In this five-part series, data analyst David Stasior demonstrates how AI can turn unstructured legal records into measurable data, testing the rule of law against empirical reality. Part 1 made the case, and Part 2 put it to work on 24 years of Massachusetts attorney disciplinary decisions. In Part 3, he turned to federal court statistics, where AI quickly surfaced rule-of-law insights buried in awkwardly formatted tables. Part 4 aimed AI at the text of federal court orders to reveal the reasoning judges use when deciding whether to end supervised release early. In Part 5, the final installment, Stasior turns from hindsight to real time.

Throughout this series, AI has tested our beliefs about the rule of law against empirical data. In Massachusetts, a disciplinary authority the public sees as a stern guardian of legal ethics functioned more like an inconsistent consumer-complaints bureau. The “rocket docket” title of fastest federal trial court rested on just 16 trials in 2025, and a fuller analysis of the courts’ own statistics handed the crown to a fellow district. Publicly available orders showed federal judges applying contradictory standards when deciding whether to end supervised release early—even the same judge applied different standards in different years.

Each finding points toward improvement. Disciplinary authorities, court administrators, and judges can use the same tools to examine their own records and identify where inconsistency undermines fairness. But all these findings share one feature: they look backward.

Historically, looking backward is often the first step in turning an unmeasured domain into a measured one. Consider medicine. In 1896, an Italian physician introduced an inflatable cuff that turned an invisible internal force, blood pressure, into a number. Decades of population studies then showed that specific readings reliably predicted strokes and heart attacks. Only after that retrospective work did the cuff become what it is today: a routine check at every doctor’s visit, where risk is caught and treated on the spot rather than discovered years later.

Now with the ability of AI to read at scale, nothing prevents law from following the same path. Putting that reading to work while cases are underway requires another step: building it into how courts check findings and manage information. Two cases, one from a civil trial and one from a small claims courtroom, illustrate what that could accomplish while a matter is still being decided.

At the end of Part 4, we described a fiercely contested civil case in state court: nearly a dozen days of testimony, millions of dollars at stake. The judge asked both parties to submit proposed findings of fact and, months later, issued his own. A layperson expects judges to write their own findings. Yet we asked you to assume this judge copied at least one sentence and to write down two numbers: out of every 100 sentences in his findings, how many were lifted, word for word or nearly so, from each side? If you wrote down your numbers, now is the time to check them.

In this case, about 55 of every 100 sentences in the judge’s findings were copied, word for word or nearly so, from the plaintiff’s proposals, and about 11 from the defendant’s. Roughly two-thirds of the findings came directly from the parties’ own drafts, a startling proportion. AI also showed that the copied material differed in kind. The sentences copied from the defendant were mostly background and undisputed facts, while those copied from the plaintiff went to the heart of the dispute. This judge largely adopted one side’s account as his own.

AI uncovered a deeper problem. Some of the copied findings contradicted the trial exhibits. Checking findings against evidence is exactly the kind of reading AI does well: it can read an exhibit, hold its contents, and then read each finding in turn, flagging apparent contradictions for review. Finding the copied sentences was the easy part, closer to the simple text matching described in Part 1. Spotting a contradiction is harder. It requires the skill Part 1 described last: understanding not just which words appear, but who is chasing whom.

In this case, AI found errors, and every one favored the proposing party. The errors consistently strengthened that party’s account. And some passed, word for word, into the judge’s findings, where they wrongly carry the weight of judicial fact.

That raises two concerns for the rule of law. First, an officer of the court submitted proposed findings that misrepresented the evidence. Part 2 quoted Massachusetts’s highest court calling truthfulness and candor “the cornerstones” of the legal profession, yet the analysis of disciplinary decisions found that sanctions for dishonesty ranged from nothing to years. Second, the judge adopted those incorrect findings, turning one party’s errors into his own. There is little reason to think this case is unique. Similar errors may sit unnoticed in countless decisions because no one has compared findings against the record at scale.

This analysis, however disappointing, still looks backward: AI read the record after the judge had ruled. It does not have to. Before signing a decision, a judge or clerk could ask AI to check every proposed finding against the exhibits. In Part 4, an AI agent running on an ordinary Mac mini extracted the reasoning from each judicial order in about five seconds. Each party could run the same check on its own submissions before filing, and on the other side’s proposals before the judge rules. The errors in this case could have been flagged before they became the court’s findings.

We hold college freshmen to a strict standard on plagiarism and give them software that checks their essays against millions of sources before they submit. Surely the judiciary can meet that seemingly low bar. AI offers an additional check: it can compare proposed findings against the record and flag contradictions before a judge adopts them. Judges can use that assistance to write findings that are their own and grounded in the evidence.

