AI Can Read Your Case. Can It Understand What Matters?
Every week seems to bring another major announcement about artificial intelligence for lawyers.
Every week seems to bring another major announcement about artificial intelligence for lawyers. CoCounsel. Harvey. Legora. Protégé. Even the major AI model providers like Anthropic (Claude) and Google (Gemini) are pushing deeper into legal workflows.
At the same time, some lawyers are building their own private technology stacks around ChatGPT, Claude and other general-purpose tools, while hundreds of legal AI companies compete for attention.
For lawyers who are curious about AI, there has never been more to explore. For lawyers who worry that AI may eventually displace some legal work, there has never been more fuel for that concern. And for lawyers who are simply tired of hearing about AI, the noise has become difficult to avoid.
Beneath all the product announcements, however, something more important is happening. Access to powerful AI is becoming less of a competitive advantage. The major models are now very capable (and are getting better) and lawyers can increasingly access similar underlying intelligence through multiple products or directly if they want to build their own system.
So, if everyone has access to powerful AI, what actually makes one litigation system better than another?
Reading a Case Is Not Understanding a Case
Large language models are very good at reading documents. Give one a complaint and it can summarize the allegations. Upload a deposition transcript and it can identify testimony on a particular topic. Provide medical records and it can extract diagnoses, dates and treatment.
Those capabilities are useful, but a lawsuit is not simply a collection of documents waiting to be summarized.
Litigation is a developing body of evidence. An allegation in a complaint is not an established fact. A discovery response may conflict with deposition testimony six months later. A medical record may reveal a prior condition that changes the causation analysis. A photograph may alter the significance of a witness’s testimony. An expert report may rely on an assumption contradicted elsewhere in the record. Information that appeared unimportant early in the case may become critical after a new witness is deposed or a new document is produced.
Experienced lawyers understand this instinctively. Two lawyers can read the same file and come away with very different views of the case. The difference is not simply their ability to read. It is their ability to recognize relationships, inconsistencies, gaps and significance.
That is the harder problem for litigation AI. AI can read your case. The more important question is whether it can understand what matters.
From Case Data to Case Intelligence
For AI to become genuinely useful in litigation, the case record has to become more than a folder of files available for retrieval. The information must be structured and connected.
The complaint contains allegations and legal theories. Discovery identifies contentions, witnesses and claimed damages. Medical records contain diagnoses, histories and treatment. Depositions add sworn testimony. Photographs and videos provide visual evidence. Experts offer opinions based on a selection of assumptions and facts. Each source adds information to the same case, and the value comes from understanding how that information fits together (or not).
Consider a plaintiff who testifies that an accident caused a particular physical limitation. That testimony may appear completely reasonable when viewed by itself. But the medical record may contain a similar complaint several years earlier. An interrogatory answer may deny prior treatment. A later physician may have recorded a different mechanism of injury. The useful insight is not simply a summary of each document. It is the relationship among them and how they fit together into the case as a whole.
Some years ago, I had the good fortune to participate in an excellent week-long trial advocacy program led by former federal district judge Herbert Stern. I remember this question from his opening lecture: “How do you determine what the jury will believe?” he asked. “It has something to do with the inner computer we all have that constantly tallies the thing we are asked to believe based on everything else we think we know.” That is close to the problem litigation AI has to solve: not deciding what to believe, but continuously testing each new fact against everything else in the record.
A complaint allegation may connect to medical history, a discovery response and deposition testimony, creating a potential causation or credibility issue. A witness statement may conflict with a photograph or another witness’s testimony, creating an impeachment opportunity. A damages claim may depend on an expert assumption that is undermined by prior records. These relationships are where much of litigation strategy begins.
That is what I mean by case intelligence: turning the litigation record into structured, connected and verifiable knowledge that helps lawyers identify what deserves attention. Structure tells a system what something is. Case intelligence helps surface why it may matter to the case.
The lawyer must also be able to inspect the result. If an AI system says testimony conflicts with a medical record, the useful response is not simply, “Trust us.” The lawyer should be able to read the testimony, evaluate the medical record, understand why the inconsistency was surfaced and return immediately to the underlying evidence. The closer AI gets to helping lawyers analyze a case, the more important that connection to the record becomes.
The Record Itself Is Changing
There is another reason litigation systems need to do more than simply read what they are given. The material entering the litigation record is changing too.
Generative AI has dramatically reduced the effort required to produce legal-looking text. That is affecting not only lawyers, but self-represented litigants. Recent academic research suggests that federal courts are already seeing the effects.
