The King’s Good Servant, and God’s First

St. Thomas More’s final declaration of loyalty echoes today. Tradition often recalls More’s last words as, “I die the King’s good servant, but God’s first.” A biographical note published by the Holy See renders the line with a different conjunction: “and.”²

That note renders it “I die the King’s good servant and God’s first,” a subtle shift with profound implications. The word and implies harmony: that one can serve both earthly justice and divine truth simultaneously, with God’s law still holding primacy. The popular retelling with but, by contrast, suggests an inherent conflict, a choice between duties when push comes to shove. This paper takes the distinction between these renderings as a point of departure. In our age of artificial intelligence (AI) in law, are we striving to be both the legal system’s diligent servants and servants of a higher truth? Or is there an unspoken but, a creeping conflict, where technological advancement butts against moral integrity?

The following pages explore the intersection of AI and the practice of law with a jurisprudential and theological lens, treating the judicial Record as something sacred: a collective memory of the law’s proceedings, akin to a communal scripture of justice. And by further examining how AI tools challenge the sanctity of that Record through deepfake evidence and hallucinated citations, the question is asked, “Who watches the Watchmen?”³

In sum, this paper argues that to harness AI ethically in law, we must exercise disciplined, human judgment rooted in moral truth. We must be the profession’s good servants, and truth’s first. Only then can we protect the integrity of the Record in the age of intelligent machines.

In any courtroom, the Record is more than filed paperwork. It is the authoritative memory of what has transpired. There is a sanctity to this Record — a moral claim rather than a doctrinal one. Judge Cardozo located a kindred intuition in the law itself: because “law shall be uniform and impartial,” there “must be nothing in its action that savors of prejudice or favor or even arbitrary whim or fitfulness.” His guarantor was adherence to precedent, without which litigants could not have “faith in the even-handed administration of justice in the courts.” The Record is that faith’s evidentiary counterpart: what the law did, preserved.⁴ We alter court transcripts only through scrupulous, formal mechanisms because a single errant word can invert the meaning of a proceeding. To treat the Record as sacred is to treat truth as sacred. Yet today, AI’s growing power to fabricate realistic fictions poses an unprecedented threat to that truth. AI-generated deepfakes — hyper-realistic fake videos, images, or audio — have begun to seep into court proceedings, putting the integrity of evidence at grave risk.

Consider a 2025 watershed case in California. In Mendones v. Cushman & Wakefield, Inc., two pro se litigants submitted what appeared to be video testimony of a key witness. Judge Victoria Kolakowski sensed something was off with Exhibits 6A and 6C. Comparing them against authentic video of the witness, she identified a lack of facial expressions, a looping video feed, and mouth movements that did not match the words, and found both exhibits to be generative-AI products in which the witness saying exactly what the plaintiffs needed was not real at all. Exhibit 7, a photographic exhibit, was also materially altered. The judge described the witness’s speech and behavior in the fabricated videos as “vastly different” from the authentic recording. These were substantive falsifications, presented to the court as the witness’s own words and conduct. The submission of this falsified evidence struck at the heart of legal truth. As Judge Kolakowski put it, such misuse of generative AI “egregiously” violated the defendants’ and the court’s trust. She imposed a terminating sanction, dismissing the plaintiffs’ entire case with prejudice. In her order, she made clear that the court has “zero tolerance” for anyone attempting to pass off AI forgeries as real evidence. The message was sent loudly: deepfakes in the courtroom are no longer a speculative threat. They are here, and this court treated their submission as the intentional filing of false evidence warranting dismissal with prejudice.⁵

One Louisiana judge, Scott Schlegel, called the Mendones case a “warning shot,” noting that while these deepfakes were crude enough to be caught, rapidly advancing technology means the next forgeries may be indistinguishable from reality.⁶ If falsehood pollutes the judicial record, confidence in the justice system’s verdicts collapses. In response, courts and rulemakers are scrambling to build oversight and verification into the system.