AI’s ability to read at scale could also become the foundation for changes in how courts are managed. A recent small claims hearing showed me where that could begin.

A relative, knowing my interest in legal analytics, invited me to join her at a small claims court hearing. Over a year before, someone had tried 20 times in a few hours to charge her credit card fraudulently. The bank holding the card blocked 16 of the attempts but allowed four through, with a total cost of thousands of dollars. Despite a rapid report of the fraud, the bank spent weeks investigating, only to issue a boilerplate denial letter padded with pages of merchant transaction reports that did not establish she had authorized the purchases. Three more appeals resulted in the exact same form letter. Given this, she sued.

The first available pretrial hearing was months away, because the court held small claims sessions only about once a month. When we arrived that morning, more than 40 people were waiting in the gallery. I was stunned that so many people had brought new claims. But as I waited for the judge to call her case, it appeared ours was the only new case with both parties present.

The docket was crowded with unfinished business. Case after case was a follow-up hearing because one or both parties had failed to complete a task ordered at an earlier hearing. The inefficiency was staggering.

In one case, a criminal court had ordered a defendant on probation to pay restitution of $50 a month. The apparent victim had not relied on restitution alone and had also obtained a small claims judgment. At the previous hearing, the court had ordered the plaintiff to account for what had been paid and the defendant to file a financial affidavit. Neither did. The judge spent 15 minutes unsuccessfully trying to piece together basic numbers in real time before scheduling yet another hearing.

In another case, a frustrated plaintiff requested a hearing to prove that a contractor could afford payments of more than $200 a month. Prior to the hearing, the contractor had not completed his required financial affidavit, instead filling one out in the gallery that morning. The judge noted that it showed income of $2,000 a month against expenses of more than $4,000. The contractor said his wife worked and made up the difference. He then asked to lower his payment to $100 per month.

The plaintiff then held up a stack of photos from the contractor’s Instagram account, saying they showed recently completed projects. The contractor responded by saying those projects were older. The judge held up a CD the contractor had submitted a year earlier, containing his business and personal finances from that time, and instructed him to provide a similar accounting. The judge left the payment amount unchanged and ordered yet another hearing. The plaintiff stormed out of the room in frustration.

Case after case followed this pattern. Each earlier hearing had ended with an order stating exactly what each party had to do. The court waited until everyone returned a month later to learn whether the order had been followed. Instead of this slow and unreliable process, AI could read the order and identify each party’s obligations and deadlines. Connected to the court’s case-management system, that information could trigger reminders by text or email.

When the contractor filed his affidavit, AI could flag missing information or figures needing explanation, such as expenses double his reported income, while there was still time to respond. The court could schedule follow-up hearings in just days and check whether the file was ready before bringing everyone back. As it is, the court resets the clock a month at a time. When the clerk told the judge that the earliest trial date for my relative’s case would be sometime in 2027, the cost of waiting was plain.

In Part 3, the chief judge of the “rocket docket” invoked the adage “justice delayed is justice denied.” In small claims, the delay is built into the calendar. A monthly hearing schedule echoes the era when judges rode circuit, traveling from town to town on a fixed schedule. Many filing and document-handling practices likewise reflect the demands of paper. Yet electronic submissions need not sit unread until the next court date. The tools to read them, identify missing information, and prompt a response are now available. We cripple how well our courts can dispense justice by relying on archaic practices.

Over five installments, AI has read roughly 4,000 disciplinary decisions, more than 10,000 federal statistical tables, a set of judicial orders on supervised release, and the findings and exhibits of a contested civil trial. Each time, it tested a belief about the rule of law against what the documents actually say. Reading is no longer the bottleneck.

Access is. A state court system refused to release electronic divorce files. Massachusetts’ disciplinary authority closes about 95 percent of attorney complaints without an investigation, leaving no public record. The federal courts keep their case records behind PACER’s paywall. The tools that could measure and strengthen the justice system are ready, but the documents they need remain locked away.

Every court and disciplinary authority can act now. Release records in bulk, so the public can measure how the law is administered rather than rely on assumptions. Read filings against the record before a decision issues, to catch copied errors before they become judicial findings. Read orders and submissions as they arrive, prompt parties to respond on time, and help organize hearings when files are ready, rather than a month later.

Part 2 ended with a choice: ignore the evidence, lower our expectations, or change. This series has argued for change, and AI has made change affordable.

Part 1 opened with a profession anxious about AI that writes. That anxiety is understandable. Meanwhile, the AI that reads has been quietly ready, able to make the rule of law visible, testable, and stronger. The technology is here. Why wait?