Introduction: A Troubling Trend
With increasing frequency, I have been seeing Colorado judges enter case-management orders that include the following language:
It is the expectation that all motions, briefs, and other written submissions filed with the court are the original work of the attorney or attorneys who sign the pleading. If Chat GPT or other AI is used to generate any written product filed with this court, a notice shall appear on the first page of the filing indicating that AI was used to generate all or part of the filing, as well as the specific portions of the filing (e.g., page number and paragraphs or lines) which contain the AI-generated content.
I understand the impulse behind this language. After the widely reported Mata v. Avianca debacle—in which two New York attorneys were sanctioned for submitting a brief citing six entirely fabricated cases generated by ChatGPT—courts nationwide scrambled to address perceived risks of generative AI in legal filings. The impulse is understandable, even commendable. But I believe the specific remedy embodied in the order quoted above is unnecessary, practically unworkable, and potentially harmful—both to clients and to the very attorneys it purports to regulate.
I write as a partner at Fox Rothschild with twenty-one years of trial experience and a substantial Colorado litigation practice. I am also a frequent and enthusiastic user of AI legal tools—specifically Harvey, a closed, enterprise AI platform designed for lawyers. I do not work in tech; I am a working trial lawyer. And from that vantage point, I believe the order quoted above gets it wrong in several important respects.
AI and the Cost of Litigation: A Generational Opportunity at Risk
Everyone complains that litigation is too expensive. For two decades, I have watched litigation costs climb relentlessly—driven by the proliferation of electronically stored information, the explosion of regulatory complexity, and the irreducible labor intensiveness of legal research, drafting, and discovery review.
AI is the first tool I have encountered in my career that has real potential to reverse that trajectory. Used properly, AI can help lawyers research more efficiently, draft more quickly, review documents more thoroughly, and deliver legal services at lower cost. That is not hypothetical; it is my lived experience.
Orders like the one quoted above threaten those gains. First, they risk discouraging lawyers from using AI at all — either because lawyers fear the stigma of disclosing AI use or because they fear inadvertent noncompliance with vague labeling requirements. A lawyer who decides not to use AI because of such an order is a lawyer whose client pays more for the same work product. Second, even for lawyers who continue using AI, the order imposes substantial compliance costs. A lawyer who must attempt to parse, at the page-and-line level, which specific words “came from” AI versus human editing is spending time on a compliance exercise that adds zero value.
The Duke Law Judicature study of judicial AI orders recognized precisely this concern, noting that such orders “may discourage the use of technology that might otherwise increase access to justice and reduce costs.”
The Vagueness and Practical Impossibility Problem
Set aside for the moment whether AI labeling is a good idea in principle. The specific order at issue demands something that may be practically impossible to provide: identification of “the specific portions of the filing (e.g., page number and paragraphs or lines) which contain the AI-generated content.”
Here is how I — and many lawyers I know — actually use AI in practice: I use AI to generate a first draft of a brief or motion. I then read that draft carefully. I edit it. I delete sections. I add new sections. I rewrite sentences. I reorganize arguments. I add citations from my own research. I revise, again, after reflection. By the time I sign and file the document, the final product is mine—the product of my judgment, my analysis, and my professional expertise. But it began as an AI-generated draft.
Which “page number and paragraphs or lines” of that final filing “contain the AI-generated content”? All of them? None of them? Some indeterminate fraction? I genuinely do not know, and I suspect neither would any other lawyer working this way. The order demands a level of granularity that simply does not correspond to how iterative, collaborative drafting works.
This leads to my central question: How is using AI to generate a first draft of a brief, which I later edit, adopt, and ultimately sign under Rule 11 penalties, any different from having an associate prepare a first draft that I later edit, adopt, and sign under the same Rule 11 penalties?
No court has ever required a signing attorney to disclose which sentences of a brief were drafted by the partner versus the associate versus the summer clerk. In each of those scenarios, the legal system relies on a single, elegant safeguard: the signing attorney’s certification, under Rule 11, that the filing is well grounded in fact and warranted by law. The lawyer who signs the pleading is on the hook for its contents. That safeguard works regardless of who—or what—prepared the first draft.
Rule 11 Already Provides the Remedy
Under Colorado Rule of Civil Procedure 11(a), the signature of an attorney on a pleading “constitutes a certificate by him that he has read the pleading; that to the best of his knowledge, information, and belief formed after reasonable inquiry, it is well grounded in fact and is warranted by existing law or a good faith argument for the extension, modification, or reversal of existing law, and that it is not interposed for any improper purpose.” If a lawyer signs a pleading in violation of Rule 11, the court “shall impose … an appropriate sanction.”
This is not a theoretical safeguard. It is already being enforced in the AI context. In Mata v. Avianca, the court sanctioned the offending attorneys under Rule 11—precisely because they failed to read and verify their AI-generated filing. Closer to home, in November 2023, a Colorado Presiding Disciplinary Judge suspended an attorney who filed a motion containing fictitious case citations found through ChatGPT and never verified. The existing system worked. No categorical AI-labeling mandate was needed to identify or punish the misconduct.
The ABA’s Formal Opinion 512 reached essentially the same conclusion, grounding lawyers’ obligations regarding generative AI in existing Model Rules — competence, confidentiality, candor, supervision, and meritorious claims—rather than proposing new categorical disclosure requirements.
