The Empire of Strings

Peter LoDuca had known his colleague for more than twenty-five years. When Steven Schwartz gave him a submission to file in federal court, he read it for flow and grammatical errors. He did not read the decisions it cited. Schwartz had done the research; LoDuca, who was admitted to practice in that court, signed and filed the paper. The division of work was ordinary enough to need little explanation until the court began asking where the decisions could be found.¹

The client, Roberto Mata, had sued an airline after a serving cart allegedly struck his knee during a flight. The airline argued that his action was out of time. Schwartz sought authorities for an answer and used ChatGPT, which supplied descriptions and excerpts of cases that did not exist. These entered the opposition filed on March 1, 2023. Two weeks later, Avianca's lawyers reported that they could not locate most of the cited decisions, and that several they could find did not support the propositions for which they had been offered. In April, the judge ordered copies produced. What followed was another signed submission, this time attaching supposed extracts from the missing opinions.²

There were opportunities to stop. The opposing brief was one; the court's orders were others. By the sanctions hearing, the difficulty was no longer confined to the original research. It included what each lawyer had done after being given reason to doubt it, what each had represented himself as knowing, and how their accounts changed under examination. Judge P. Kevin Castel found subjective bad faith, assessed the two lawyers separately, and imposed sanctions on both and their firm. The familiar summary—that a machine invented cases and lawyers were fined—leaves out much of the human activity for which the fine was imposed.³

It also leaves out an institutional success. The invented opinions did not become precedents. An adversary searched for them; the court searched and asked for evidence; the account of their production was eventually brought into the record. The machinery was expensive and belated, but it distinguished a filed assertion from the law the assertion claimed to report. Any argument about what computation has made impossible must begin by noticing what these people were still able to do.

The output nevertheless acquired consequences without being accepted as law. Someone had to answer the submission, examine its sources, issue orders, and attend to explanations that turned out to require explanations of their own. The client had entrusted a claim to his lawyers and received a controversy about their conduct. A document can enter an institution's business without succeeding in the purpose for which it was submitted. That is already a considerable power to give a piece of writing.

Under another name

There was no shortage of names. The fabricated opinions had case titles, citations, judges, and apparent courts of origin. One supposed Eleventh Circuit decision carried a docket number belonging to a different proceeding. Its attributed judicial panel included a judge of the Fifth Circuit. The court's clerk confirmed that it was not an authentic ruling. These were available particulars against which the assertion could fail, although their presence had first helped give it the appearance of a document one need only locate.⁴

A citation is an economical instruction to look elsewhere. The economy depends on there being an elsewhere, and on the reader being able to distinguish it from a second performance of the same assertion. In this case, the request for the source elicited more material purporting to be the source. That material was not a damaged copy of a real opinion. Greater fidelity in transmitting it would have preserved the invention more accurately.

The decisive addition was supplied outside the generated text. A lawyer took the material into a submission, and another lawyer put his name to it. LoDuca's confidence rested on a long professional relationship. That confidence was not absurd merely because it was misplaced here; much competent work depends upon people accepting work done by colleagues. But the particular act of filing carried duties that familiarity could not discharge. The court found no evidence that he had asked about the nature and extent of the research or examined the cited authorities. He had attended to the prose at the point where the institution required him to attend to its grounds.¹

This is where the distinction between a sentence and an undertaking becomes useful, provided it is not made too grand. Austin had already directed attention to the procedures, persons, and circumstances through which words acquire conventional force. He also distinguished an act that fails to occur from an act performed insincerely or badly. Bad advice remains advice; an undertaking does not necessarily vanish because its author had no intention of honoring it.⁵ The distinction allows a rather unsentimental view of institutional language. The act may have taken place, and precisely for that reason its defects can become someone's responsibility.

The lawyers' submission was an act in this sense even though the supposed precedents were fictions. Its authority to place a position before the court came from the people and procedures through which it was filed. Its entitlement to persuade depended on something else. Nothing in the act of signing could make a nonexistent judgment exist, and nothing about the judgment's nonexistence erased the act of presenting it. Both facts were needed to understand what had happened.

It would therefore be misleading to say that the generated words lacked all binding force until someone made them true. They were false and had already implicated the people who adopted them. Nor was the remedy to discover a more solemn way for the machine to assert its sincerity. The court had lawyers before it, a rule governing their conduct, and a record of what they had done. It could proceed without deciding whether the producer of the first text could believe a proposition or break a promise.

