AI in the Courtroom: What Works, and What Gets Lawyers Sanctioned

The failures get the headlines: AI-generated fake evidence, or invented precedents that surface mid-case. Those stories travel because they are easy to catch and fun to write up. The quiet successes outnumber them many times over, but nobody counts those. A clean first draft of a motion, or a contradiction in the testimony caught at the right moment, never makes the news.

The question stopped being “should we use AI” a long time ago. It is now “how do we use it without getting burned.” The firms that have wired AI into their daily litigation work show what that looks like.

Sizing up the odds is the most important stage of any dispute. It drives everything that follows: whether to go to court at all, what terms to settle on, and where to draw your red lines.

Here AI’s real edge is that it has no stake in the outcome. It is not attached to your theory, it will tell the client something they don’t want to hear, and it never gets tired. Ask it for a second opinion, then a tenth, then the view from the other side of the table, a skeptical judge, or an appeals court, and it gives you each one.

Asking a model to “rate our chances” gets you nowhere. Put it in the opponent’s seat and make it attack the weak points in your own case. A prompt like this turns it into a relentless devil’s advocate:

From a second opinion to a decision tree

Next, put numbers on it. A decision tree breaks the dispute into forks. At each decision node you pick an action, litigate or settle; then chance takes over, win or lose, with each branch carrying its own probability and dollar outcome. Multiply the probabilities by the outcomes and you get the expected value of each path, the average dollar result weighted by the odds. That lets you weigh a guaranteed settlement against a risky trial on the same scale.

Trees like this once meant drawing by hand, often in pricey specialist software. Now AI roughs out the layout, suggests probabilities from comparable cases, and recomputes the expected value the moment any input changes.

You don’t have to build the tree by hand. Copy the prompt, drop in your own case, and the model will ask for what’s missing, build the tree, and find the settlement figure at which a full trial stops being worth it:

Once a dispute is unavoidable, AI shifts from adviser to workhorse. It carries the heavy, repetitive load: first drafts of filings, the chronology, summaries of a sprawling record, the hunt for contradictions in the other side’s arguments. The lawyer spends the time that frees up on strategy and on checking the work.

Below are 10 core litigation tasks, each with notes and a rating for how much AI actually helps. The pattern is simple. The closer a task stays to your own materials, the bigger the payoff. The more it reaches outside for sources, case law most of all, the bigger the risk.

Any assistant is only as good as the context you give it. An AI’s answer is only as good as what it can see. In a dispute, that context is the whole case, not a single document, and the more complete and clean it is, the better the model performs.

What goes into the context of a dispute

How to assemble the context and share it

You can pull the context together in one place with the “Projects” feature in the major AI services, or with Claude Cowork. Upload documents straight into a project, or let Cowork connect to your cloud storage (Google Drive, Box, and the rest) and draw the materials in without manual copying. The whole case then lives in one window.

It can also be shared, which matters just as much. On team and organization accounts, a project opens to colleagues, so everyone works from one vetted base instead of a dozen scattered copies.

Context in hand, you still can’t trust the output on faith. AI has one dangerous habit: it delivers a wrong, often invented answer with total confidence and makes it sound right. A recent economics paper puts the real bottleneck not in how smart the machine is but in how fast humans can check its work. The more an AI produces, the more it costs, and the more it matters, to verify. In a dispute that bites hard, because anything headed for court has to be checked.

Three things deserve special attention:

These are where AI slips most: it invents cases that never existed and quietly rewrites the wording of the ones that did.

A formal check with AI

You can hand the first pass to the AI, as long as you keep it on a tight leash. Skills and workflows help: you spell out once what to check and how, and the model runs every document through that checklist. A skill like this takes each citation, hunts for the original source, and marks it confirmed, not found, or distorted. That doesn’t replace a human review, but it clears the first layer of errors and saves real hours.

A check by a live lawyer

The last stage is a human review. A machine answers to no court and no client, and it cannot be disbarred. So the final word, and the signature, belong to the lawyer, who weighs not just whether a source exists but the strategy, the tone, and whether the argument fits. The sequence is simple: the AI draft clears a formal check first (a skill or workflow can automate that), then the lawyer’s substantive review, and only then does it get signed and filed.

First things first: check everything by hand before it leaves your desk. In U.S., U.K., and other common-law courts the professional duties, and the personal liability, run higher than many lawyers expect, and the court holds you, not the tool, responsible for every word you file. Catching an AI error yourself, before anyone else does, is always the goal.

