.
Verify citations from AI
If you use a chatbot for legal questions yourself, check every provision and ruling in an official source. See also our guide: AI as a search engine.
Follow whether the case is final
Proceedings can take time. Go back to the sources after months, because the assessment often changes.
More on checking information: step-by-step fact-checking.
FAQ: the most common questions
A trace of a conversation with an AI assistant was found in a judgment from Oristano. Counsel argued that the judge had handed the decision to a machine. According to Italian commentaries, the Corte d’appello in Cagliari (ord. 182/2026) did not consider this proof of delegation at the cautelare stage. Media reported different details, which is why we show the discrepancies in the table above.
Not in that sense. The ordinanza is not general permission for AI in adjudication. It dismissed the objection because a trace alone doesn’t prove the judge assigned the decision to a machine. The case is not closed.
Yes, but as support. Art. 15 of Law 132/2025 reserves to the judge the interpretation of the law, the assessment of facts and evidence, and the issuing of decisions.
In the sources we checked, there is no general obligation. The UK guidance doesn’t require a description of research preparation. This is one of the gaps lawyers are debating.
According to Recital 61, AI systems assisting courts in researching facts and law are high-risk, and final decisions are to remain a human activity. Purely ancillary tasks, such as anonymisation, are excluded.
In the sources we checked, we found no separate law on AI in adjudication. The Ministry of Justice is running pilots, and parliamentary question no. 18057 asks about implementing AI in courts. Guarantees come from the Constitution and general rules on judicial independence.
The Italian Supreme Court (Cassazione penale, sez. III, no. 14631/2024) noted defence counsel’s reliance on a ChatGPT answer and rejected the argument. A chatbot is neither a source of law nor evidence.
They are citations of non-existent judgments that a model generates because it predicts plausible-sounding text rather than checking a database. Such cases have been recorded in Florence, Turin and the USA.
The Italian minister, in reply to a parliamentary question, pointed to the possibility of disciplinary action in case of gross negligence, e.g. citing non-existent rulings. Using the tool itself is not misconduct.
In our view, what matters is: who decides, whether the output was verified, whether the data was safe and whether the parties know. Hiding use and failing to verify are a problem; editorial help on its own is not necessarily.
Probably yes for supporting tasks: transcription, summaries, search. This is our speculation. Delegating judgement and hiding use will most likely remain unacceptable.
Current legal frameworks in the EU and worldwide do not provide for this. UNESCO and CEPEJ guidelines, as well as the Italian law, point to human oversight. The Colombian court held that AI cannot replace a judge’s reasoning.
As a starting point, not as a source. Check every provision and judgment in official databases. Don’t paste confidential data into public tools.
Compare several sources, including legal commentaries, and check the case number. Discrepancies between media outlets are a signal to be careful.
Get the “How to Search?” guide
Practical methods for searching and verifying information, step by step, with dozens of examples.
See the guide → ☕Support us on Buy Me a Coffee
We create these materials in our spare time. Every bit of support helps us write more, and more often.
Support → ▶️Watch us on YouTube
The same topics in a shorter video format, with concrete examples on screen.
Subscribe →More from our blog
Sources: Interia: Italian judge pasted an AI dialogue into a judgment; Wiadomości WP: the court in Cagliari; Salvis Juribus: AI, sentenza, delega della decisione; Mondo Professionisti: L’IA non manda in fumo la sentenza; CMS: AI e magistrati, limiti, rischi e usi consentiti; Ministero della Giustizia: reply to a parliamentary question (15.01.2026); EC News: Corte di Giustizia Tributaria di Palermo; Studio Cataldi: the Florence case; Leggezero: la prima volta di ChatGPT; Cybersecurity360: sentenze fantasma in Cassazione; Domani: Corte costituzionale colombiana; FedScoop: federal judges’ erroneous rulings; Judiciary UK: AI Judicial Guidance (04.2025); AI Act, Recital 61; CEPEJ: European Ethical Charter; Prawo.pl: AI and adjudication by judges; Polish Sejm: parliamentary question no. 18057. We describe rulings based on commentaries; we have not read their full texts. Legal assessments, forecasts and ethics are our analysis, not legal advice.