Responsibility need not reside where the wording originated. An institution can compose a notice from other people's work, or authorize a process to issue notices within settled limits. Whether a particular notice is effective, mistaken, unauthorized, or actionable depends on that arrangement. A rule that attaches consequences to issuance can cover words nobody read individually before they were sent. The institutional question is what has been authorized and who must answer when the authorization is exceeded or the information is wrong. Requiring an identifiable human author for every sentence would answer a different question, and sometimes prevent a useful division of labor without making the surviving labor any more responsible.

The work already assigned

An objection ought to be allowed its full force here. Mata did not reveal a gap in the law that awaited a theory of computational witness. Rule 11 already required reasonable inquiry before a submission and imposed continuing obligations when its positions were later advocated. The opinion opened by acknowledging legitimate assistance from people, books, databases, and reliable artificial-intelligence tools. It located the failure in duties the lawyers already had.⁶

That explanation is sufficient for the sanctions. To rename it a protocol would not deepen it. The usefulness of following the passage from generation to adoption lies elsewhere: it tells us where the new facility entered an old division of work, what was assumed to have been done, and who eventually performed the omitted inquiry. Schwartz produced the research, LoDuca presented it, the adversary challenged it, and the court obtained an account. These were distinct operations even where the rules attached responsibility to more than one participant. Calling all of them human oversight would conceal almost everything worth knowing about the oversight.

Consider the limited success of the signature. It made an adopter visible and helped expose him to consequences. It did not establish that he had read the sources. The duty existed before the inquiry took place, and in this instance the duty was breached. A system which recorded the signature perfectly could still have recorded a perfectly attributable failure. That is a reason to examine what the signature certifies and how the required work is done, rather than to abandon signatures or mistake their presence for the work itself.

The subsequent response mattered too. A duty to correct cannot be completed at the instant of issue, since some reasons for correction arrive later. Rule 11's treatment of later advocacy recognizes this ordinary difficulty. A person can assume responsibility for a statement and then encounter evidence which changes what responsible conduct requires. The record of the first act helps identify him; it cannot determine whether he has answered the new evidence. In Mata, continued reliance after warning was central to the court's judgment.³

This persistence of responsibility is more demanding than a system that merely remembers who clicked a button. It is also more familiar. Editors publish corrections, laboratories investigate discrepancies, lawyers reconsider arguments. The relevant design problem begins with the actual duties and available means of performing them. If a proposed computational arrangement can make the sources easier to examine or a correction easier to bring to everyone who relied on the first result, it has improved something real. It acquires no such merit from recording more fields whose contents nobody is required or able to use.

There is a temptation to make the lawyer's failure stand for a general human incapacity to resist fluent machines. The record does not warrant that conclusion. It shows warnings missed, inquiry avoided, and inaccurate accounts offered under pressure. A theory that excuses these acts as the inevitable effect of a persuasive interface would give the interface too much credit and the professionals too little responsibility. Equally, telling the next office to exercise judgment will be of limited use if its organization leaves no time to inspect a source or no authority to suspend an action. Those conditions must be investigated in the office concerned; they cannot be supplied by the moral of a celebrated case.

A source within reach

A different encounter with ChatGPT began with an operation that worked. In May 2023, the journalist Frederick Riehl pasted passages from a legal complaint into the system and received accurate summaries. He then supplied a link to the complaint. The system said it could not access the internet; after further prompting it produced a different account, falsely introducing allegations of embezzlement against Mark Walters. Riehl had access to the actual complaint. He investigated and established that the account was false. In the defamation action that followed, a Georgia trial court granted summary judgment to OpenAI, addressing the particular interaction, the recipient's knowledge, fault, and damages.⁷

The accurate summaries make the failure more instructive. Success with supplied passages had not established the reliability of a materially different operation. Riehl could compare the new answer with information he already had and examine the document for himself. The source was within reach. His work prevented the generated account from becoming his account of what the complaint said. This was an available outcome, although it is a less convenient emblem than either an infallible assistant or an irresistible deceiver.

The two proceedings concerned different duties and cannot be reduced to a common liability rule. They do show why the path taken by an output matters. A false passage may be investigated as a possible error, adopted as an assertion, submitted as evidence of misconduct, or quoted in a judgment explaining why it was false. Its words can remain the same while the act performed with them changes. An archive containing all those appearances would still need to distinguish their uses.

There is substantial achievement in making language available for work it was previously too costly to undertake. A person can ask an unfamiliar question without first learning the vocabulary of an index. A draft can expose a possibility that its recipient would not have found unaided. A competent summary can spare the labor of an initial reading and leave more attention for the passages that require judgment. These are reasons to improve the machinery and make its capacities more widely available. An account of institutional responsibility that begins by despising the assistance will misunderstand why institutions want it.