In August 2025, the federal rulemaking process produced a proposal for a new rule,⁷ Federal Rule of Evidence 707, specifically to govern certain machine-generated evidence. That version, discussed here in the context of the paper’s original composition, is narrower than it first appears. It reaches machine-generated evidence offered without an expert witness that would be subject to Rule 702 if a human witness supplied it. In that circumstance the evidence may be admitted only if it satisfies Rule 702(a)–(d): the output must “help the trier of fact,” rest on sufficient data, use reliable principles, and apply those principles reliably to the facts. The rule addresses reliability, not the authenticity of every AI-generated photograph, video, or recording; the Committee left suspected deepfakes to Rule 901. Even so understood, it would force the proponent of expert-like machine output to prove its reliability rather than simply assume it.

Notably, the August 2025 version of proposed Rule 707 explicitly excludes “the output of simple scientific instruments”⁸ from its scope — the Committee sought to avoid unnecessary litigation over familiar instruments such as a thermometer or electronic scale. That exclusion would not dispense with otherwise applicable evidentiary requirements. But where expert-like output comes from a complex AI system and no expert is offered to vouch for it, the proposal would require the proponent to establish its reliability. The judiciary is effectively admitting we can no longer naïvely trust what our eyes and ears perceive from media evidence; we need new legal scaffolds to preserve truth in the age of AI deception. State legislatures are acting as well, such as in Louisiana’s Act 250 of 2025, which requires attorneys to exercise reasonable diligence in verifying evidence before offering it and to disclose when they knew, or should have known through that diligence, that it was false or artificially manipulated. The Act also establishes pretrial procedures for raising reasonable suspicions about opposing exhibits and disclosing known or reasonably suspected manipulation of one’s own exhibits; those procedures exclude demonstrative exhibits.⁹

Such measures attempt to bolster the walls around our sacred Record. They raise a fundamental moral point: knowingly presenting a deepfake in court as authentic evidence is not just another evidentiary issue; it is a form of lying. It contravenes the commandment against bearing false witness as surely as any perjured testimony. In a very real sense, then, the judicial Record demands a kind of reverence and truthfulness that has theological echoes. Just as Scripture in the religious sphere must not be adulterated, so the records of justice must remain uncorrupted by falsehood. AI may be a powerful tool, even a Godsend in some applications; however, if it tempts us to put “but” where “and” should be (to serve litigation strategy but not truth), then it becomes a stumbling block.

In the film Good Will Hunting, therapist Sean Maguire (played by Robin Williams) sits on a park bench and confronts the young prodigy Will with a harsh truth. Will is a genius who can cite obscure facts from books without limitation. “If I asked you about art, you’d probably give me the skinny on every art book ever written… But I’ll bet you can’t tell me what it smells like in the Sistine Chapel,” Sean says. Will has read volumes, but he has never stood beneath the ceiling. He can quote Shakespeare, “Once more unto the breach, dear friends”, but he has never been near a war, never held his best friend’s head in his lap and watched him gasp his last breath. And yet he acts as though he understands. The moral is clear: there is a chasm between information and wisdom, between data and understanding.¹⁰

That chasm is directly relevant to AI’s role in law. Today’s AI systems are like Will Hunting in some striking ways. They have consumed countless books, cases, and articles; they can regurgitate astonishing amounts of information and even write plausible essays on legal topics. But do they understand a client’s pain, a witness’s credibility, or the moral weight behind a legal rule? An AI can summarize Michelangelo’s biography, but it cannot smell the Sistine Chapel. It can scrape the text of thousands of precedents, but it has never felt the anguish of a wrongful conviction or the solemn duty of a juror deciding a life-or-death case. In theological terms, one might say AI lacks a soul. Pope Benedict XVI warned that technology becomes self-sufficient when too much attention is given to the “how” and not enough to the “why,” confining judgment within efficiency and utility and obscuring the moral dimension of what we do.¹¹ AI, for all its computational brilliance, tends toward exactly that narrowing unless we imbue its use with human wisdom.

