From prompt to precedent: how a Mexican court turned ChatGPT into a judicial tool

Daniel Morán
Pérez-Llorca, Mexico City
daniel.moran@perezllorca.com

In 2025, a Mexican federal appellate court used a standardized prompt and ChatGPT to calculate court-ordered security in an actual judicial proceeding. After applying the methodology in five cases, two binding precedents were published in January 2026: one establishing minimum requirements for the ethical and responsible use of artificial intelligence in judicial proceedings, and another validating its use to calculate security in amparo proceedings. The decisions matter not merely because a court used ChatGPT, but because they begin to define the conditions under which AI may become part of judicial methodology and the safeguards litigants may need when its output affects their procedural or financial position.

An ordinary case. An unusual response.

The case arose from an amparo indirecto proceeding — a form of constitutional review in Mexico — connected with a mortgage foreclosure action. A litigant claiming defective service obtained a preliminary stay preventing enforcement of the challenged decision.

The district judge conditioned the stay on court-ordered security of approximately USD 2,500. The claimant challenged the amount as arbitrary.

Mexican Supreme Court precedent had already established a methodology for calculating security, linking potential damages to inflation-driven loss of purchasing power and economic losses to the return the money would have generated.

The appellate court agreed that the lower court’s reasoning was insufficient. Instead of simply performing the calculation itself, however, it incorporated a standardized ChatGPT prompt into its methodology. The operator would provide the relevant case data, while the prompt instructed the system to work with updated inflation figures and the Mexican Interbank Equilibrium Interest Rate (TIIE) from official sources and apply the formulas established by precedent.

Using a principal amount of approximately USD 163,000 and an eleven-day period, the AI-assisted calculation produced security of approximately USD 18,700, more than seven times the amount originally ordered.

Yet the court did not increase the security. Because only the claimant had appealed, doing so would have worsened her legal position, contrary to non reformatio in peius. The AI-assisted result therefore entered the court’s reasoning without ultimately increasing the appellant’s financial burden.

What happens when the next AI-assisted calculation actually determines the amount that a party must post or pay? 

What Did ChatGPT Actually Contribute?

According to the subsequent precedents, AI may reduce human error, increase transparency and traceability, promote consistency and improve procedural efficiency.

The choice of tool nevertheless deserves examination.

Once the legally relevant inputs and parameters have been identified, the exercise appears largely mechanical: obtain particular economic information and apply formulas that have already been established.

What does a tool such as ChatGPT add to a calculation that should, in principle, produce the same result from the same data and formulas? 

That question is not itself an argument against using the tool. Generative AI may offer advantages: a natural-language interface can make a complex methodology accessible without programming expertise, adapt a single prompt to different variables and produce a step-by-step explanation for judicial review.

But those advantages must be considered alongside the technology’s characteristics. Outputs may depend on the model version, available functionalities, the way data are obtained and the particular interaction between user and system.

The judgment leaves one especially important issue unclear: whether the relevant inflation and interest-rate data were first obtained and independently verified by court personnel and then supplied to ChatGPT, or whether the system itself was asked to locate them.

The distinction matters. In the first scenario, ChatGPT functions mainly as a calculation and explanatory tool. In the second, it also participates in gathering external information that feeds the calculation. The risks involved and the degree of human verification required are not the same.

That is why the real novelty of the Mexican precedents is not merely that a court used ChatGPT. It is that the court attempted to subject that use to explicit legal standards.

From experiment to binding standards

The methodology was applied in five cases during 2025. In January 2026, that repetition resulted in two binding precedents which, as far as we have been able to determine, are Mexico’s first specifically devoted to judicial use of artificial intelligence.

The more significant of the two establishes a general framework. In the absence of specific regulation, judges using AI tools must observe, at a minimum, four principles:

1. proportionality and do no harm;
2. protection of personal data;
3. transparency and explainability; and
4. human oversight and decision-making.

The second precedent applies that framework to the calculation of security in amparo proceedings, provided that AI remains an aid and does not replace the judicial function.

