“The AI Did It” Is Not a Defense

The next time a Texas lawyer types “/s/” beneath a court filing, the signature may carry more weight than it does today. Under a proposed rewrite of Texas Rule of Civil Procedure 13, the signer would represent that the signer “reviewed the filing and verified its accuracy,” including that its “legal citations, quotations, and holdings are not fabricated.”

The proposal is aimed principally at the improper use of artificial intelligence. But its most sensible feature is that it does not ask what software touched the draft. It asks whether the final filing is accurate. The court receives a certification from a human being, not an explanation about a machine.

Texas Would Regulate the Filing, Not the Tool

The Supreme Court of Texas gave the amendments preliminary approval on June 26, 2026. They are not yet effective. Written comments are due September 1, 2026, at rulescomments@txcourts.gov, and the Court says it may revise the text before issuing a final order. The Court currently expects an October 1, 2026 effective date.

The proposed rule neither bans AI nor requires a lawyer to disclose its use. The accompanying comment says verification matters “regardless of whether artificial intelligence was used.” It also draws an important boundary: the rule is not intended to govern real disputes over the interpretation, application, or persuasive force of existing authority. A weak argument is not the same thing as a fabricated authority.

A companion amendment to Texas Rule of Appellate Procedure 9.1 would apply a similar certification to appellate filings and expressly add “record citations” to the items that cannot be fabricated. That is an operational clue. Verification cannot stop after confirming that a case name exists; it must reach the quotation, the proposition, and the record.

The sanctions structure is measured. The proposal preserves a presumption of good faith, requires notice and a reasonable opportunity to respond, and requires proportionality. It makes sanctions discretionary and directs courts to use lesser sanctions before contempt, striking, or dismissal unless the court explains why lesser measures would not work. This is not a strict-liability trap. But it would turn source verification into an express representation made by the signer.

Fake Cases Were Only the First Wave

The first famous warning was Mata v. Avianca, Inc., where lawyers submitted nonexistent judicial opinions containing fake quotations and citations generated by ChatGPT, then continued to stand by them after the opposing party and the court questioned their existence. The court imposed a $5,000 penalty and required notices to the client and to judges falsely named as authors. Those errors were spectacular, but comparatively easy to detect: look for the case, and there is no case.

The Fifth Circuit’s 2026 decision in Fletcher v. Experian Information Solutions, Inc. exposed the harder problem. Counsel’s reply brief used real case names and citations but attributed language and propositions to them that they did not contain. The court identified 16 fabricated quotations and five additional serious misrepresentations of law or fact. It found that counsel used generative AI to draft a substantial portion, if not all, of the brief and failed to verify it.

The Fifth Circuit described this newer failure as “false quotes or statements of law attributed to real cases, rather than the more easily recognizable fake cases.” That is the real verification problem. A citation checker that asks only whether the case exists will miss the most convincing hallucination: a real authority wearing invented language.

The court imposed a $2,500 sanction under Federal Rule of Appellate Procedure 46(c) and its inherent authority. The response to the error also mattered. The court said that if counsel had accepted responsibility and been more forthcoming, it likely would have imposed a lesser sanction. The lesson is twofold: verify before filing, and if an error escapes, correct it promptly and candidly.

There is a useful irony in Fletcher. The Fifth Circuit had considered an AI-specific rule that would require filers to certify either that no generative AI was used or, if it was used, that a human checked the generated text. It declined to adopt that rule because existing sanctions rules were sufficient. Texas’s proposal splits the difference: it requires no AI disclosure, but it makes the human verification duty express and tool-neutral. That is a durable approach. Today’s software name will age quickly; the duty to verify will not.

The Rewrite May Reach Beyond AI

The AI language will draw the headlines, but it is not the only change that deserves scrutiny. Current Rule 13 couples groundlessness with bad faith or harassment. Proposed subsection (c)(2) appears to separate those concepts, making the signer represent that a filing is not groundless, not brought in bad faith, and not brought for harassment. If read that way, groundlessness alone could become inconsistent with the certification even without bad faith or harassment.

The proposed comment does not discuss that possible shift. It may be intentional, or it may be a drafting artifact. Either way, it is too important a change to arrive by punctuation. Texas lawyers commenting on the proposal should examine the sanctions threshold as carefully as the AI sentence.

What Verification Should Look Like

The proposal does not prescribe a workflow, and it should not. But a reasonable verification protocol is not mysterious. Search results, headnotes, AI summaries, and draft memoranda are leads. The source itself is the authority.

Open the Authority

Confirm that every cited decision exists and that the court, date, reporter, docket number, and disposition match the filing.

Check the Quotation

Compare quoted text word for word against the opinion. Recheck brackets, ellipses, and any pinpoint reference.

Check the Proposition

Read enough context to confirm that the authority actually supports the sentence. A party’s argument, dissent, headnote, dictum, or case decided under a different procedural standard is not the holding.

Check the Record

Match each factual assertion to the cited transcript, exhibit, declaration, or docket entry. The proposed appellate rule makes this obligation express.

Assign Human Responsibility

Identify the lawyer who performed the final citation and record check. A workflow owned by everyone is usually owned by no one.

The Signature Still Belongs to the Lawyer

AI can accelerate research, organize a record, test an argument, and improve a draft. It can also state fiction with the confidence of a senior partner who has not yet opened the case. The proposed rule does not confuse usefulness with trustworthiness.

Texas’s proposal remains preliminary. Comments are due September 1, 2026, and the Court expects the amendments to take effect October 1 only after a final order. The bar should use that comment period to refine both the verification language and the broader sanctions threshold.

But the central point is already sound. AI may help write the brief. It cannot sign it. The lawyer can, and that is the point.

Primary Sources

Supreme Court of Texas, Misc. Docket No. 26-9054 (June 26, 2026), preliminarily approving the proposed amendments to Texas Rule of Civil Procedure 13 and related rules.
Texas Rules of Civil Procedure, current Rule 13.
Fletcher v. Experian Information Solutions, Inc., No. 25-20086 (5th Cir. Feb. 18, 2026).
Mata v. Avianca, Inc., 678 F. Supp. 3d 443 (S.D.N.Y. 2023).