Recently, however, a few flavor of this design feature of LLMs (it is not a flaw in them, it is a design feature which I explain below) created another version of this issue and the New Mexico Supreme Court was not having it.

Stephen Aarons is a New Mexico defense lawyer working in Santa Fe. He is not a newbie. He has more than 40 years as a lawyer according to records there. I don’t say this to be nice - I say this because I believe it - I am sure he is a smart, capable lawyer. I think the majority of folks sanctioned for AI related issues like this one are also smart and capable.

Mr. Aarons was working on the appeal of his client’s murder conviction. He loaded a transcript of his client’s murder trial into an AI tool. This is, by the way, a pretty smart thing to do to enable the AI to search for gaps in the evidence, contradictions, and even corroboration. One slight fact I omitted, however, the transcript that Mr. Aarons uploaded to an AI tool was computer-generated. What does “computer generated” mean in this context? Who knows. Many things can be considered computer generated, including this substack article considering I am typing it on a laptop computer using software loaded on that computer. But, in this context, it means, a transcript likely produced by loading an audio recording of the trial.

He loaded that transcript along with other documents from the case file. He expected what he later called a “bulletproof” summary of the proceedings. He relied on the output of that AI tool and filed a brief last year. There was recently a hearing about the case and as the Santa Fe New Mexican reported the hearing which included content from what Mr. Aarons relied upon and filed, referenced several witnesses who do not exist. It attributed fictional testimony to witness who does exist. I don’t know, but that seems like a problem. The New Mexico court labeled that problem a false statement of fact to a tribunal in a homicide appeal.

The justices found him in contempt, removed him from the case, ordered him to pay $5,000 to a client-protection fund in thirty days, and referred him to the Disciplinary Board. Of course, his client did not even prevail on the appeal or have his hearing yet. The brief filed on behalf of the client was struck. The appeal now goes to a public defender and begins anew.

As of this morning (August 23, 2026) the court’s order is not posted. Everything that follows is from the New Mexican’s account of the hearing.

Aarons told the Court he assumed the AI tool would summarize the transcripts properly. He did not assume it would invent people or quotes, or quotes of existing witnesses, etc. In a written response he admitted “undue reliance upon generative artificial intelligence” and said he did not realize the tool would generate, “apparently out of whole cloth, fictitious witnesses, testimony, quotations and authorities in a form that appeared coherent and plausible.”

LLMs Invent Stuff

If you get anything out of this article, it should be this. LLMs are designed, I repeat, designed to hallucinate and make stuff up “out of whole cloth.” This is not a flaw in their design, it is an intentional design decision. They are sycophantic (in case you have not noticed). Whatever you state as your desired strategy and outcome, the AI is designed to cheer you on like a groupie. Given that philsophical bent, it is unsurprising and intentional that when the LLM cannot find information to support your requested prompt seeking validation - it fabricates it.

I have consistently called LLMs, 20% fabricating next word guessing machines. You should also repeat that in your head as you read every output of any AI tool, OpenAI, Anthropic, Google, Meta any open source tool, anything. They won’t acknowledge a 20% hallucination rate. But, what is the basis for whatever quotes hallucination rate they do admit? I can tell you. It’s their use of code analysis and generation tools built by AI to analyze their LLMs. This is not a fox in the henhouse. This is a fox marketing, selling, supplying the materials for, installing and then guarding the henhouse assuring the owner they will not lunch on the hens.

LLMs do not know which witnesses were or were not in the courtroom. They do not know that fabricating answers is wrong. They do know which words often follow which other words. Hand them a long, messy trial record and ask for a clean summary, and they will fill gaps the way they fill every other gap: with the most statistically comfortable sentence. Sometimes that sentence names a person who was not in the trial at all.

Mr. Aarons’ excuse for missing this in his filed pleadings is plausible, but only if he is entirely unaware of what has been happening in the world of hallucinated cases the past 24 months or so. The fact that this case happened and I am writing about it indicates he was. acon was not interested in a teachable-oops framing.

One of the Justices reportedly asked Mr. Aarons, “Do you watch the news?” The problem of lawyers relying on AI hallucinations is, she said, “an above-the-fold story every single day.” Seems like the court was judging whether he buried his head in the sand — “an intentional choice to be uninformed” — or he took a gamble. “Neither of those are consistent with the code of conduct.”

Justice Michael E. Vigil made the point we should already be making to ourselves. Using the tool is fine, he said, the way using a law student to help draft is fine. “But when you put your name on the brief and file it, you are attesting you have checked the brief and it is accurate.” Aarons, Vigil said, “assumed without checking.”

This is a murder conviction. The client is serving a life sentence. The client was apparenlty not informed about Friday’s hearing. Chief Justice Julie Vargas said the justices were concerned the defense lawyer lacked concern for his client.

Restarting an appeal, homicide or otherwise, is not merely a paperwork inconvenience. It is time. If the client being defended was improperly convicted in the end, they are spending months and perhaps longer in prison awaiting a restarted appeal. Time is wasted by the prosecution, the courts and their respective staffs. The time a victim has to await justice, if that is the case, is extended.

