Stephen Aarons ChatGPT Sanctions Turn Fake Witnesses Into a Legal Crisis

Stephen Aarons received a $5,000 sanction after his ChatGPT-assisted appeal brief presented false testimony and witnesses who never appeared at trial. The New Mexico Supreme Court also held the veteran defense attorney in direct contempt. The Stephen Aarons ChatGPT sanctions expose a failure far more serious than another invented legal citation.

Aarons represented Oscar Renee Sandoval, who received a life sentence after a murder conviction. According to the court, the appellate brief misrepresented both the trial record and legal authorities. Some passages described testimony from fabricated witnesses, including supposed observations about the shooter’s clothing and appearance.

That difference changes the stakes. Earlier AI sanctions often involved nonexistent cases or quotations. This filing allegedly altered the factual story of a criminal trial, where the client’s liberty depended on an accurate record. It turned an unreliable summary into purported evidence and carried that material into the state’s highest court.

The Stephen Aarons ChatGPT Sanctions Went Beyond a Fine

The court treated the filing as a threat to the appeal itself, not simply an embarrassing research mistake.

The New Mexico Supreme Court imposed the sanctions on September 9, 2026. The justices held Aarons in direct contempt and ordered him to pay $5,000 to the State Bar of New Mexico Client Protection Fund.

The court also referred the matter for attorney discipline. Aarons cannot appear before the state Supreme Court while that process remains pending. That restriction separates the case from incidents resolved through warnings or continuing education.

The existing appellate briefs were struck. Sandoval’s appeal was reassigned to attorney Kim Chavez Cook through the New Mexico public defender system. The underlying appeal remains pending, according to case reporting published by Reuters.

Those remedies recognize two distinct injuries. The first concerns the court, which received a filing containing material that did not match the record. The second concerns Sandoval, whose direct appeal was delayed and forced to restart with new counsel.

Sandoval had pleaded not guilty to murdering the mother of his children. He was convicted and sentenced to life imprisonment in 2025. His capital appeal raised issues involving evidence, witness disclosure, expert testimony, confrontation rights, and the effectiveness of trial counsel.

The New Mexico judiciary’s term case list confirms those appellate questions. They required careful comparisons between the trial record, preserved objections, and controlling law.

Instead, the court said the submitted brief contained false police testimony and statements attributed to people who had not provided them. It also included inaccurate descriptions of legal authority.

The distinction between evidence and argument matters. A fabricated precedent can distort the legal rule governing a case. Fabricated testimony can rewrite what jurors heard, which witnesses existed, and why a verdict should be reversed.

A false factual summary can then contaminate every argument built above it. An ineffective-assistance claim, for example, depends on what counsel did and what the record actually contained. An invented witness can create a failure that never occurred.

The court’s response therefore reached the filing, the lawyer, and the representation. Striking the briefs removed the contaminated work product. Appointing new counsel gave the defendant a clean path forward. The disciplinary referral moved the professional questions beyond the appeal.

That structure explains why the Stephen Aarons ChatGPT sanctions deserve attention outside New Mexico. The court did not announce that lawyers must avoid generative AI. It enforced the older rule that a lawyer remains responsible for every filed assertion.

The episode also tests a familiar defense against AI mistakes. Aarons called the incident an honest error and said he did not understand how extensively ChatGPT could invent facts. The court focused instead on the unverified document bearing his signature.

A signed brief is not a brainstorming note. It is a formal representation to judges, opposing counsel, and a client. Once generated text enters that document, its origin does not reduce the lawyer’s responsibility.

An AI Summary Became a False Version of the Trial

The central failure occurred when generated prose was treated as a faithful record without line-by-line verification.

At an August 21 hearing, Aarons described giving ChatGPT a computer-generated transcript and other case materials. He expected the system to produce what he called a “bulletproof summary.”

That expectation reversed the proper burden of proof. A generative model produces plausible text from patterns and context. It does not certify that each sentence appears in an uploaded transcript.

An AI hallucination is a confident output that lacks support in the relevant evidence or source. The term can sound harmless, but its consequences depend on where the output travels. A wrong summary in private notes creates confusion. The same summary in an appellate brief becomes a representation to a court.

The brief reportedly attributed statements to Danny Stanton and Linda Stanton that the court found false. It also referred to Mariah Chavez and Teresa Marquez in descriptions concerning the shooter’s clothing and appearance. The court identified Marquez as a fabricated witness.

One passage apparently said the shooter wore dark pants and a white shirt. The statement sounded like an ordinary factual detail, which made it easy to absorb into the larger narrative. Yet the court found no valid record support for it.

