AI chatbots excel at translating dense legal language into plain speech, but they remain unreliable as sources of legal citations. Court records confirm the pattern: by November 2026, more than 2,200 rulings worldwide had documented instances where a judge or tribunal found that a party submitted content fabricated by an AI, and some of those cases are Spanish.
- Understand, don’t decide: AI is useful for explaining legal texts, organizing facts, and preparing questions for a lawyer—not for delivering verdicts or drafting filings without professional review.
- Hallucinations reach courtrooms: A registry maintained by Damien Charlotin counted 2,292 rulings as of November 12, 2026, in which courts found or implied that a party relied on invented content. Of those, 1,553 are from the United States and 8 from Spain.
- Verification is free and fast: Spain’s BOE offers consolidated legal texts, article-level version selectors, and electronic codes that let anyone confirm whether a cited norm exists, says what the AI claims, and remains in force.
What AI Can and Cannot Do for a Legal Doubt
Used well, AI saves you the most thankless part of a legal consultation: arriving without knowing what to ask. It can translate a statute or contract clause into everyday language with a practical example, summarize your case into a chronology with gaps flagged, prepare a list of questions for a lawyer or consumer service, and point you toward the law that governs a given issue so you can read it at the source.
What it should not do is issue a verdict on your case or produce a filing for submission without professional review. OpenAI’s usage policies, in the version applicable from November 20, 2026, prohibit creating or offering AI tools that give others personalized advice requiring a license without the involvement of a qualified professional, and prohibit automating high-impact actions in sensitive domains—including legal—without human review.
If cost is the barrier, Spain’s Law 1/1996 on free legal assistance includes, in Article 6, free advice and orientation prior to proceedings for those who meet the requirements.
Fabricated Citations and the Cases That Reached Court
The failures follow a pattern: invented citations, articles from another country’s laws, and judgments that never existed. And it is not only laypeople—lawyers who failed to review what the machine gave them have also been caught.
In Spain, the reference case is an order from the Superior Court of Justice of Navarra dated September 4, 2024 (ECLI:ES:TSJNA:2024:38A). A lawyer filed a complaint citing an article from Colombia’s Penal Code as if it belonged to the Spanish Penal Code. According to the order, the lawyer explained that ChatGPT had provided the text through improper handling of the system in his office and apologized immediately. The court archived the piece without a fine because the matter was “certainly novel” and because of that rapid reaction, but it wrote that careless use of these technologies “is susceptible to constituting a clear example of procedural bad faith.” The fine contemplated by Article 247 of the Civil Procedure Law ranges, the order recalls, from 180 to 6,000 euros.
The same order cites a U.S. precedent, the Avianca case, in which a federal district judge in New York imposed a $5,000 sanction on June 22, 2023 for improper use of ChatGPT in court filings. It also cites a January 2024 study finding that legal hallucinations appeared in 69% of cases with ChatGPT 3.5 and 88% with Llama 2. Those figures come from models of that era, but the conclusion bears repeating: models do not always know when they are making something up.
Navarra saw a later episode. According to reporting from March 26, 2026, the Social Chamber of the same court archived another sanctioning piece against a lawyer who, in an October 2025 appeal, included eight invented citations of rulings from the Constitutional Court, the Supreme Court, and the superior courts of Navarra and Madrid. Again there was no fine, due to the speed of her apologies.
There is a second problem that weighs heavily in legal matters: chatbots tend to agree with you, and if you ask “I can claim this, right?” they are likely to say yes.
How to Verify AI Claims in the BOE
Three checks suffice: that the norm exists, that the article says what the AI claims, and that the version is current. All three are free on the BOE website.
1. Find the consolidated text. When the AI cites a law, look it up in the BOE’s consolidated legislation section. The consolidated text incorporates all amendments. The header shows the date of its last update: in the Trademark Law, for example, it reads “Last update, published 28/07/2022.” If the chatbot cites a law that does not appear, or with a number that does not match, stop there.
2. Read the exact article. Do not settle for the summary. Go to the article the AI mentions and compare. The Navarra case illustrates the point: the cited provision did not correspond to the Spanish Penal Code, something a quick read would have caught.
3. Check the version. Next to many articles, the BOE offers a “Select wording” selector that shows each version with its entry-into-force date. It is the key tool against a very common AI error: giving you an old wording as if it were current. The “Analysis” section of each norm also indicates which other provisions have amended it.
The Electronic Codes of the BOE’s Digital Legal Library gather the main norms in force by branch of law, are updated permanently, and allow subscription to alerts through Mi BOE.
For judgments, the site is different. If the AI cites one, search for it in the General Council of the Judiciary (CENDOJ) search engine by court, date, or number. If it does not appear, treat it as invented until someone proves otherwise.
The lesson from the Navarra cases extends beyond the legal profession. If professionals of the law, with years of training and ethical responsibility, had invented citations slip past them, the risk for the rest of us is greater. AI is a good first step to understand your legal problem. Verification at the source is the second step, and it is not optional. And for anything that will be signed or filed before a court or administration, the third step remains a person who answers for what it says.