Prosecutors can potentially have the same problem with different files. Body-worn camera review is currently manual, but will increasingly become a hybrid model - automated analysis, human review. Audio files of 911 or jail calls the same. A computer-generated bodycam or audio file transcript is a guide, a valuable time saving search tool, but the bodycam or audio itself is the record. An LLM can just as easily invent what an officer said. It can invent a civilian witness who didn’t exist. It can put a real detective on the stand saying something that is not in the footage and not in the transcript.

New Mexico still has no generative-AI filing rule. A committee has been sitting for about a year. I don’t think they need one. They merely need to remind lawyers that no matter what tools, paralegals, associate lawyers, etc. they used to create a pleading, whomever signs it is ultimately saying “I confirmed the accuracy of everything in here.”

The same day as the hearing, Catherine Lynch, a spokesperson for the First Judicial District Attorney, said the office has no written AI policy and “we do not rely on AI to perform these [filing] tasks.” The statewide public defender, through spokesperson Maggie Shepard, is “working on finalizing a formal AI policy.” Again, what policy would this be? In the end, it should be as I stated above.

Here is a working rule that balances the benefits of AI with these risks (which are entirely avoidable). Do put trial transcripts, briefs, expert reports, etc into AI tools (provided you can secure confidentiality of that content). There are so many valuable things you can do with an AI analyzing all that information that is beyond what lawyers have the time and attention to achieve.

The safety value here? Ensure that when you provide this information to the AI and ask it to investigate one of those 10 items (or others you generate) it provides a citation to the place in the record used to reach whatever conclusion.

That output can be gold for your preparation on either side of a case, criminal or civil. Imagine the AI output says in part, “opposing counsel claimed witness A would testify as to these facts __________”. (citation to pages of transcript during opening statement). “The witness actually said the opposite in three places during their testimony.” (Citation to other pages of the trial transcript during that witness’ testimony). With this, you can now review those AI claims and confirm or not their accuracy. If accurate, you just found something that would have taken hours to uncover manually, if at all.

Do not just deposit that material into an AI tool and ask it to summarize and then, rely on quotations or claims or statements of evidence in that summary without checking the citation to the record form which they supposedly arose.

This is not a hard rule to comply with. It’s what we have always done as lawyers, relying on others for case law, arguments, research, etc. But, the best of us have always confirmed. Just keep doing that and your use of AI will be all benefit and no risks realized.

When I was a 3rd year summer law clerk the firm I was working with believed in dictation. Staff typed stuff that lawyers dictated to them. And, when that document draft was typed out the expectation was that we, the lawyers, would dictate any changes to it and return that tape to the staff person who typed it. So, being interested and learning about tech since middle school, you can imagine the answer to this question - did I get a job offer from that firm for post-graduation? Ha, no way.

We are many years from that time, but the situation in the legal field is much the same. As I have written about here before, there are now more than 700 cases and counting of lawyers sanctioned for just typing whatever into some AI chatbot and copy/pasting the case law, quotes, deposition quotes, whatever into their pleadings and filing it. Nothing says technological un-sophistication like that. This is not a critique of intelligence of lawyers. Even those caught in these sanctionable “the AI did it” moments are not by virtue of that error not smart. What they are is not aware.

I have always been someone interested in and constantly learning about technology. Simultaneously, I have always been someone who did not purchase, use, subscribe to the latest gadget just because it was the latest gadget. AI is no different. There are many legalAI products, services proliferating in the market. The ones you have heard of merely have the biggest marketing budgets so far. However, the real question is, now that AI is available to so many entrepreneurs, are there any true “moats” left for any of these companies sufficient to get your monthly subscription over some other competitor?

Let’s start small. Lawyers do these tasks, in whatever order, in their daily practice.

1. Speak on the phone

2. Speak in person (clients, lawyers, experts, judges)

3. Draft pleadings

4. Draft letters

5. Research legal issues

6. Ponder creative solutions to client problems

7. Negotiate.

This is not the sum total of what all lawyers do every day, but it’s a good starting list of the basic tasks common to most lawyers.

What does AI have to say about any of those? Automation is basically the answer. There is no world, currently, where you can set it and forget it and let a Large Language Model (LLM) just draft something, anything for you to file, send or otherwise rely on. LLMs are automation machines. In some cases, they can do it so well they perform tasks the lawyers just haven’t developed a workflow for because it is too time consuming. This is especially true of solo and small firm lawyers who are business owners and employees, HR, sometimes cleaning staff, IT…you get the point.

AI is the term of the day, so I use it. But, LLMs are not anything I. They are not intelligent. The way to think about them is akin to a magician making an elephant disappear. When you are in the audience, it all looks amazing and you have no idea how that happened. But, those facts do not equate to the elephant actually disappeared. Keep that in mind when watching the output of AI. It looks a lot like it is thinking, reasoning, solving puzzles and problems with creativity, etc. But, LLMs are simply next word guessing machines. They do not “know” what the sentence, “The United States has a President but not a King” actually means. They are simply very good at mimicry in response to questions from you and I. The results they provide are very fast, very data intensive predictions of what the most likely next word is to complete a sentence and then the next sentence and so on.