One 2026 study analyzing approximately 2.8 million federal civil filings found that the rate of pro se plaintiffs increased from 11.33% before widespread generative AI adoption to 16.94% afterward. A separate 2026 study using more than 4.5 million non-prisoner federal civil cases and 46 million docket entries found a similar increase, with the pro se share reaching 16.8% in fiscal year 2025.
Those studies do not establish that AI caused the increase in self-represented litigation, but they do provide evidence that AI is lowering the practical barrier to producing filings that look and sound like legal documents. That creates a new challenge for lawyers and courts. More text does not necessarily mean better legal analysis, more citations do not necessarily mean real authorities, and more polished pleadings do not necessarily mean more meritorious claims.
Sometimes the risks are stranger still. Recently, a Connecticut judge discovered that a self-represented litigant had embedded hidden white-text instructions in court filings designed to influence an AI system that might read them. The instructions reportedly told AI tools to agree with the litigant’s position. The judge treated the technique as a prompt-injection attempt and revoked the litigant’s electronic filing privileges.
This development illustrates a broader point. For litigation AI, being grounded in the record is no longer enough. The system must also be skeptical of the record. Litigation has always involved unreliable allegations, inconsistent testimony and disputed evidence. AI introduces additional failure modes: hallucinated authorities, machine-generated arguments, manipulated content and even instructions aimed at the software rather than the human reader.
A litigation platform cannot assume that everything entering the case should simply be ingested and trusted. It must help evaluate it.
Verification Has to Become Part of the Workflow
Hallucinated legal citations are the most obvious example. Despite years of warnings, courts continue to encounter filings containing authorities invented by generative AI.
A California appellate court recently sanctioned an attorney for an appellate filing that contained fictitious citations. Though a paralegal had reviewed the citations, the court emphasized that lawyers remain personally responsible for verifying the authorities they submit, even when AI or staff members are involved in the research process.
The issue has become significant enough that one federal magistrate judge has proposed a nationwide federal rule addressing verification of citations in filings involving generative AI.
The professional obligation is clear. And technology can make satisfying that verification burden easier. If an incoming pleading cites twenty cases, why should the lawyer have to begin by wondering whether all twenty exist? Verification can become part of the litigation workflow itself.
Just as an incoming pleading can be checked for questionable citations, an outgoing draft can be checked before it is filed. A factual assertion generated from the case can remain connected to the evidence supporting it. An inconsistency can link directly to both sources. An insight can be inspected rather than simply accepted.
This is not about allowing AI to replace lawyer judgment. It is the opposite. The purpose of case intelligence is to give the lawyer a stronger foundation on which to exercise judgment.
What Will Differentiate Litigation AI?
Legal AI products will continue to improve. New models will appear. Foundation models will become faster, cheaper and more capable. That progress is inevitable, but for litigation the underlying model will increasingly be only one component of the system.
The harder questions are different. What does the system know about the case? Can it distinguish an allegation from evidence? Can it connect testimony to medical history, discovery and other testimony? Can it recognize when new information conflicts with what came before it? Can it verify the information entering and leaving the case? Can it show the lawyer exactly where the insight came from? And can it do all of this as the record grows and changes?
Those questions are particularly important in civil litigation defense, where a single matter can involve thousands of pages of medical records, extensive written discovery, multiple depositions, photographs, expert materials, and years of factual and procedural history.
At esumry, that is the problem we are focused on. We are building around a structured case record so that pleadings, discovery, medical evidence, testimony and other case materials do not live as isolated documents. As the case develops, the information can be connected, analyzed and traced back to its source. We are also adding citation verification to help identify questionable authorities in both incoming and outgoing pleadings.
These are not attempts to automate away the lawyer. They are meant to streamline the searching, checking, reconciling and reconstructing that consumes lawyer time before good judgment can even begin.
AI can already read a case remarkably well. The more important question is whether it can help a lawyer understand what matters, and show its work to be trusted. That will require more than increasingly powerful AI models. It requires structured case knowledge, evidence-linked analysis and verification designed around the way litigation actually develops.
The lawyer still provides the judgment.
The technology should supply better case intelligence to inform it.
Schedule a demo to see how esumry helps defense teams turn case documents into structured, accessible litigation knowledge that supports trusted, faster, consistent pretrial work.
About the Author
James Chapman is a co-founder of esumry and a defense litigator. He writes about the intersection of AI, litigation strategy, and legal operations.
Using esumry, privilege is protected with ZDR (zero data retention), and case analysis is fast, strategic, and secure. Create timelines, tag testimony, assess credibility, and get ahead of how the other side will use the record—before they do.