This is a harder problem than detecting an obviously absurd answer. Generated legal prose often uses the correct tone, expected vocabulary, and familiar argumentative structure. Those qualities make unsupported details appear settled.

A lawyer reading quickly can recognize the overall story and miss the invented connective tissue. Names, clothing descriptions, threats, and witness observations can look consistent with surrounding material. Fluency disguises the missing provenance.

The workflow also blurred two different documents. A transcript records testimony and courtroom events. A model-generated summary is a new text created from that source. It must be treated as an interpretation until every material claim is traced back.

Uploading source material does not eliminate hallucinations. Long records can exceed a system’s effective attention, contain transcription errors, or invite unsupported synthesis. A model can also combine nearby details into a statement no witness made.

This risk increases when instructions ask for a definitive narrative. Terms such as “complete,” “bulletproof,” or “comprehensive” can reward apparent completeness over explicit uncertainty. The output may fill gaps instead of flagging them.

A reliable legal workflow needs claim-level provenance. Every factual sentence should point to a transcript page, exhibit, docket entry, or judicial finding. Every quotation should match the source exactly. Every cited case should exist and support the stated proposition.

That process cannot end with asking the same chatbot whether its summary is accurate. A generator can defend its own unsupported output with further invented explanations. Verification requires returning to independent source material.

The lesson applies beyond law. Researchers, analysts, consultants, and executives increasingly use AI to compress long records. The danger rises when a summary becomes the only version that later reviewers see.

Teams need to preserve the original evidence beside the generated synthesis. A structured knowledge blending workflow can help connect notes with source documents. However, retrieval alone does not establish that every generated claim is correct.

The decisive safeguard remains human review against authoritative records. AI can help locate passages, organize issues, and draft working notes. It should not transform an unverified inference into a factual assertion.

The Lawyer’s Signature, Not ChatGPT, Carried the Duty

The primary conflict is between delegated drafting and nondelegable professional responsibility.

Aarons told Reuters that he had practiced criminal defense for more than 40 years. He said he was unfamiliar with the extent of AI hallucinations when he accepted Sandoval’s appeal in 2025.

“I am remorseful but hopeful that the disciplinary board takes into account it was an honest mistake,” he said. He characterized the episode as a lesson for professionals using an unstable technology.

The court reached a harsher assessment. Its order said Aarons admitted that he did not verify the factual claims or legal authorities before signing and filing the brief. It also said he failed to tell his client about those deficiencies.

The justices found that he demonstrated insufficient remorse and concern for Sandoval. Justice C. Shannon Bacon challenged the claim that a lawyer could remain unaware of widely reported AI failures. Her questions emphasized that hallucinated legal material had already become a recurring public issue.

This tension does not require treating ChatGPT as blameless. The system produced language that, according to Aarons and the court, included unsupported material. OpenAI did not provide Reuters with an immediate response to the sanctions.

However, a chatbot cannot enter an appearance, sign a brief, advise a client, or answer a contempt order. Those duties belong to licensed counsel. The legal system assigns accountability to the person authorized to act.

The American Bar Association reached the same basic conclusion before this case. Its AI ethics opinion says lawyers must understand a generative tool’s benefits and risks.

Formal Opinion 512 connects AI use to existing duties involving competence, confidentiality, client communication, supervision, candor, meritorious claims, and reasonable fees. It does not create an AI exception to those duties.

The guidance also describes generative text systems as prediction tools. They produce statistically probable outputs from patterns in data. That design differs from a database that returns an authenticated court document.

A lawyer can delegate research or drafting to associates, paralegals, contractors, and software. Delegation changes who performs the first pass. It does not change who must supervise the work or verify the filing.

This case shows why disclosure alone would not solve the problem. A label saying “AI assisted” might alert the court, but it would not make fabricated testimony accurate. Judges cannot become fact-checkers for every filing produced with software.

Tool bans face similar limits. A lawyer can introduce false information through bad memory, careless dictation, untrained staff, or ordinary copy-and-paste errors. Courts care about the accuracy of the signed filing, regardless of production method.

The better dividing line concerns verification. AI assistance can remain useful when each material claim is checked against the record. It becomes dangerous when fluency substitutes for that check.

Client communication adds another layer. A criminal appellant should understand when defective briefing has put an appeal at risk. According to the court, Aarons did not tell Sandoval about the brief’s factual and legal misrepresentations.

That omission matters because the client bears the practical consequences. Counsel faces sanctions and discipline, but the appellant experiences delay, uncertainty, and a disrupted attorney relationship.