The gain can also alter the distribution of work. Where it becomes easy to produce a finished-looking answer, an office may receive more material than its previous reviewing practice was designed to handle. It may change who drafts, what a supervisor examines, or how quickly a decision is expected. None of this follows inevitably from the use of a language model. It describes choices through which a gain in production can become a different arrangement of authority. We need evidence about those choices before declaring that a deployment has saved labor overall, or merely transferred the difficult portion to someone less able to refuse it.

Better technical arrangements can change the inquiry. Retrieval may put the actual decision beside the summary; a citation check may expose a missing case before filing; a permissions system may prevent a draft from being issued as a final notice. Such arrangements can be built from existing components. Their value depends on what they check and which actions their checks govern. Finding the opinion does not show that it supports the argument. A permission to issue does not establish the accuracy of what is issued. These distinctions permit useful improvements to be recognized without asking any one of them to stand for the rest.

The phrase Empire of Strings names the attraction of this facility: language can be produced, recast, and carried into almost any office whose business is conducted in words. It does not name a class of machines condemned to remain outside institutions. The same output can enter a governed process or an improvised one. The question is how the institution receives it, and whether the person who eventually bears its consequence can discover what that reception amounted to.

In Mata, the court required the lawyers to send its opinion and the relevant record to their client and to the judges whose names had been attached to fictitious decisions. It declined to compel an apology, observing that compulsion would not make one sincere.⁸ The remedy was modest in its understanding of what an institution could require. It could make people give an account, attach a cost to their conduct, and place the record before those directly affected. The inward act remained theirs.

A generator that returned only genuine and relevant authorities would have avoided the particular inventions. It would not have relieved the institution of deciding what a submission was entitled to do. Even an accurate description can acquire an unwarranted consequence if the office receives it under the wrong category, or lets a classification decide more than its evidence supports. We have followed the words to the point where someone adopts them. The next difficulty concerns the terms in which the institution is prepared to understand what it has adopted.

---

Footnotes

1. Mata v. Avianca, Inc., no. 22-cv-1461, Opinion and Order on Sanctions (S.D.N.Y., June 22, 2023), findings of fact 2 and 5–6, pp. 3–5. The opening follows the court's findings and its account of LoDuca's testimony; it supplies no imagined scene or private motive. ↩ ↩²

2. Mata, findings 3–11, 13–23, 38–40, pp. 3–10, 16–17. The first filing cited generated summaries and excerpts; the later affidavit attached purported extracts, not authenticated full decisions. The client's injury is stated as an allegation. The court's sanctions opinion, rather than a reconstruction of the merits action, governs this chapter. ↩

3. Mata, opening pp. 1–3; conclusions of law 10–15 and 21–25, pp. 24–31. In this court-initiated proceeding, the opinion applied subjective bad faith, not a rule that every mistaken filing warrants sanctions. The sanctions rested on the respondents' separately assessed conduct, including conscious avoidance and false or misleading representations. ↩ ↩²

4. Mata, findings 24–29, pp. 10–13. The defects discussed are the court's verified examples. A copied identifier was not evidence that the purported case existed. ↩

5. J. L. Austin, How to Do Things with Words (Oxford: Clarendon Press, 1962), lectures II–IV, pp. 14–17, 26, 39–43; primary-text transcription. Austin distinguishes misfires, abuses, and other criticisms of an act, including bad advice. The application to a generated submission is this chapter's argument; Austin does not establish a novel computational doctrine. ↩

6. Mata, p. 1; Federal Rule of Civil Procedure 11, especially (b) and the 1993 advisory committee note concerning later advocacy. Reasonable inquiry depends on the circumstances; it is not a guarantee that every assertion will prove correct. ↩

7. Walters v. OpenAI, L.L.C., no. 23-A-04860-2, Order Granting Summary Judgment (Superior Court of Gwinnett County, Georgia, May 19, 2025), pp. 2–7, 9–22. The accurate summaries concern pasted passages, not access to the URL. Pages 4 and 7 record the recipient's verification; p. 20 expressly states that he did not republish the output. Page 5 introduces the court's three independent grounds; fault is treated on pp. 9–18 and damages on pp. 18–22. The passage makes no claim that every recipient would respond similarly or that the decision grants immunity for inaccurate output. ↩

8. Mata, conclusions 31–34 and decretal provisions, pp. 33–34. The penalty was $5,000, imposed jointly and severally on the respondents. Notification, deterrence, compensation, and apology were distinguished by the court. ↩