Even considering a robot trained to fold towels and is presented with a fitted bed sheet to fold. It may well figure out how to do it efficiently. But, it was not learning. It was predicting based upon its folding training data what would work. Inside the brain of the robot is simply math, calculations, probabilities and then action. Now, many of you might be saying, “well, isn’t that what we do from being babies to adults?” That is what it seems like we are doing, but in reality, we are learning. No one can teach a baby to walk. It seems other people walking, their way is very different from a baby’s stunted, wobbly beginnings. But, then, something clicks and rapidly that walking becomes smooth, less scary for the parents and so forth.

Let’s refocus from tangent-land. My overall point of this post is do not worry that you do not know about or have not used the latest whatever Legal AI thing. The point is not to just use a thing. The best way to approach Legal AI is to sit for 30 minutes and write down any currently manual tasks you are performing. Include in that any tasks that take more than 4 steps right now. Then research whether a tool, AI or not, exists to reduce that inefficiency. In some cases, there will not be one. In some cases there will be one for $1,000 a month license - yeah, no thanks. In other cases, a cheap tool will exist that saves you an hour or more a month. There’s your arbitrage.

I charge $X per hour. The tool costs 25% of x and saves you an hour, you just made money. So, go for that tool, spend a little time to learn it and experiment. Not working? Pivot fast. We are long past the era of learning some overly complex unintuitive interface and workflow just to be stuck with it because - sunk cost.

LLMs are great tools. I use them all the time. I am amazed at the speed with which they can accomplish tasks. But, I have not yet seen a task they perform that either I or some other human cannot also perform - just much more slowly and perhaps struggling with some accuracy at scale. Automation. That is what LLMs represent for lawyers right now and likely will for years to come.

For now, embrace the opportunity to automate away some tedious stuff from your personal and professional life and, as is often said in our overly tech world - go outside and touch grass.

Measuring hallucination rates, which will never be zero if you know how LLMs were created and how new versions are produced, seems narrowly useful. It also presumes lawyer/users of AI tools are or should be simply signing and filing whatever comes out of an AI tool. That is a bad idea now and will be a bad idea 50 years from now unless the rules governing the conduct of lawyers change to make them not liable for AI outputs in their signed filings.

Is Hallucination Rate The Only Thing To Measure?

I do not write this to dismiss the people doing benchmark work. The Stanford LegalBench team, the Vals AI grading rubrics, the Atticus Project’s CUAD dataset — these are serious efforts by serious people, and they are trying to move the field forward to provide lawyer/users some way to make decisions about tools. Reliable benchmarks would be valuable to everyone involved.

The problem is more specific. The artifact that legal AI produces like a brief, a motion, a clause, a redlined contract is not a thing whose quality can be evaluated by inspection alone. A person who does not regularly deal with contracts can review some AI output of a redlined agreement and say ,”looks really good.” Meanwhile, an experience contracts lawyer sees the same output as flawed.

Legal AI producing a motion or appellate brief for filing is creating an instrument designed to produce a future event: a judge ruling for your client. The only fully honest answer to “is this AI output good?” is whether the future event went your way. Benchmarks measure proxies for that. The proxies are useful. They are not the thing. The map is not the territory.

This article is intended to examine what current legal-AI benchmarks actually measure, where they help, where they break, and what I think the real metric is.

What the Benchmarks Actually Measure

Three benchmark families dominate legal AI in 2026.

The first is LegalBench, the Stanford-led academic project that compiled 162 tasks covering six types of legal reasoning, hand-crafted by legal professionals around the IRAC framework — Issue, Rule, Application, Conclusion (Stanford LegalBench white paper). The tasks are short, structured, and individually defensible: rule recall, rule application to fact patterns, issue spotting, conclusion derivation. LegalBench is the closest thing the field has to a public, neutral measurement, and the published frontier-model scores cluster in the high 80s.

The second is Vals AI, a private benchmarking firm that produces head-to-head evaluations of legal AI products. Vals runs its own rubrics, authored and peer-reviewed by practicing lawyers from contributing firms like Reed Smith, Paul Hastings, Paul Weiss, and Ogletree Deakins.1 Its February 2025 study compared CoCounsel, Harvey, Vincent AI, and Oliver across seven tasks — Data Extraction, Document Q&A, Summarization, Redlining, Transcript Analysis, Chronology Generation, EDGAR Research. Harvey topped five of six entered categories, with a 94.8% on Document Q&A against a 70.1% lawyer baseline (LawSites coverage of Vals; Artificial Lawyer).

The third is CUAD, the Contract Understanding Atticus Dataset, which the Atticus Project released in 2021. CUAD is 13,000+ labels in 510 commercial contracts pulled from EDGAR, identifying 41 clause types that matter in M&A diligence (Atticus Project CUAD). It is the dominant academic benchmark for contract review. Transformer-based models now achieve respectable scores on it, and every contract-AI vendor in the market has either tested against it or quietly trained on something that looks a lot like it.

There is real value in this work. LegalBench tells us, with some confidence, that a frontier model can issue-spot at roughly the level of a competent first-year associate on isolated fact patterns. Vals tells us that on document-extraction tasks defined by lawyer-authored rubrics, the best legal-AI products genuinely beat a control group of contract lawyers asked to do the same work. CUAD tells us that clause classification on M&A contracts is approximately a solved problem. None of these results should be dismissed.