The distinction between knowledge and wisdom has concrete implications. One concern gaining attention is AI “model collapse,” which an IBM explainer describes as declining performance when generative models are trained recursively on generated content. When generated material replaces or overwhelms the original training data, a model’s performance can deteriorate. Errors get amplified like a copy of a copy, and the model begins to “forget” the tails of the distribution, those uncommon, nuanced cases that don’t appear often in the AI-generated content. Over successive generations, such a model loses diversity and accuracy, producing increasingly nonsensical or generic output. The danger is not synthetic data as such, but a training process that progressively loses contact with the original distribution. IBM’s explainer warned that generative models “trained solely on their predecessors’ output” are beset by “irreversible defects” and “eventually become useless,” while the underlying study found that preserving access to original human-generated data substantially reduced the degradation. Other research shows that retaining and accumulating original data alongside synthetic data can avoid collapse in the settings studied. The danger still offers a striking metaphor: a snake eating its own tail.¹²

In the legal domain, this offers an analogy about compounding omissions: one AI summarizing another AI’s transcript summaries may carry forward errors while losing details. That is not itself the recursive training studied in model-collapse research, but it illustrates the importance of returning to the underlying record. The law thrives on nuance, that rare case, the unexpected fact pattern, the creative argument. A model that loses rare features of its training distribution may omit the very details that justice might hinge on. We might get a fluent answer from the machine, but like Will Hunting’s textbook knowledge, it could be hollow. A legal brief written by a generative AI tool might sound like an impressive synthesis of precedent, while subtly missing the point that a seasoned attorney or judge would catch in a heartbeat. The human lawyer or judge must remain the discerning soul who can tell surface knowledge from substance. An AI can inform, but it takes human understanding to interpret.

Recall that in the Book of Genesis, God grants humans dominion over creation but not a license for negligence. We are to “till and keep” the garden, which suggests responsibility and care. In the garden of justice, AI may be a new tool for cultivation, but the “keeping” remains our duty. Lawyers and judges must cultivate a stance of respectful skepticism toward AI. Use it, yes; be enlightened by its rapid research, certainly; but verify every suggestion against the real world and one’s own seasoned understanding.

Even as judges grapple with deepfakes in evidence, another AI phenomenon has been bedeviling the legal profession: the plague of hallucinated citations. This refers to the tendency of generative AI tools to confidently produce references to cases, statutes, or other sources that simply do not exist. These tools don’t lie with malice, but they generate text based on patterns, and if that pattern suggests a string of words that looks like a legitimate citation, the AI will merrily output it. The results can be eerily plausible. A fake case name, a docket number, even a snippet of “quoted” language that sounds on-point, yet it’s pure fiction. For an unsuspecting lawyer or client, such hallucinated case law is a trap set by the AI’s lack of understanding. And in 2023–2025, we have seen that trap catch multiple attorneys in its jaws, with embarrassing and even career-threatening consequences.

The saga that first drew global attention was the now-infamous Mata v. Avianca, Inc. incident in New York. On March 1, 2023, Peter LoDuca filed an affirmation in opposition to Avianca’s motion to dismiss Roberto Mata’s personal injury action. Steven A. Schwartz, his colleague at Levidow, Levidow & Oberman P.C., had researched and written the affirmation; LoDuca signed and filed it. The opposition included six nonexistent decisions. When Avianca’s counsel and the court could not locate the authorities, the truth came out: Schwartz had relied on ChatGPT for legal research, and the AI had fabricated the citations.¹³

