In short: Latin American courts have gone from one viral ChatGPT judgment to a working body of law on AI in under four years. They agree that a human must decide. Where they disagree is on what happens when AI gets it wrong. Colombia's Supreme Court fined a lawyer 15 monthly minimum wages for citing law an AI invented. A month later, El Salvador's Constitutional Chamber found a pasted chatbot reply in a citizen's petition and imposed nothing.
One question, five courts
On 31 January 2023 a Colombian judge, Juan Manuel Padilla, quoted ChatGPT in a ruling on the health coverage of a child on the autism spectrum. It went around the world. Most coverage of AI in Latin American courts still starts and ends there.
A lot has happened since. We wanted to know what the courts have decided, so we asked the same question of every high court we could query, and we read what came back:
When may AI be used inside a judicial proceeding, and what happens when it goes wrong?
We pulled the answers straight from the official case-law databases of Colombia's Constitutional Court, Supreme Court and Council of State (the country's top administrative court), Mexico's federal judiciary, Peru's Constitutional Tribunal and El Salvador's Supreme Court. We cite only rulings we pulled and read ourselves.
A confession first. AI was not our first question. We started with habeas data, the Latin American constitutional right to see and correct the data held about you. It seemed the richer comparative topic, and it broke the search in the first ten minutes. Four of the five courts returned the maximum 1,000 results, and the top hits were about pensions and tax records. Of everything we tried, AI was the one question with a small, exact, on-topic result set at every court. That tells you something too: this body of law is new enough that you can still read all of it.
What each court decided
Colombia's Constitutional Court: allowed, inside a tight box
The landmark ruling is T-323 of 2024, decided on 2 August 2024. A judge had used ChatGPT 3.5 while writing a decision about a child's health care. The Court found no violation of due process, because the AI was used "luego de haberse fundamentado y tomado la decisión": after the decision had already been reasoned and made.
It did not let the judge off, though. The Court held that the duties of transparency and responsibility "no se cumplieron a cabalidad" (were not fully met), and it drew the line in unusually concrete terms. AI may help with administrative work, drafting support and "corrección y síntesis de textos". It must not be used for "creación de contenido ni interpretación de hechos o pruebas y, mucho menos, la solución de casos": creating content, interpreting facts or evidence, and least of all deciding cases.
The Court set out twelve principles and ordered Colombia's judicial council to write a guide. In January 2026, in T-008 of 2026, it declined to rule on AI-drafted administrative acts and sent that question to the administrative courts.
Colombia's Supreme Court: misuse is sanctionable
On 13 February 2026 the Civil Chamber issued AC739-2026. A lawyer had filed a brief built on law and precedent that did not exist: the fake-citation problem, with a price tag attached. The Court fined him 15 monthly minimum wages (the statutory range is 10 to 50) and referred him for professional discipline. The sentence that matters:
"Trasladar un texto generado por IA sin contraste alguno constituye ceguera voluntaria (dolo)." Copying AI-generated text without checking it is wilful blindness, and wilful blindness counts as intent.
That has teeth. "I didn't know the case was fake" stops working as a defence, because not checking is treated as knowing. In June the Criminal Chamber rejected the lawyer's challenge to the sanction (STP13149-2026), and in STC9687-2026 the Civil Chamber applied T-323's rule on hallucinated reasoning "en todos los contextos", in every context.
Colombia's Council of State: the court discloses its own AI
The Council of State returned 257 hits for "artificial intelligence", far more than any other court. Most of them are not about AI at all. A ruling from its Tax Section dated 3 September 2026 ends with a note: Copilot and GPT-4 were used "con el único propósito de efectuar una revisión ortográfica, gramatical y léxica", only to check spelling, grammar and vocabulary. It cites the judicial council's agreement PCSJA24-12243.
We read one such ruling closely and saw the pattern in the search results. We did not open all 257. It is the quieter half of the story: a top court now tells litigants, in the ruling itself, which AI it used and for what.
Mexico: AI for the arithmetic, never for the reasoning
Mexico's contribution comes from a federal circuit court, the Second Collegiate Court in Civil Matters, published in the Supreme Court's official Semanario Judicial. It did not come from the Supreme Court itself, and the difference matters: circuit jurisprudence binds within that circuit, not nationwide.
In two jurisprudential theses published on 9 January 2026 (registries 2031639 and 2031640), the court held that AI tools "pueden emplearse válidamente" (may validly be used) to calculate the bond (the garantía) a party must post in amparo proceedings, Mexico's constitutional rights action. Its reasoning is that such calculations "no forman parte del núcleo decisorio jurisdiccional": they are not part of the decision's core, but auxiliary work that technology can support. It went further than any other court we read and called the practice "recomendable" for courts adopting best practice.
It also set four minimum conditions: proportionality and harmlessness, personal-data protection, transparency and explainability, and human supervision and decision. AI may help with the numbers "sin que se alcance al razonamiento jurídico", as long as it never reaches the legal analysis.
Both began as non-binding theses in August 2025, written by the same court about its own use of AI.
El Salvador: allowed, and no sanction for citizens
On 13 March 2026 El Salvador's Constitutional Chamber ruled on an unconstitutionality petition (case 57-2025) against Article 46 of the country's telecommunications-interception law. The filing still contained the chatbot's reply: "Perfecto. A continuación, te presento un texto…" ("Perfect. Below is a text for you…").