But notice what each of them does, and what each of them does not do. LegalBench measures whether a model can answer a question that a legal academic designed to have a defensible right answer. Vals measures whether a model can produce an output that a lawyer-graded rubric will award points to. CUAD measures whether a model can correctly label a clause type. None of them measures whether the model’s output, used by a real lawyer in a real matter, produced a better legal outcome than the alternative the lawyer would have written alone. That is the gap I lawyers need to take seriously. This is the gap that matters. It’s the gap that legal benchmarks no matter how well designed, might never be able to reach.

The Useful Work Benchmarks Do

I want to be fair to the benchmark project before tearing into it, because the case for benchmarks is real and I do not want to lose it.

Benchmarks establish a floor. Before the Stanford RegLab hallucination paper, Magesh et al., published as a preprint in May 2024 and peer-reviewed into the Journal of Empirical Legal Studies in 2025 the dominant vendor claim was that retrieval-augmented legal AI did not hallucinate. After the paper, that claim died. Lexis+ AI hallucinated on 17% of test queries. Westlaw AI-Assisted Research hallucinated on 33%. Even Ask Practical Law hallucinated. Bare GPT-4 hallucinated on 43% (Stanford HAI summary). The vendors did not appreciate the paper. Some of them tried to re-evaluate the methodology. None of them, ultimately, has been able to keep using the words “hallucination-free” without footnotes (Wiley/JELS final version).

Can LLMs be modified to never hallucinate? Yes. Would that kill their ability to perform well in creative tasks like poetry writing, business analysis, etc? Yes.

This is what good benchmarks do. They translate a marketing claim into a falsifiable proposition. They give a state bar disciplinary panel something to point at. They give a solo with a $500-per-seat-per-month decision an external reference point. The 2026 follow-up HAQQ study that graded 3,000 answers found that 24% cited or misapplied law that did not support the claim, and that every model tested fabricated or misapplied at least one citation (HAQQ benchmark). That is information you cannot extract from a vendor demo or a sales call. The benchmark made it visible.

Benchmarks also do useful internal work inside a firm. If you build a contract-review pipeline using a local model and a RAG index of your prior contracts, you need some way to know whether the new prompt template you just deployed broke something. A held-out test set of contracts with known answers is the right tool for that job. The benchmark there is not a public ranking; it is a regression test. That is exactly what software engineering has done with unit tests for forty years, and it is exactly what legal-AI workflows should be doing now.

The honest summary of the pro-benchmark case is this: benchmarks are very good at telling you when an AI tool is bad, very good at policing vendor marketing, and useful for internal quality control. They are not very good at telling you when an AI tool is good.

Where Benchmarks Break

The benchmark project has four structural problems, and you should understand all four before you make a purchasing decision based on a leaderboard.

The first is contamination. Every published benchmark eventually leaks into the training data of the next model. MMLU, the dominant general-knowledge benchmark, has been so thoroughly absorbed into training corpora that scores are no longer reliable measures of underlying capability; researchers have documented gains of up to ten percentage points on seen versus held-out versions of the same test, and StarCoder-7b once scored 4.9x higher on leaked versus clean data (MindStudio on benchmark gaming). LegalBench is six years old. CUAD is five. Any model trained in 2025 has plausibly seen both, and you have no way to know how much of the reported score is genuine reasoning and how much is memorization. The frontier benchmark teams know this and rotate test sets. Most of legal AI does not.

The second is Goodhart’s Law: when a measure becomes a target, it ceases to be a good measure (TDWI explainer). Once Harvey’s marketing department puts the Vals AI numbers in a slide deck, every subsequent Harvey release is optimized for the Vals This is akin to not teaching high school students subject matter at all, but instead inspecting the questions on the SAT and teaching those subjects and test taking skills for four years. That optimization may produce real capability gains. It may also produce surface-level improvements that look like capability on the benchmark and feel like nothing on a real deal. The Kimi K2 case showed a 50% claimed versus 29.4% independently measured gap on HLE. Chatbot Arena has been gamed to the point that model developers openly trade tricks for inflating Arena scores. There is no reason to think legal benchmarks are immune.

The third is vendor opt-out. When Vals AI ran its October 2025 legal research benchmark neither LexisNexis nor Thomson Reuters opted in (Legal IT Insider). The two market leaders, who between them probably account for more than half of all paid legal-AI research seats in the United States, simply declined to be measured. Vals graded what it could; the benchmark is genuinely useful for the products that participated; but the headline question, how do the actual market leaders perform on rigorous, lawyer-authored research rubrics?, was unanswered. This is not Vals’ fault. It is a structural feature of any benchmarking regime that depends on vendor cooperation. The vendors with the most to lose can always walk.

The fourth is the validity problem, and it is the deepest. “Ecological validity” is the term researchers use for the gap between the experimental task and the real-world task it is meant to predict. A LegalBench rule-application task gives the model a clean fact pattern, an isolated rule, and asks for an isolated conclusion. A real legal matter gives the lawyer a confused client, a thousand pages of email and texts, a strategic objective that may shift mid-engagement, an opposing counsel with leverage, a judge with a docket, and a budget. The two situations are not the same situation. A model that aces LegalBench is probably better than one that fails LegalBench, but it is “probably” doing a lot of work in that sentence (Reality Check (arXiv)).