Professional accountability must therefore cover the whole workflow. It starts with selecting an appropriate tool and protecting confidential information. It continues through source validation, substantive review, client communication, and final certification.

ChatGPT Fake Witnesses Crossed a More Dangerous Line

Invented testimony moves AI hallucinations from faulty research toward contamination of the evidentiary record.

The best-known early sanctions case involved Mata v. Avianca. Lawyers submitted nonexistent judicial opinions generated through ChatGPT in a federal personal injury dispute. A judge imposed a $5,000 penalty in 2023.

That case established a durable warning. A chatbot can invent case names, quotations, docket numbers, and legal reasoning. Courts expect attorneys to confirm those authorities through recognized research systems.

The New Mexico episode expands the problem. Aarons’ filing reportedly included fabricated witnesses and false descriptions of testimony. Those errors did not merely misstate what another court had decided. They created facts inside the client’s own case.

Appellate courts generally review an existing record. They do not conduct a new trial or accept a lawyer’s factual additions as evidence. Briefs must ground their narratives in transcripts, exhibits, rulings, and preserved objections.

An invented factual detail can influence how judges understand prejudice. It can make excluded evidence seem more important, strengthen an identification argument, or change the apparent value of cross-examination. These effects can persist even after correction.

The problem also burdens opposing counsel. Prosecutors must compare the brief with the record and identify discrepancies. Judges and clerks must determine whether a strange statement reflects a genuine transcript passage or generated fiction.

This verification cost spreads across the system. One attorney can save drafting time while transferring review work to the court, the opposing party, replacement counsel, and the client. Sanctions attempt to place that cost back on the responsible professional.

A 2026 Georgia murder appeal showed another route for contamination. In Payne v. State, fictitious or misattributed citations appeared in a prosecution filing and a proposed order signed by the trial court.

The Payne decision vacated the trial court’s order and returned the matter for a new one. The episode demonstrated how generated errors can migrate from an advocate’s draft into a judicial document.

That migration risk makes provenance essential. Legal documents are frequently reused across motions, proposed orders, internal memos, and appellate briefs. Once an unsupported passage enters the chain, later authors may assume someone already verified it.

The New Mexico filing appears to have followed a similar trust shortcut. Aarons used ChatGPT to create a condensed representation of the proceedings. He then relied on that representation without fully returning to the transcript.

The resulting document had the surface form of appellate advocacy. It identified witnesses, described testimony, invoked precedent, and connected facts to legal claims. The form made the content appear professionally reviewed.

This is why “human in the loop” provides little assurance by itself. A human can glance at a draft, correct grammar, and still miss factual fabrication. The phrase matters only when the reviewer has time, expertise, and direct source access.

Effective review also needs visible boundaries between source text and generated text. Quoted transcript passages should retain page references. Summaries should identify their supporting ranges. Unsupported model commentary should remain outside the factual statement.

For high-stakes matters, a second reviewer should test the document independently. That reviewer should sample sources based on risk, then expand the check when any discrepancy appears. Fabricated names should trigger a review of every factual representation.

The same principle applies to legal authorities. Reviewers must open each cited decision and confirm the quoted language, holding, jurisdiction, and relevance. A real case cited for a false proposition remains a serious error.

The Stephen Aarons ChatGPT sanctions show that courts are moving past curiosity about which tool produced a mistake. The operational question is whether the lawyer built controls strong enough to keep generated fiction from becoming filed fact.

The Sanctions Do Not Prove AI Has No Place in Legal Work

The strongest criticism targets uncontrolled reliance, not every use of generative AI by lawyers.

It would be easy to treat this case as evidence that legal professionals should never use ChatGPT. That conclusion reaches beyond the court’s findings and ignores the range of tasks grouped under “AI use.”

A model can help create issue lists, reorganize attorney-written notes, compare document structures, or suggest search terms. Those uses still require judgment, but they do not all carry the same risk as summarizing a murder transcript.

Risk depends on the output’s purpose and destination. A private outline is provisional. A client letter requires careful review. A factual statement in a criminal appeal demands verification against the official record.

Tools also differ. A general chatbot produces open-ended text, while some legal systems retrieve from bounded databases and show supporting authorities. Even systems with citations can misquote, misread, or select irrelevant material.

The New Mexico court did not conduct a scientific assessment of ChatGPT’s reliability. It examined one lawyer’s conduct and one defective filing. Readers should not convert that finding into a universal error rate.

