Dispute
Legal AI needs a process memory
Legal AI needs a process memory

LexBox

Legal AI Needs a Procedural Memory
We talk about AI hallucinations but miss another risk in litigation. The debate about AI in law has largely focused on the risk of incorrect answers and fabricated legal cases. That is an important discussion. But for those working in dispute resolution, there is another challenge that has received considerably less attention: AI must not only be able to provide a good answer here and now, but it must also be able to hold together a case that changes over time.
Litigation and AI
A dispute is not a single question to be answered; it is a process that can go on for months or years. During that time, new material is added, claims are adjusted, and grounds are developed. New circumstances are introduced, evidence is invoked or dismissed. The opposing party concedes certain information and disputes other points. A piece of information in a new document can alter the meaning of something submitted several months earlier.
To a lawyer, it is self-evident that all of this is interconnected, but for AI, it is not as obvious.
A correct answer can still be wrong for the case
Much of today's AI usage is still based on the lawyer asking a question and receiving an answer: summarize this document, analyze this legal issue, or write a draft based on this material.
That can work very well.
The problem arises when a locally correct answer does not align with the case as a whole.
Suppose that at the beginning of a case, a lawyer uses AI to analyze the plaintiff's claim regarding when a certain agreement was entered into. A few months later, new evidence arrives that changes the timeline. Subsequently, the argumentation is adjusted, and the opposing party also concedes a circumstance that was previously disputed.
If AI is later used to generate a new draft, it is not enough that it can formulate a persuasive legal text. It needs to work from the current procedural status and understand which information is still valid, which has changed, and how it impacts the rest of the case.
Otherwise, the result may be linguistically persuasive and legally plausible – but based on a picture of the case that is no longer current.
Litigation is a web of dependencies
This becomes particularly clear in dispute resolution because information must not merely be stored, but must also be placed in relation to other information.
Facts are linked to operative facts. Operative facts are linked to grounds. Grounds must support claims. Evidence must substantiate relevant circumstances. The opposing party's position affects what is actually in dispute.
When one part changes, several other parts may therefore be affected.
This is also where the difference between using AI for text production and using AI as support throughout a legal process becomes clear.
Summarizing 30 documents is one thing. But understanding that a new email contradicts an earlier entry in the timeline, that the contradiction affects an invoked operative fact, and that this in turn affects how a certain ground should be assessed, is something else entirely.
For that type of work, it is not enough for AI to be good at language. It needs to work with structure and context.
Legal AI needs a procedural memory
When we talk about the AI's memory, it is therefore not just about how many documents or tokens a model can read simultaneously. A long context window is not the same thing as a reliable procedural memory.
What is crucial is whether the system can maintain an up-to-date and traceable view of the case: which circumstances are invoked, what is in dispute, which claims and grounds apply, which evidence belongs to what, and what has changed over the course of the proceedings.
Furthermore, it must be possible to understand where the information comes from. For the lawyer, it must be possible to trace back to the source material, verify the AI's conclusion, and make the final legal assessment themselves. That is a considerably higher bar than receiving a well-formulated answer from a language model.
Hallucinations are only part of the quality issue
Great emphasis has rightly been placed on hallucinations, about which we have also shared an article. An AI system that invents a legal case or a source is obviously problematic in legal work.
But a system does not need to invent anything to lead the lawyer astray.
It can use an outdated piece of information that was subsequently corrected. It can fail to notice that the opposing party has changed their position. It can treat two contradictory statements as if both were still valid. Or it can analyze a document correctly without understanding how that document alters the case as a whole.
That type of error is harder to detect because the result can still look reasonable.
And the more AI is integrated into lawyers' day-to-day work, the more important the question of continuity becomes.
The next step is not just better answers
AI will likely become increasingly better at legal information retrieval, summaries, analysis, and text production.
But for procedural law, we should set higher standards than that.
We should ask how the AI system handles changes over time. How it distinguishes between current and outdated information. How facts, claims, grounds, and evidence are interconnected. How a lawyer can verify the source behind a conclusion. And what happens when new material alters the previous picture of the case.
The important question, then, is not just whether AI can read a dispute, but whether AI can follow it.
Because in a legal process, it is not enough to remember what was written in the documents. One must understand what still applies.
Try LexBox if you want support in your work.
On the page about LexBox Workspace, we show how the lawyer can go from a new statement of defense to source verification and a draft.
START YOUR NEXT MOVE