The Real Metric for Legal Writing Is Persuasion

Here is the part of this argument I want to plant a flag on. Legal writing is not a fact-recovery exercise. It is a persuasion exercise. The brief exists to make a judge rule for your client. The motion exists to make a judge issue an order in your favor. The demand letter exists to make a counterparty pay or settle. The complaint exists to survive a motion to dismiss. There is no benchmark in the legal-AI literature that measures any of those outcomes against an AI-drafted alternative, because, and this is the part that should be uncomfortable for the benchmark industry, those outcomes cannot be measured cleanly.

The right experiment, if you could run it, would be a randomized controlled trial. Take 1,000 similar motions across similar courts. Randomly assign half to be drafted by a lawyer working alone and half to be drafted by a lawyer working with a specific AI tool, holding the matter facts and procedural posture constant. Measure win rates. Run the same experiment with a different tool and a different lawyer pool. Compare. You would have a benchmark that actually measures the thing the tool is for.

You cannot run that experiment. The institutional review board would not let you randomize representation. Clients would not consent. Judges would not cooperate. Matter facts cannot be held constant. Lawyer skill is the dominant variable and cannot be blinded. Even if you could overcome all that, a single judge’s docket has too few comparable matters to power the statistics in any reasonable window. The metric is correct, the metric is honest, and the metric is unmeasurable.

What Gets Reported Instead

Vals’ “all-pass” rubric score, LegalBench accuracy, citation-fidelity percentages, head-to-head expert preference judgments is what could be measured, not what should be measured. There is real risk in confusing the two. A demand letter that scores beautifully on a citation-correctness rubric and ineffectively on the recipient is a worse tool than one that scores poorly on citations and gets the check cut.

Predictive analytics like Lex Machina, Premonition, judge-specific persuasion engines point at the real metric without quite reaching it (ABA on predictive analytics). They forecast outcomes given case characteristics. They do not forecast outcomes given AI-drafted versus human-drafted briefs in the same case. That comparison is still missing from the public literature, and until someone funds it, every claim that “AI improves legal writing” rests on lawyer self-report and the structural assumption that benchmark proxies generalize. They might. They might not.

The Contract Problem Is Worse

If the brief-writing metric is hard, the contract-drafting metric is structurally worse. A well-drafted contract has two jobs: to allocate risk in a way the parties accept at signing, and to allocate risk in a way that holds up six, twelve, or twenty years later when something the parties did not anticipate happens. The second job is the harder one, and it is the one no benchmark touches.

The Atticus CUAD score on a contract tells you whether a model correctly labeled the indemnification clause. It does not tell you whether the indemnification clause, as drafted, will be interpreted in your client’s favor by a Delaware Chancery Court judge in 2032 when the counterparty is acquired by a private equity sponsor and the integration goes badly. The Vals redlining benchmark tells you whether the AI’s proposed changes match a lawyer-rubric of what a competent redline looks like. It does not tell you whether the redlined contract will avoid the dispute that the unredlined contract would have produced.

The contract is an instrument designed to govern a future state of the world that does not yet exist. The metric for whether it is well-drafted is whether the parties never end up in dispute over what it means or, when they do, the dispute resolves in the client’s favor cheaply. That is observable in principle, but the observation window is years long. The confounders are everything that happens to the relationship in the meantime, and the baseline question of “what would the dispute have looked like under the version a different model would have drafted?” has no answer in this universe.

The contract-AI industry has started to acknowledge this obliquely. Vendors increasingly position themselves around “dispute prevention.” They emphasize their tool’s ability at flagging inconsistent language, surfacing risk clauses, alerting on post-execution compliance (ICertis on contract drafting; American Arbitration Association). The framing is honest about where the value is. But there is still no published study comparing dispute rates across AI-drafted versus human-drafted contracts of equivalent complexity, because the same problem applies. The dispute window is too long, the confounders are too many, and the counterparties did not consent to a controlled trial.

So What Should We Actually Measure?

I want to end constructively, because “benchmarks are flawed and outcomes are unmeasurable” is not a useful place to leave a working lawyer. Here is the metric stack I have settled on for my own practice and the AI tools I build for solos.

Use the Stanford hallucination paper, the Vals rubrics, and a head-to-head comparison like the Cambridge Journal of Legal Information evaluation of ChatGPT-4, Copilot, DeepSeek, Lexis+ AI, and Llama 3 (Cambridge Core) as a sanity floor. If a vendor will not participate in independent benchmarking, treat that the way you would a witness who refuses to be cross-examined.

But, in the end, as with forest fires, only you can prevent filing documents with hallucinations.

Outcome tracking with humility. Track win rates, settlement rates, and time-to-close across AI-assisted and non-AI-assisted matters. Do not believe small samples. Look for trends across two- and three-year windows. Acknowledge that lawyer skill, judge assignment, and matter facts dominate, and that AI is at best a marginal contributor. The point is not to prove the AI works; the point is to detect, early, the matter where it stops working.

A standing rule against confusing benchmark wins for outcomes. If a vendor tells you their model scored 94% on Vals or 89% on LegalBench, that is information about the model on the test, not information about the model on your work. The right response is “that is interesting; let me run our internal regression set and see.”