If a judge believes a lawyer is not reading or verifying AI output before signing and filing it, the appropriate response is to enforce Rule 11 against that lawyer, not to impose a categorical labeling requirement on every lawyer in every filing. We already have Rule 11.
Privilege and Work-Product Concerns
Beyond the practical and policy objections, I have a more fundamental concern: AI-labeling mandates risk invading attorney-client privilege and work-product protections.
The emerging case law on AI and privilege is instructive — and unsettled. In United States v. Heppner, the Southern District of New York held that a criminal defendant’s use of the consumer version of Anthropic’s Claude AI destroyed any privilege claim, because Claude is not an attorney, and Anthropic’s privacy policy permitted data collection and third-party disclosure—eliminating the confidentiality necessary to support privilege.
But other courts have reached different conclusions. In Warner v. Gilbarco, Inc., the Eastern District of Michigan held that AI systems are “tools, not persons,” and that disclosure to an AI tool is not disclosure to an adversary — meaning materials created using AI can still qualify for work-product protection. And in Morgan v. V2X, Inc., a Colorado federal court agreed with Warner and extended that reasoning, finding that AI use does not automatically waive work-product protection and that substantive AI prompts and outputs constitute protected work product reflecting mental impressions and legal strategy.
The critical distinction running through this case law is between open, consumer-facing AI platforms (like the public versions of ChatGPT or Claude, whose terms of service and data practices can defeat confidentiality) and closed, enterprise AI systems (like Harvey, which operates under contractual confidentiality protections and does not train on or publicly disclose client inputs). The Morgan court itself crafted an AI-specific protective-order provision recognizing this distinction: no party may input confidential information into an AI platform unless the provider contractually prohibits using inputs for training or disclosing them to third parties.
Here is why this matters for the AI-labeling order at issue: I frequently receive AI-produced work product from clients who use enterprise AI systems under contractual confidentiality protections. When I incorporate that material into a filing and am required by court order to identify precisely which “pages, paragraphs, or lines” contain “AI-generated content,” I may be forced to reveal which portions of my filing derive from privileged client communications or protected work product. The very act of parsing “AI-generated” from “human-generated” content is arguably revelatory of legal strategy and thought process.
Colorado’s Own Rules Show a Better Path
Colorado already has a more thoughtful model. Effective January 8, 2026, the Colorado Supreme Court adopted AI-specific amendments to the Colorado Rules of Professional Conduct—making Colorado the first jurisdiction in the country to do so. Those amendments are deliberately technology-neutral and measured. A new Scope section provides that “[t]echnology, including artificial intelligence and similar innovations, plays an increasing role in the practice of law, but that role does not diminish a lawyer’s responsibilities under these Rules.” Revised comments to Rule 1.1 direct lawyers to keep abreast of the benefits and risks of relevant technologies as part of the duty of competence, and clarify that “[r]eliance on technology does not diminish the lawyer’s duty to exercise independent judgment in the representation of a client.”
This approach—reinforcing existing duties rather than creating new format-specific labeling mandates—is the right one. It tells lawyers: you remain responsible for everything you file, regardless of what tools you used to create it. That is Rule 11 in different words. It is technology-neutral, it is workable, and it does not create the practical, privilege, and chilling-effect problems inherent in page-and-line AI labeling.
Even the AI-specific standing order entered by Judge Nina Wang in the U.S. District Court for the District of Colorado—a federal order explicitly addressed to generative AI use—requires only a certification regarding AI use and accuracy verification, not the granular page-and-line identification demanded by the state-court order at issue. Judge Wang’s order expressly states that all filings remain subject to Federal Rule 11 and applicable ethical rules—confirming that the existing framework is the operative safeguard. The state-court order at issue is thus an outlier, more burdensome than even comparable Colorado federal practice.
Conclusion: Trust the Tools We Already Have
I do not suggest that courts should ignore the challenges posed by generative AI. The Mata case was a real scandal, and the Colorado disciplinary suspension shows that AI misuse has real consequences. Lawyers must use AI responsibly, and courts must hold them accountable when they do not.
But Rule 11 already provides that accountability. Professional discipline already provides that accountability. Colorado’s new Rules of Professional Conduct already reinforce that accountability in AI-specific terms. What the order at issue adds is not accountability — it is bureaucracy. It imposes a vague, potentially impossible compliance obligation on every lawyer in every filing, raises costs for clients, chills the adoption of technology that could make justice more accessible, and risks forcing disclosures that invade privilege and work product.
My recommendation to courts considering AI-specific orders is simple: look to Colorado’s 2026 amendments to the Rules of Professional Conduct as a model. Reinforce existing duties. Hold lawyers accountable under Rule 11 when they fail to verify their filings. Sanction and discipline those who submit fabricated material. But do not require every lawyer, in every filing, to attempt the impossible task of parsing which specific sentences “came from” a tool versus from the lawyer who used that tool, edited its output, and signed the final product under penalty of sanctions.
After twenty-one years of trial practice, I can say with confidence: AI, used properly, makes me a better, more efficient advocate for my clients. Orders like the one at issue do not protect the court or the public from bad lawyering. They just make good lawyering harder — and more expensive.
Henry Baskerville is a partner at Fox Rothschild. He has practiced as a trial attorney for twenty-one years, with a substantial litigation practice in Colorado. He is a frequent user of Harvey AI and an advocate for responsible AI adoption in legal practice.