Even so, the most careful review can miss one. If an error slips through, don’t wait for your opponent to point it out to the judge. Disclose it first, and offer the correction yourself. A court is far kinder to the side that owns a mistake than to the side that stays quiet about it.

Getting ahead of it turns a potential disaster into a case study in crisis management: you stay in control, you show good faith, and you keep the one thing that matters most, the court’s trust.

Brevity has always been a virtue. In the age of AI it is a survival skill. A model can pad a filing to a hundred pages in minutes, and parties lean on that constantly. The problem is that no judge can actually get through it.

A lean filing is far more likely to be read in full and to land. Resist the urge to “strengthen” your position with one more minor point; it just blurs your real arguments and drains the patience of a judge who is already buried. The job is to fit in every argument that carries weight while holding the judge’s attention. More pages mean more arguments, and worse odds the judge reaches the end. The sweet spot is where those two lines cross.

Confidentiality is the bedrock of the lawyer-client relationship. A client trusts that whatever they hand their lawyer stays with the lawyer.

Lawyers and clients have passed information through cloud services for decades, though: email like Outlook, messaging apps, file storage. In each of those, the provider can technically reach the data. That rarely worried anyone, because two things covered it: the end-to-end encryption the popular messengers promise, or a clear term in the user agreement that the service won’t use what passes through it.

AI puts that old problem in a new light. Some providers may feed conversations straight into training the next version of their model, which raises a fair worry: could a client’s private data resurface as a future model learns?

So the confidentiality question today comes down to one thing: does the model train on the data in the chat you’re working in? The market has settled into two answers.

Professional norms are already forming around this. A client has every right to ask a few pointed questions, and the lawyer should have the answers ready:

1. Is AI used at all in providing the legal services, and does the client consent to it?
2. Is the data anonymized before it is uploaded?
3. Which models, and which providers, are used?
4. Does the firm have a direct contract (a subscription) with the model provider?
5. Does the lawyer guarantee the model does not train on the uploaded data?
6. Is a Zero Data Retention (ZDR) agreement in place?

More and more, these come up when the engagement is signed.

Confidentiality is the lawyer’s duty to keep a client’s information from outsiders. Privilege goes further. It is a protection the law grants the lawyer-client relationship and backs with hard guarantees in litigation. But it rests on that same confidentiality, and the moment the material reaches an outsider, the protection can fall away. Uploading a client’s data to a public AI is, for these purposes, handing it to an outsider.

Privilege and the work-product doctrine

Common-law systems offer two protections. Attorney-client privilege shields confidential communications with a lawyer made to get legal advice; it belongs to the client and is lost on voluntary disclosure to any third party. The U.S. Supreme Court called it “the oldest of the privileges for confidential communications known to the common law” (Upjohn Co. v. United States, 449 U.S. 383 (1981)). There is also the work-product doctrine (Hickman v. Taylor, 329 U.S. 495 (1947), and Rule 26(b)(3) of the Federal Rules of Civil Procedure), which covers materials prepared in anticipation of litigation. That protection is sturdier: it is generally lost only on disclosure to a litigation adversary. Under ABA Formal Opinion 512 (2024), a lawyer needs the client’s informed consent before entering information about the client into a self-learning AI, and a boilerplate clause in the engagement letter does not count as that consent.

Compare the prompts a lawyer writes to build a case: where they carry counsel’s own thinking, courts have shielded them as work product (Tremblay v. OpenAI, N.D. Cal. 2024). The lesson is simple. The lawyer should do the AI work on a matter, through enterprise tools with real confidentiality safeguards. And it is worth telling clients, for their own sake, not to “consult” public chatbots.

Courts and regulators are catching up to a world where most lawyers use AI every week. Rules on AI in disputes are landing at several levels at once: courts decide how existing rules apply, bar regulators police professional ethics, and legislators take up the delivery of legal services through AI. Expect those rules to keep getting more detailed. Here is a snapshot across the main forums: international arbitration, the U.S. courts and bar regulators, the EU, the U.K., and a few other jurisdictions.

The rules change almost monthly, there is no single standard, and there probably never will be one. Get in the habit of checking five things before every filing: the specific judge’s standing order or the tribunal’s rules, the AI service’s policy on your data, any data-protection limits on what you can upload, any duty to disclose that you used AI, and, of course, hallucinations in whatever the model produced.

Disclaimer

This material is for information only and is not legal advice. The figures, case details, and ratings are illustrative and should be verified independently. The decision tree and the payoff ratings are notional. For advice on using AI in disputes, on assessing the merits of a case, or on internal policies, contact Buzko Krasnov.