The closing thought, which I think is right but cannot prove: benchmarks are a floor, not a ceiling. They police the worst, they discipline vendor claims, and they let you move faster on internal quality control than you could without them. They cannot tell you whether the brief will win, whether the contract will hold, or whether the client is better off. Those answers come from the work, the judge, the counterparty, and a longer time horizon than any leaderboard can measure. Use the benchmarks. Do not mistake them for the verdict.

The phrase “personal computer” which morphed into just “pc” came about because the original computers were not so personal. They were large, room-filling monstrosities, using reels of magnetic tape to record and recall data. Eventually, that technology got reduced in size, increased in speed and storage and was stuffed into a box that someone could put on a desk and the personal computer was born.

Until the personal computer era, data was “out there” on what was called a main frame computer. You wanted to do some computing, you had to go to where that big beast was and wait in line effectively. Personal computers changed all that. Now, all your data was on a device on your desk or table and companies began selling software packages to you on disks to help you do things. Early versions of word processors, spreadsheets and even image creation/editing tools were picked up at local stores on sets of disks and brought home to be used.

Internet Changes Personal and Application Data Storage

The Internet changed software delivery. Now, companies could produce software and you could just download it directly to your computer and use it. Still, it was operating on data on your computer and anytime the software had an update, you had to go to a website, click some things and download that update.

That era was the use of software “on-prem” as tech nerds call it. Meaning, the data and software were stored on your premises. Data security was limited to preventing someone from gaining access to your desktop computer and that was fairly straightforward. True hacking, then and now, is typically not done by geniuses figuring out how to bypass all your controls. It is still the old fashioned social engineering. Tricking users into thinking they are tech support or a relative or whatever and the user gives up credentials.

Your Data Leaves On-Prem

Then, the Internet got fast and stable enough so users could just purchase subscriptions to software that ran somewhere else, not on-prem and they could use to from that central location. Obvious advantages to the software company. They could more quickly fix bugs, add features all to one location and make it automatically available to all users. The last step which we have been in for a decade or more is the cloud era. Cloud is a made up term for “someone else’s computer.” There is no could out there. There are just racks and racks of computers (called servers) that hold user data and hold the applications processing that data. All of us users got used to the convenience of it all. New computer, data is out there, no need to transfer from old computer. New iPhone, same convenience. Worries about securing your data, gone - the application provider does that.

Your Data Becomes The Product

Facebook was the most notable practitioner of making the user the product. More precisely, make the user’s data the product. Google and so many other companies, the same approach. Here is a bunch of convenient and free stuff, just give us all your data. From that pile of data, machine learning and now LLMs aggregated it, sliced and diced it and sold if off to uncountable advertisers and others to try and sell stuff.

AI Had To Be The Cloud - Initially

The first widely used LLM, ChatGPT, had to be an application in “the cloud.” What was really happening was that it was such a large piece of software that no laptop or desktop had sufficient storage (hard drive space) or RAM (Memory) to run it. We all connected to it from our machines providing it data both public and personal.

Like Microsoft and Facebook and Google before it, OpenAI realized immediately the value of collecting all that data and also user behavior with its software (LLM).

I am not the first to notice that OpenAI has established a pattern for years now of noticing user behavior, i.e. huge upticks in token usage, then evaluating the business causing the spike in usage and copying that business as their own offering.

LegalAI Follows Suit

There are now hundreds of legal ai tech companies at this point. The most well known ones are using the same approach: Software in the cloud, lawyers/users have to put their data in the cloud for the AI to work on it.

Solo and Small Firm Lawyers - Whose Data Is It?

If you practice as a solo or in a firm of fewer than ten lawyers, you have probably had this argument with yourself in the last six months. On one side: the frontier cloud models — Claude, GPT-5, Gemini 3 — and the legal products built on them (e.g. Lexis+AI, CoCounsel, Harvey, Spellbook). On the other: purchasing your own equipment, installing it in your office, running open source models on it (read, zero token cost) where every prompt and every client document stays on your office network. Security, lower cost, stable access to AI.

The conventional answer — “use the cloud, it’s better and cheaper” — was true in 2023. It is not obviously true in 2026. Frontier models still win on raw capability, but the gap to the best open-weight models has narrowed to single digits on benchmarks that matter for legal work, cloud vendors’ terms of service have started behaving like a moving target, and a federal magistrate in Manhattan has already issued a preservation order that put hundreds of millions of ChatGPT conversations behind a litigation hold the users never agreed to. Meanwhile, a computer for your office that can run useful open source models now costs about a year or so of single, expensive Legal AI license.

The rest of this article walks you through the seven variables that actually matter — cost, confidentiality, ToS volatility1, regulation, the performance gap, speed, and real-world solo usage — and ends with a hybrid architecture suggestion that gets you most of the upside of both worlds.

The Cost Question Is Not What You Were Told

Comparing a $20/month ChatGPT Plus subscription to the price of a server with a GPU2 and concluding the cloud always wins is the wrong comparison. What we do as lawyers is rarely time-sensitive like it would be in emergency medicine, or disaster response, air traffic control, etc. The comparison is not the Ferrari to the Honda. The comparison is, what work do you need to accomplish? LLMs are great data analysts and especially so with text. But, the most powerful LLMs are not required for summarization, text parsing3, motion response drafting, etc. In short, the majority of what most lawyers do can be accomplished in 24 hours or less with open source models running on a computer in your office with client data never leaving your control.

