A Court Fined the Signature, Not the Chatbot. Your Assignment Has the Same Gap.

The New Mexico Supreme Court did not punish that lawyer for using ChatGPT. It punished him for signing.

That distinction is the entire story, and almost every account of it has buried the distinction under the chatbot. On September 11 the court entered a dispositional order of direct contempt against Stephen D. Aarons, a Santa Fe attorney, in the appeal of a murder conviction. He was fined five thousand dollars, referred to the Disciplinary Board, and barred from appearing before the court pending the outcome. At the August 21 hearing Justice C. Shannon Bacon asked him, according to reporting on the hearing, “Counsel, do you watch the news?” and observed that lawyers relying on AI hallucinations is an above the fold story every single day.

She is right. It is. Which is exactly why the order underneath the headline is worth four minutes of your morning, because the court wrote down something more useful than a warning about chatbots.

Read paragraph five

The order records the AI use almost in passing. Paragraph three notes that the attorney “used the generative artificial intelligence (AI) tool ChatGPT in the preparation of the brief in chief.” No adjective. No alarm. It is a fact in a list of facts.

The finding lives in paragraph five:

Respondent admitted to the Court that he did not verify the factual claims and legal authority in his AI-generated brief before signing it and filing it with the Court, and that he did not inform his client of this failure.

Before signing it. The court did not build its sanction on the tool. It built the sanction on the gap between generating a document and putting your name on it, and on the fact that nothing happened in that gap.

There is a version of this story where the lesson is “do not use AI for legal briefs.” That version is comfortable and it is wrong, and it will be wrong in your discipline too. The order describes a workflow with a production step and a filing step and no step in between.

What was actually in the brief

This is where it stops being abstract, and it is worth being precise because the precision is the argument.

The brief did not contain slightly wrong citations. It contained testimony from four witnesses who do not exist: the order names Officer Michelle Amarillo, Officer Sanchez, Manal Al-Jibury, and Teresa Marquez. It also contained false testimony attributed to people who do exist, including an account of threats a witness supposedly received and a description of the shooter’s clothing and appearance that nobody gave. On top of that it misrepresented the holdings of two real New Mexico cases, State v. Lopez and State v. Manus.

So the failure mode was not a citation that pointed at nothing. It was a document that read exactly like a brief, cited real cases, quoted named officers, and described a defendant’s clothing. Every surface signal of a competent filing was present. The attorney told the court he expected the tool to produce, in his words, a bulletproof summary of the trial.

Hold that phrase. A bulletproof summary. That is not a person misusing a tool. That is a person who believed the output was already checked.

Who paid

The five thousand dollars is the number in the headlines. It is not the interesting cost.

The order strikes all briefing in the appeal. It appoints the Law Office of the Public Defender to represent the defendant and directs new counsel to enter an appearance. It states the case will proceed with the intention that it be heard during the court’s 2026 to 2027 term.

Read that as a schedule. A man appealing a murder conviction had his appeal reset to zero and handed to a lawyer who had never seen the file, and the calendar moved. He did not use ChatGPT. He did not skip a verification step. He was the one downstream of somebody who did.

The court also found something the coverage mostly left out. Paragraph seven concludes that the attorney “demonstrated a lack of remorse and a lack of concern for his client.” That is a finding about a person, not about a model.

Your assignment has the same gap

Here is the part that is about you, and it is not the part you are expecting.

Almost every assignment in higher education has exactly two events in it. There is a produce step, where the student makes the thing. There is an evaluate step, where you grade it. Between those two there is a submit button, and the submit button is a signature, and no one has ever told a student that is what it is.

That was survivable when producing the thing was slow. The slowness was the verification. A student who spent six hours finding sources had, as a side effect of the six hours, read them. The checking was free because it was structurally unavoidable.

Production is no longer slow. The side effect is gone. And nothing was added to replace it, which is how you end up with prompt and pray students: not cheating, exactly, but submitting work whose truth value they have never personally established, because no step in the assignment ever asked them to.

The New Mexico order is what that looks like when the stakes are real and somebody downstream checks. In a course, nobody downstream checks. The error surfaces in the next course that assumed the prerequisite, or in a clinical placement, or on a licensure exam, or in the first month of a job. By then it is unattributable and nobody learns anything.

What a verification step looks like

You do not need a detector and you do not need a policy paragraph. You need one graded step between produce and submit. Four things make it real.

It names what the signature covers. Not “I affirm this is my own work.” That claim is now ambiguous and students know it. Something checkable: I have read every source I cite, and each quoted sentence appears in the source at the page given.

It produces evidence. A one page verification memo. Three sources, the exact quoted line, the page, and one sentence on what the source actually argues. If the student cannot produce the memo, the work is not finished, in the same way a lab report is not finished without the data.

It carries points. An ungraded step is a suggestion. Give it real weight, and grade the memo independently of the argument, because a student can reason well from sources they never opened and that is precisely the thing you are trying to catch.

It is the same shape every time. Verification is a habit, not an assignment feature. If it appears once in week nine it teaches nothing. If it appears in every submission from week one, by the end of the term the student has a process. Review first, always.

Notice what none of that requires. No detection software, no proctoring, no accusation, no arms race. The student is free to use whatever tool they like, which is the honest position anyway, because the claim they are signing is about verification rather than authorship.

The part nobody wants to say out loud

Adding a verification step to one assignment takes an afternoon. Adding it to every assignment in a course, with a rubric line that grades it consistently across ninety submissions, is a different kind of work, and it is the kind that quietly does not happen in week eleven.

That is a real constraint and it deserves a real answer rather than an exhortation. Build it once, properly, on one assignment. See whether it changes what you get back. It will, and the change is usually visible in the first batch.

If it works and you want the same step in all of them, the Assessment Builder inside the Faculty Toolkit writes assignments, rubrics and tagged question banks against your Course Record, so the verification line reads the same in week one and week fifteen without you retyping it. Do the first one by hand regardless. You need to know what the step feels like before you scale it.

Stephen Aarons had one verification step to perform. It sat between a document and his signature, it would have taken an afternoon, and the reason we are all reading his name this week is that nothing in his process required him to take it.

Go look at your next assignment. Find the moment where the student signs. Ask what, exactly, they are signing for.

More on this: join the Harvest Kernel Learning Community, where this week’s discussion is about where the checking step lives in your own course.

Sources

• Supreme Court of the State of New Mexico, State v. Sandoval, No. S-1-SC-40845, In re Stephen D. Aarons, Dispositional Order of Direct Contempt, filed September 2026. Full text: DocumentCloud. All quotations above are taken from this order.
• Emma Roth, “Lawyer fined $5K over AI-hallucinated witnesses in a murder case,” The Verge, September 11, 2026.
• “New Mexico lawyer fined for using AI-generated brief containing fabricated testimony,” The Guardian, September 11, 2026.