Nor does the case prove that AI caused every defect independently. Aarons selected the materials, prompted the system, used its output, and signed the result. Public reporting does not provide the complete prompt history or every intermediate draft.

That uncertainty matters. Without the full interaction record, observers cannot reconstruct which claims the model generated, which passages Aarons edited, or whether transcript defects contributed. The court’s order establishes responsibility, not a complete technical postmortem.

Still, those gaps do not excuse the filed inaccuracies. They clarify where product and process questions remain open. Better logging could help firms determine how an unsupported statement entered a document.

Legal AI vendors can reduce risk by attaching source references to every factual sentence. They can warn when an answer lacks evidence, separate quotations from summaries, and block exports containing unresolved citations.

Those controls would help, but they cannot replace substantive review. A source link can point to a real document that does not support the generated claim. Automation can make verification easier without making it optional.

Law firms also need policies based on task sensitivity. Summarizing public background material does not carry the same consequences as interpreting privileged evidence. Drafting a routine checklist differs from describing disputed testimony.

A practical policy should identify prohibited inputs, approved systems, retention rules, review standards, and escalation points. It should specify who checks citations and who confirms factual assertions against the record.

Training must go beyond showing lawyers how to write prompts. Users need to understand confident error, context limits, incomplete retrieval, data exposure, and automation bias. Automation bias occurs when people over-trust output because a system produced it.

Time pressure deserves attention as well. AI promises faster drafting, but verification can erase some of that gain. Firms must budget review time honestly instead of assuming the first fluent draft is nearly complete.

Supervisors should also avoid measuring success only through speed. That incentive encourages staff to conceal uncertainty and skip source checks. High-risk AI work needs accuracy metrics tied to evidence.

The skeptical conclusion is therefore narrower and stronger. Generative AI can assist legal work, but only when firms treat its output as unverified drafting. The person filing the document must still prove every material claim to themselves.

What Courts and Legal Teams Should Watch Next

The next test is whether sanctions produce verifiable workflow changes across legal practice.

The first signal will come from the disciplinary proceeding involving Aarons. The New Mexico board can examine competence, candor, client communication, and the circumstances surrounding the filing.

A public disciplinary decision would clarify whether the court views the incident as isolated negligence or broader professional misconduct. A stronger penalty would reinforce the direct-contempt order. A limited remedy would emphasize remediation and intent.

The proceeding may also reveal information absent from news reports. Investigators can examine the drafting process, communications with Sandoval, transcript handling, and the timing of Aarons’ discovery. Those facts will define the case’s value as a precedent.

The second signal will be Sandoval’s replacement appeal. New counsel must rebuild the arguments from the actual record after the earlier briefs were struck. That process will show how much substantive damage the fabricated material caused.

If the corrected brief raises similar legal claims using authentic evidence, the episode will look like a severe workflow failure with a recoverable appeal. If the claims change substantially, the AI summary likely distorted the case’s theory.

The court’s eventual treatment of Sandoval is equally important. Sanctioning counsel protects institutional integrity, but the system must avoid punishing the client for a lawyer’s defective filing. New representation addresses that risk without resolving every delay.

The third signal will be the adoption of evidence-linked drafting controls. Courts can require certifications, while firms can demand transcript citations for each factual statement. Vendors can expose the source behind every generated passage.

The most useful controls will be auditable. A checkbox stating that someone reviewed the draft offers limited protection. A claim map linking each assertion to a transcript page creates evidence of real verification.

Courts will also decide whether existing professional rules provide enough leverage. The ABA framework already covers competence, candor, supervision, and confidentiality. The New Mexico order suggests judges can act without creating AI-specific duties.

New rules still have potential value when they define procedures. They can require disclosure for particular uses, preservation of generated drafts, or personal certification of quotations and citations. However, disclosure should support verification rather than replace it.

The broader industry should watch for repeated factual fabrication in criminal cases. Fake precedents are now familiar. Invented testimony presents a more direct challenge to the integrity of appellate records.

For knowledge workers, the immediate action is simple but demanding. Identify every generated factual claim, reopen the controlling source, and record where the support appears. Escalate any mismatch before the text reaches a client, regulator, executive, or court.

The Stephen Aarons ChatGPT sanctions are not a warning about imperfect prose. They show what happens when an AI-generated summary becomes an undocumented substitute for evidence.

Ask one question before relying on any generated account: Can another reviewer trace every consequential sentence to an authoritative source? If the answer is no, the draft is unfinished. Build the verification trail now, preserve the original records, and give reviewers enough time to challenge fluent language. The next serious AI failure will probably look plausible before it looks false.