The Chamber adopted Colombia's T-323 criteria by name and held generative AI "admisible en los procesos constitucionales", admissible in constitutional proceedings. It found the duties of transparency and oversight breached. It dismissed the petition as inadmissible, because its central argument was "la conclusión de un razonamiento que carece de premisas", a conclusion with no premises behind it. The AI-drafted case failed on its merits. But the Chamber imposed no sanction, because filing such a petition is "una facultad ciudadana", a citizen's right, unlike the work of lawyers and judges.
Peru: silent, as far as the database shows
Peru's Constitutional Tribunal returned two results. Both were false positives: a police officer's tattoo case and a dispute about works on a church square. "ChatGPT" and "algoritmo" returned nothing. Peru's top constitutional court has not written on AI in proceedings yet, or has not published it in a way the search can find.
Brazil we could not include at all. The official source we used publishes legislation, not case law.
Încearcă HAQQ AI gratuit
Experimentează redactarea și cercetarea juridică cu inteligență artificială
Where they agree, and where they split
Side by side, the rulings agree more than they disagree. Every court that has spoken says three things:
- A human decides. AI can support, never substitute.
- Disclose it. Transparency is a duty, not a courtesy.
- Check it. Unverified output is the user's responsibility.
The real split is over consequences, and it runs along a line most coverage misses: who made the mistake.
| Colombia | El Salvador | Mexico | |
|---|---|---|---|
| AI allowed? | Yes, auxiliary only | Yes | Calculations only |
| Where the line sits | Never interprets evidence or decides | Follows T-323 | Never the analysis |
| Lawyer cites fake law | Fined 15 minimum wages | No ruling yet | No ruling yet |
| Citizen uses chatbot | No ruling yet | Breach, no sanction | No ruling yet |
Colombia treats an unchecked AI citation from a professional as intent. El Salvador treats an unchecked chatbot text from a citizen as a procedural lapse. Both positions are defensible, and they will collide as soon as a self-represented litigant files AI-drafted work in Bogotá, or a lawyer does it in San Salvador.
What the research itself taught us
Getting this far was harder than the rulings suggest, and the difficulty is the point.
- Search is primitive everywhere. The case-law search behind these courts matches every word anywhere in the full text. It is not phrase search, and results come back newest first, not by relevance. That is how a tattoo case ends up in an AI search.
- Every court publishes differently. The Council of State gives you the legal problem, a yes-or-no answer and the thesis. Peru gives you a publication date and nothing else, not even the date of the ruling. Mexico publishes abstracted theses, not the judgments behind them. El Salvador redacts party names.
- Coverage splits in silent ways. Colombia's Council of State database changes format in December 2021. Search only the newer one and you lose twenty years without any error message.
- Freshness varies. When we pulled the data, most sources were a day old and Mexico's was three weeks behind.
- The documents are huge. T-323 alone runs to about 438,000 characters.
Finding a ruling is easy now. Knowing whether five differently built databases are telling you the same thing is where the week goes.
Our take
We build legal AI, so this body of law is written about us as much as about judges and lawyers. We think the courts have the bar in the right place. AI output is a draft. A human owns the decision. Every source has to be real, and checkable by the person signing.
The Colombian Supreme Court's framing should make every legal-AI vendor uncomfortable in a useful way. If copying unchecked AI text is wilful blindness, then a tool that makes checking hard is a tool that sets its users up for sanctions. Citations that link to real sources, honesty about what the model did and did not check, a hard line between drafting and deciding: in Colombia, that is now the difference between a useful tool and a fine of 15 minimum wages.
Key Takeaways
- Latin American courts agree that AI may assist in proceedings, but a human must decide, disclose the use and verify the output.
- Colombia's Supreme Court treats filing unchecked AI-generated law as wilful blindness and fined a lawyer 15 monthly minimum wages (AC739-2026).
- El Salvador's Constitutional Chamber found a pasted chatbot reply in a citizen's petition and imposed no sanction (57-2025).
- A Mexican federal circuit court called AI "recommendable" for calculations but barred it from the legal analysis (2031639, 2031640).
- Comparing courts is still hard: every database searches, dates and structures its rulings differently.
Sources & further reading
- Corte Constitucional de Colombia, Sentencia T-323 de 2024
- Corte Constitucional de Colombia, Sentencia T-008 de 2026
- Corte Suprema de Justicia de Colombia, AC739-2026 (13 February 2026)
- Corte Suprema de Justicia de Colombia, STP13149-2026 (16 June 2026)
- Corte Suprema de Justicia de Colombia, STC9687-2026 (3 June 2026)
- Consejo de Estado de Colombia, Sección Cuarta, radicado 11001031500020260521600
- Semanario Judicial de la Federación, tesis 2031639
- Semanario Judicial de la Federación, tesis 2031640
- Sala de lo Constitucional de El Salvador, jurisprudence portal (Inconstitucionalidad 57-2025)
- Verfassungsblog, on the 2023 Colombian ChatGPT ruling
- Counterparty due diligence in seconds: what US public records show
