In re Paula M. Miller: Texas Court of Appeals Warns of AI-Hallucinated Citations in Mandamus Petition | Advocate Prakhar

Case at a Glance

Background

Paula M. Miller filed an original proceeding on petition for writ of mandamus before the Texas First Court of Appeals. The petition was denied without further discussion of its merits. What prompted the Chief Justice to write separately was the quality of the legal citations in the mandamus petition: the attributions and quotations to case law had characteristics that strongly suggested AI-generated hallucinations — language that sounds authoritative but does not correspond to what the cited cases actually say.

Mandamus petitions in Texas are procedurally demanding: they require showing a clear abuse of discretion by the lower court and no adequate remedy at law. AI-generated arguments for mandamus relief are likely to be particularly unreliable because the specific factual and procedural context of each mandamus petition is highly individualised, making generic AI outputs poorly suited to establishing the elements the petition must prove.

The AI Issue

Chief Justice Adams’ concurrence frames the AI issue in terms of the duty of candour to the court under Texas rules. Filing a document with fictitious or misleading citations — whether generated by AI and not checked by a human, or otherwise — is a “serious breach of candour.” The Chief Justice quoted the foundational statement from United States v. Hayes (E.D. Cal. 2025): “Citing nonexistent case law or misrepresenting the holdings of a case is making a false statement to a court. It does not matter if generative AI told you so.” The concurrence then articulated a specific obligation: “Always ‘trust and verify’ the accuracy and validity of your results.”

What the Court Decided

• Mandamus petition denied without further discussion of the underlying merits.
• Chief Justice Adams wrote a concurring opinion specifically to address the AI-hallucination problem in the petition.
• Formal warning issued: counsel who use AI for legal research must “trust and verify” — confirming that cited cases actually contain the quoted language and stand for the proposition cited.
• Trusting AI for legal research without verifying violates counsel’s ethical duty of candour to the court.
• Future offenses will result in striking the brief and reporting counsel to the State Bar of Texas.

“Trusting AI for legal research without verifying that the cases actually contain quoted language and stand for the proposition for which they are cited violates counsel’s ethical duty to this Court. And that unfortunately will require the Court to take the necessary corrective action including, but not limited to, striking the offending brief and reporting counsel to the State Bar.”

— Chief Justice Terry Adams, Texas First Court of Appeals, April 23, 2026

The India Angle

Indian Law Equivalent

In India, writs of mandamus are issued under Article 226 of the Constitution (High Court) and Article 32 (Supreme Court) to direct public authorities to perform their legal duties. The evidentiary and legal standards for a writ petition are high: the petitioner must demonstrate a clear legal right to the relief sought and a corresponding clear legal duty on the respondent. AI-generated mandamus petitions that fabricate or misrepresent the case law establishing these rights and duties would be caught quickly by court staff or the opposing party — Indian High Courts have extensive experience reviewing writ petitions and can identify generic or factually disconnected legal arguments.

Bar Council Rules

BCI Rule 49 (duty against misstatement) and Rule 52 (accurate citation) apply directly. The Chief Justice’s “trust and verify” formulation captures the essence of these rules in a memorable phrase that Indian advocates should internalise. Trust (using AI as a starting point for research) is permissible; verify (checking every cited case against authenticated primary sources) is obligatory. The failure to verify is where the professional obligation is violated.

Practical Advice for Indian Advocates

• Mandamus and certiorari petitions require highly specific factual and legal justification for extraordinary relief — AI tools generate generic arguments poorly calibrated to the specific facts of the case; the verify step is especially important here.
• When a Chief Justice or senior judge writes separately to warn about your citations, that warning is on the public record and will be seen by any court where you subsequently appear — treat a concurrence warning with the same seriousness as a majority opinion sanction.
• The “trust and verify” mnemonic from this concurrence is a useful professional practice standard: use AI to identify potentially relevant authorities (trust), then independently verify each authority before citing it (verify) — never rely on AI’s representation of what a case says without reading the relevant section.

Quick Takeaways

• A warning issued in a judicial concurrence carries the same prospective force as one in a majority opinion — courts track their own prior statements when assessing future violations.
• “Trust and verify” is the governing standard for AI use in legal research: using AI as a starting point is permissible, but submitting AI output without independent verification is a violation of the duty of candour.
• Texas is the latest state in a growing national consensus — AI hallucinations in court filings will result in striking briefs and State Bar referrals for counsel who fail to verify their citations.

Deep Dive: The Concurrence as Warning — Judicial Communication in the AI Era

The In re Paula M. Miller concurrence is a fascinating example of how appellate judges are using their separate writing power to address AI-hallucination problems that do not rise to the level requiring a formal sanctions order but that the court believes must be addressed on the record. Chief Justice Adams chose not to vote to sanction counsel in this proceeding — perhaps because the petition was denied without discussion and the warning itself was deemed sufficient for a first offense. But the choice to write separately ensures that the warning has precedential value: it is published, citable, and will appear in any future case where this counsel — or any other Texas attorney — commits the same offense.

This use of the concurrence as a vehicle for forward-looking judicial guidance is consistent with the practice in other jurisdictions. The Florida Fourth DCA’s Gouveia concurrence (which called AI reliance “unacceptable, and sanctionable”), the Texas First Court of Appeals’ Miller concurrence, and similar writings from other courts collectively build a body of judicial statements that, even without carrying majority precedential weight in the traditional sense, establish the norms against which future conduct will be assessed.

For Indian advocates, the lesson is that judicial communication about professional conduct takes many forms beyond formal sanctions orders. High Court and Supreme Court judges in India frequently include observations and dicta in their judgments about counsel conduct, the quality of legal arguments, and the standards expected in their courts. These observations — even when embedded in judgments deciding other issues — are read by the Bar and by lower courts and shape professional expectations. An advocate who ignores such judicial observations does so at professional peril.

The “trust and verify” standard articulated by Chief Justice Adams is deceptively simple but captures an important insight about the specific nature of the AI reliability problem. AI systems are trustworthy in some respects — they can efficiently identify potentially relevant legal areas, generate plausible arguments, and structure complex documents. They are unreliable in a specific and predictable way: they confabulate specific factual details, including case citations, quoted passages, and specific holdings. The verification step is precisely calibrated to catch the type of error AI makes, while preserving the efficiency benefits AI provides in the drafting and research process.