In one example, the affirmation cited a case called Varghese v. China Southern Airlines¹⁴ with a full citation and a compelling parenthetical summary, all of it fake. U.S. District Judge P. Kevin Castel was not amused. In a stinging sanctions order in June 2023, he noted that while there is “nothing inherently improper” about using AI as an aid, lawyers have an ethical “gatekeeping role” to ensure the accuracy of their filings.¹⁵ The attorneys in question had utterly failed in that duty and worse, when first questioned, they doubled down, insisting the cases were real (even submitting surreal ChatGPT-generated excerpts that purported to be the text of the fake opinions!). The court imposed a single $5,000 penalty jointly and severally on Schwartz, LoDuca, and Levidow, Levidow & Oberman P.C., and ordered them to provide the sanctions order to Mata and to the judges falsely identified as authors of the fabricated opinions.¹⁶ Judge Castel’s message, now echoed by many others, was clear: if you use AI, you must verify its output. Blind reliance on a chatbot is no excuse for promulgating falsehood in court. Blind reliance can breach duties of competence and reasonable inquiry; knowingly presenting falsehoods, or failing to correct known falsehoods, also implicates one of the most sacred duties of an attorney, candor to the court.

One might have hoped that the Mata debacle would be a singular cautionary tale. Surely, after that media storm, no attorney would repeat such an error. And yet, in the months that followed, hallucinated citations continued to surface in courtrooms across the United States. The problem, however, is not confined to the U.S. By mid-2025 the problem had gone global. Matthew Lee reported in Counsel that more than fifty AI-hallucination incidents had been reported internationally in July 2025 alone.¹⁷ They spanned jurisdictions and levels of experience: not just junior lawyers or solo practitioners, but seasoned attorneys at reputable firms. Perhaps the most telling example was Johnson v. Dunn.¹⁸

In that case, experienced attorneys from a major firm filed two discovery motions containing five fabricated citations. One supervising lawyer had used ChatGPT to obtain the citations and inserted them through revisions without verifying them; the lawyer who drafted, signed, and filed the motions incorporated those revisions without independently reviewing the added authorities; and the partner whose name appeared in the signature block did not check them either. Judge Manasco characterized this conduct as “serious misconduct that demands a serious sanction,”¹⁹ explaining that fabricating legal authority constitutes an “extreme dereliction of professional responsibility”²⁰ that threatens “the fair administration of justice and the integrity of the judicial system.”²¹ The court found the attorneys’ actions to be “particularly egregious,”²² driven by “recklessness in the extreme,”²³ and “tantamount to bad faith,”²⁴ noting their “troubling indifference to the veracity”²⁵ of their filings and their “utter disregard for the truth”²⁶ of statements presented to a federal tribunal.

Because these attorneys had access to proper research tools, repeated warnings about AI hallucinations, and clear internal firm policies they ignored, the court rejected any suggestion that this was excusable error. It imposed sanctions far beyond the modest fines seen in earlier AI cases, holding that “no lesser sanction will serve the necessary deterrent purpose.”²⁷ The court publicly reprimanded the lawyers, disqualified them from further participation in the case, ordered them to provide the sanctions order to their clients, opposing counsel, and the presiding judge in every pending case in which they were counsel of record, required distribution to every attorney in their firm, and referred them to state licensing authorities. Their failure to conduct even the “most cursory of investigations,”²⁸ combined with each signer’s personal responsibility for a filing bearing his name, supported the court’s findings that the misconduct went beyond mere recklessness and was tantamount to bad faith.²⁹

In contrast, courts have credited admissions, apologies, and corrective safeguards when calibrating sanctions — though not by treating the conduct as innocent. In Versant the court found counsel “at least reckless” and still imposed fees and costs, mandatory AI-ethics continuing legal education, and fines of $1,000 and $500. In Ramirez the court inferred subjective bad faith from counsel’s complete failure to verify four nonexistent cases, yet set the sanction at $1,000, the low end of the range, because of her prompt admission, apologies, and new internal review.³⁰ Some, like King David before the prophet Nathan, repent; others double down and are judged accordingly.