The decisions also attempt to draw a substantive boundary: AI may facilitate numerical reasoning, but it must not displace human legal reasoning in interpreting or applying the law.

The distinction is intuitively attractive. In practice, it may be harder to maintain. Before a calculation can be performed, someone must decide which amount is legally relevant, what period should be used, which rate applies and what assumptions should be incorporated. Those judgments are not purely mathematical.

The more useful question may therefore be whether a task can be structured so that every legal judgment remains visible and under human control while technology is limited to executing those judgments.

Four safeguards and three procedural battlegrounds

The technical battleground: what makes the process truly auditable?

Disclosure that AI was used — or even disclosure of the prompt — may not be enough. Meaningful review may require knowing which version of the tool was used, what data were entered, which sources were consulted, what output was produced and how the result was verified.

Legal auditability does not require perfect technical reproducibility. A reviewing court need not make the tool generate exactly the same words months later. But the record should contain enough information to reconstruct and challenge the process that actually influenced the decision.

The objective is not to reproduce the machine’s internal state. It is to make the court’s use of the machine reviewable.

The cognitive battleground: what does human oversight actually mean?

The precedents insist that the judge retain final decision-making authority. But human oversight is not self-defining.

Were the economic inputs independently verified? Were the calculations repeated outside the AI tool? Could the decision-maker explain the methodology without relying on the AI-generated presentation? Or did review consist primarily of approving an output that appeared convincing and precise?

There is a meaningful difference between human approval and meaningful human verification. Oversight operates as a safeguard only if it is capable of detecting what the tool may have done incorrectly.

The procedural battleground: can the parties effectively challenge the process?

Transparency allows the parties to know what happened. The ability to challenge determines what they can do about it.

Not every tool used internally by a court requires prior adversarial scrutiny. Judges routinely conduct research, calculations and internal analysis without circulating them before ruling.

The problem may be different, however, where an AI-assisted methodology introduces information from outside the record, depends on contestable assumptions, cannot be evaluated solely from the judgment, or directly determines what a party must post or pay.

In those circumstances, transparency after the decision may not be enough. The relevant issue is whether the affected party has a meaningful procedural opportunity to identify and challenge the data, assumptions, methodology and result.

Why this matters beyond Mexico

The precedents are domestic, but the questions they expose are not.

International commercial disputes increasingly involve financial information, confidential business records, personal data and evidence originating in multiple jurisdictions. If judicial AI tools are used to process or retrieve that information, questions about what may be entered into an external system, how it is protected, what must be preserved in the record and how the resulting analysis can be challenged become procedural questions, not merely questions of technology policy.

For cross-border litigants and international counsel, the Mexican precedents illustrate an emerging challenge: once AI becomes part of judicial methodology, principles of reasoned decision-making, procedural fairness, confidentiality and reviewability must be translated into operational safeguards.

The real test Is sill to come

Mexico’s first binding precedents on judicial use of artificial intelligence may represent the beginning of a much broader process.

AI is unlikely to remain confined to the calculation of security. As courts incorporate these tools into additional tasks, new disputes will force them to define what must be preserved in the record, what degree of human verification is sufficient, when AI use must be disclosed, and when a party must be allowed to challenge the data, assumptions or methodology employed.

Above all, courts will eventually confront the question that these first cases did not have to resolve in practical terms: what happens when an AI-assisted result, even after being reviewed and approved by a judge, actually changes what a party is required to do?

The first precedents answered one question: whether artificial intelligence may enter the judicial process.

The more difficult questions begin when its output matters to the outcome.

---

Notes

1. Second Collegiate Court in Civil Matters of the Second Circuit, Complaint Appeal (Queja) 404/2025, judgement dated 24 October 2025.

2. Binding precedents II.2o.C. J/1 K (12a.), digital registry no. 2031639, and II.2o.C. J/2 K (12a.), digital registry no. 2031640, published on 9 January 2026 and binding as of 12 January 2026.