Published pricing in early 2026 looks roughly like this: CoCounsel runs $220–$500 per attorney per month, with a fully bundled Westlaw-plus-AI seat closer to $400–$600 (Legal Prompts). Lexis+ AI runs $175–$400 before content add-ons. Harvey is priced at $1,000 per seat and up. The mid-market tools — Spellbook, Irys, the various solo-focused products — cluster at $69–$149 per attorney per month (Irys pricing comparison).

Now the local math. A used RTX 4090 (24GB) sits at $1,600–$2,300; an RTX 5090 (32GB) at $2,500–$3,200. But, consider that these would last the typical law office with efficient AI operation for 3 years at least. To properly consider that expense, amortize it over 3 years. An AI workstation is $3,500–$5,000 all-in. At that tier you can run open source models like Mistral, Llama and Qwen at the same production speeds as OpenAI and Anthropic models (PromptQuorum hardware guide). The electricity cost is $9–$30 per month at U.S. average rates.

Against a $300/month subscription the workstation pays off in about a year. Against $600/month, seven or eight months. Against a $1,000 Harvey seat, five. Once paid off, the marginal cost of a million additional tokens is rounding error of about a thousandth of a cent in electricity for a Llama-class model, against roughly two dollars per million tokens for a GPT-4.1-class API call (Apatero hardware guide). For a solo running tens of thousands of tokens daily that compounds.

Cloud is cheaper if you barely use AI, much more expensive if you use it heavily. But, with an on-premise AI setup, you can dive into all the features that AI can perform for you. You can experiment. You can try new tools and techniques all the while spending $0 on tokens. At LegalAI.com we recently moved a development version of our offering to entirely on-prem using free open source models. The testing is going well and all the text processing goodness is easily reproducible using free open source models. The future of not sharing client data, not paying for AI and not worrying about the changing Terms of Service or governmentally controlled model access are over.

Confidentiality: The Issue The ABA Made Inescapable

In July 2024 the ABA issued Formal Opinion 512, the first ABA-level ethics guidance on generative AI (ABA news). The confidentiality holding is short and consequential: lawyers must know how a generative AI tool uses the data they feed it and put adequate safeguards against unwitting or unauthorized disclosure in place. Boilerplate consent in an engagement letter, the ABA said expressly, is not enough; informed consent before client confidences go into a third-party model is the floor (UNC Law Library on Op. 512).

This rule quietly tilts the field. With an on-premises open-source model, you can document for any client that no data left your network. The diligence question becomes physical: admin access, office keys, backup policy. These are problems you already know how to solve under Model Rule 1.6.

With a cloud model, the diligence question is contractual and depends on which contract you signed. ChatGPT Plus and the consumer APIs are not built for privileged data and the vendor says so. ChatGPT Enterprise, Team, and the OpenAI API’s zero data retention option are: under the Services Agreement effective May 31, 2025, OpenAI does not retain prompts or outputs sent through the ZDR API and must delete customer content within thirty days of termination (OpenAI Services Agreement). Anthropic’s Claude Enterprise terms are structurally similar. The cloud option is available on terms that satisfy 512, but only if you subscribe at the right tier and read the contract, not the marketing page. None of this says anything about sending PHI to these tools and whether they are handling it consistent with HIPAA. Even if they are, HIPAA requires lawyers storing or transmitting PHI to a third party to have a signed Business Associate Agreement (BAA) with that vendor. Do you have one with your AI vendor? Have you ever had one with your AI vendor?

The uncomfortable gap is between what solos are actually using (consumer ChatGPT, free Gemini, the chatbot in a productivity suite) and what those tiers actually promise. The 2025 ABA TechReport found 62–64% of solos and 2-9 attorney firms preferred ChatGPT, well above their preference for legal-specific tools (LawSites coverage). That is a confidentiality and HIPAA violating disaster many lawyers walked right into.

The Terms of Service Are a Moving Target

The single most underappreciated risk in cloud AI for lawyers is that the vendor can change the rules underneath you, and a court can change them on top of the vendor.

On May 13, 2025, Magistrate Judge Ona Wang in SDNY issued a preservation order in New York Times v. OpenAIdirecting OpenAI “to preserve and segregate all output log data that would otherwise be deleted on a going forward basis.” The order swept in ChatGPT Free, Plus, Pro, and the standard API. ChatGPT Enterprise and the ZDR API were carved out. Judge Stein denied OpenAI’s objections in late June 2025; the preservation obligation was wound down in late September 2025 (OpenAI’s response). In November 2025, Judge Wang ordered OpenAI to produce 20 million de-identified ChatGPT logs to the news plaintiffs (National Law Review analysis).

Pause on that. For roughly four and a half months, conversations users had with ChatGPT under a published thirty-day deletion policy were held under court order in violation of that policy, without notice. Enterprise customers were carved out because their contracts were stronger. Consumer and standard-API users were not. If you were pasting deposition outlines into ChatGPT Plus during that window, those drafts existed somewhere they were not supposed to exist, and a different judge could plausibly have ordered targeted production rather than aggregate disclosure.