What moral, then, do we draw from this rash of AI-related sanctions? First, it reinforces the timeless principle of professional responsibility: lawyers owe a duty of competence and truthfulness to their clients and the courts. If a lawyer delegates research or drafting to a junior attorney or a paralegal, the lawyer must supervise and ultimately vouch for the work. The same must hold true if the lawyer “delegates” a task to an AI. Under the Model Rules of Professional Conduct, using an AI tool doesn’t absolve a lawyer of blame for errors; rather, it increases the onus to double-check the output. Rules 5.1 and 5.3 govern a lawyer’s supervision of other lawyers and of nonlawyer people, not an AI system as a legal person. But the ABA’s Formal Opinion 512 explains how those duties bear on generative AI: firms should adopt policies, train and supervise the people who use these tools, and exercise reasonable diligence about outside AI providers.³¹ In plainer terms: trust, but verify — or better yet, don’t trust until verified.

At the root of all this is a call to character. The legal profession has long held itself out as a learned calling, emphasizing integrity, diligence, and service. AI doesn’t change those core values; it tests them. Will lawyers succumb to the sin of sloth, letting a machine do their thinking and accepting its answers on faith? Will they give in to vanity, trusting that a high-tech tool makes them infallible? Or will they approach this innovation with prudence and honesty, integrating it into their work and keeping truth’s service first? The answer will define the ethos of the next generation of lawyers.

The phrase “Quis custodiet ipsos custodes?” — who will guard the guards themselves — was posed by the ancient poet Juvenal in a very different context, yet it resonates in this moment of jurisprudential evolution.³² We have built machines to aid our judgment; now we must guard against the machines misguiding our judgment. This is not a counsel of despair about AI, but rather a call to responsibility. In Catholic thought, human beings are created imago Dei, in the image of God, endowed with reason and free will, capable of discerning good and evil. Our technologies, no matter how advanced, do not carry that divine imprint. Therefore, the duty to uphold truth and justice can never be abdicated to them. The rhetorical dance between “and” and “but” in these renderings of More’s statement encapsulates the tension every lawyer faces in a new form today. We want to say we are good servants of innovation, and God’s first — that embracing AI can go hand in hand with upholding our highest ethical duties.

In many ways it can. AI, used prudently, can indeed democratize legal services, reduce drudgery, and even flag human biases (an AI reviewing sentencing disparities might prompt reform, for example). Such benefits can be in service of justice, aligning with the moral law’s demands for fairness and care of the vulnerable. This is the “and”, the hopeful vision where our temporal tools further eternal values. Yet we must be candid that there is also a “but” looming: moments where convenience or competitive pressure might seduce legal professionals into cutting corners; into serving the client or court but forgetting the truth. An advocate might be tempted to let a false but helpful AI-generated point slide by, rationalizing that it’s what the client wants to hear. A firm might prioritize speed of output over rigorous fact-checking, effectively saying we serve winning cases but not always the whole truth. These are the forks in the road where one must choose “God’s first” in the and sense, the courageous choice to put integrity above immediacy.

For many lawyers, the immediate test is professional courage in ordinary decisions rather than martyrdom. In the micro-choices to verify a source, to disclose an AI’s involvement when required, or to reject a client’s request to use a dubious AI-generated strategy, the spirit of More lives on. It is the spirit of professional responsibility animated by personal virtue. The Book of Proverbs counsels, “Buy truth, and do not sell it; buy wisdom, instruction, and understanding.” In the marketplace of cutting-edge legal AI, we must not sell out truth for speed or efficiency. Rather, we buy, that is, invest in wisdom and understanding alongside these new scribes. If we do so, we can truly say we are the legal system’s good servants. If we falter, if we let but creep in to justify lies or laziness, then all our AI marvels will become a new Tower of Babel, impressive in height, perhaps, but founded on confusion and destined to fall.

The choice, mercifully, is ours. With disciplined, ethically informed judgment, we can harness AI to enhance justice without degrading the soul of the law. The Record will remain sacred and reliable. The watchers (human and artificial) will each play their part. And the collective memory of our law, unmarred by falsity, will continue to guide and inspire those who seek what is right. In the final analysis, what is at stake is nothing less than the moral credibility of the legal system. If we are vigilant and virtuous, the law will continue to be a reflection (albeit imperfect) of the divine order, where, even aided by machines, the human spirit strives to do the common good, to speak the truth, and to serve the people. And in that endeavor, may we all be found faithful.