A locally hosted Llama or Mistral cannot be subpoenaed out from under you, because there is no third party in the chain. That alone is a strong argument for keeping the most sensitive work on a model you control end-to-end.

The vendor-side ToS volatility is the second piece. OpenAI revised its Usage Policies on October 29, 2025 to explicitly prohibit “personalized legal, medical, or financial advice” unless overseen by a licensed professional. The change came with no negotiation and no grandfathering. Foundation model providers can — and routinely do — tighten use restrictions, change pricing, deprecate models, and add safety filters that break workflows you depend on. A Llama or Mistral checkpoint on your disk will run identically in five years.

Anthropic recently updated its ToS with these little client confidentiality destroying bits of language4.

• Standard API & Chat: Standard API logs and consumer chats are retained for up to 30 days and then automatically deleted.

• Model Training: Standard API inputs/outputs and deleted user chats are never used to train Anthropic’s models.

• Flagged Data: If a chat or prompt is flagged for violating the Acceptable Use / Usage Policy, Anthropic may retain that data for up to 2 years to aid investigations and enforce safety guidelines.

• Legal Requirements: Data can also be retained longer if required to comply with the law or if part of a safety investigation.

The Regulatory Horizon Is Mostly Behind the Cloud

The EU AI Act is the most consequential AI regulation in force right now. After the May 7, 2026 Digital Omnibus agreement, the main high-risk obligations for Annex III systems were pushed from August 2026 to December 2027 (Bird & Bird analysis). The general-purpose AI (GPAI) provisions and certain prohibitions have been in force since 2025. Foundation model providers now face direct fines of up to €15 million or 3% of global revenue as punishments for GPAI non-compliance (Lyceum on GPAI obligations).

Two things follow for U.S. solos. First, foundation-model compliance costs will be passed through to you. Cloud pricing has held on competitive pressure, but the trend is up. Second, the U.S. picture remains fragmented but is converging. State bar opinions in California, Florida, New York, and Pennsylvania largely track ABA 512. Several states now require disclosure when AI is used in pleadings, and federal standing orders requiring AI-use certifications are common. None of this directly punishes open-source use, and several state opinions explicitly note that local deployment reduces confidentiality risk.

The regulatory risk to cloud legal AI is not that it will be banned. It is that disclosure, audit, and explainability obligations on foundation model providers will create deployer-level friction including more vendor-due-diligence forms, more model cards, more contract amendments every time a vendor swaps the underlying model. A lawyer with a local AI setup has none of that overhead for concerns.

Pricing Will Change

I am not the first to notice that neither OpenAI nor Anthropic whom are both contemplating IPOs in 2026 have not released the actual cost to them to produce each token of AI software usage. One can speculate that they have been subsidizing token costs to users. Anthropic’s CEO recently said in an interview that without significant AI growth of his company over a sustained period, it will go bankrupt.5

That likely means that immediately post-IPO you should expect the cost to use these AI tools will increase dramatically. At some point, the cost of cloud based AI tools will be prohibitive versus its value, especially for solo and small firm users.

The Performance Gap Is Small

LegalBench, the Stanford-maintained legal reasoning benchmark, places the top foundation (read: exclusively cloud based) models, Claude Fable 5, Gemini 3 Pro, GPT-5.5 in the 86–89% accuracy band. The strongest open-weight entry, Llama 3.1 405B, sits within a few points of GPT-4o and Claude 3.5 Sonnet, and Llama 3.1 70B is on the Pareto frontier of cost versus performance (Vals.ai LegalBench).

The gap between the best closed model and the best open-weight model you could run yourself is roughly 4–6 points on all the representative legal benchmarks.6 That is not zero, and on the hardest tasks like multi-hop rule application and novel statutory reasoning the frontier models are clearly stronger. As I have said many times on X.com and here, so what? A Ferrari and a Honda can take you to the store to get groceries. Do you need a Ferrari to perform legal work or a Honda? As much as we lawyers value our expertise and for some of us our writing ability, the reality is that the easiest tasks for LLMs is text manipulation. Image creation, video creation/editing, etc. are very processor intensive for LLMs. All the things lawyers do to prepare a first draft of a document are basic skills for any LLM.

Remember that the largest LegalAI companies, Legora and Harvey, were started more than 3 years ago. They were attracting customers then to the benefits of tools from OpenAI that today’s open source models can out perform easily. Since LegalAI tools that underperform today’s open source models were valuable enough that lawyers were paying Legora and Harvey back then, it stands to reason that they are even more useful today.

But for the bread and butter of solo work like summarizing depositions, reviewing contracts against a checklist, drafting demand letters, first-pass document review, generating discovery requests, etc a quantized (reduced size) Mistral Small 3.1 or Qwen 3 32B is in the same ballpark as a frontier model from 6 months ago, and that older frontier model was already good enough that lawyers were paying $200 a month for access.

These numbers are also zero-shot. With an automated vector database setup like legalai.com pointed at your own document corpus like your matter files, your drafting templates the performance gap disappears for work where the answer comes from documents you already have. The Geek Law Blog made the point directly when DeepSeek R1 was released: the reasoning is “phenomenal,” and local deployment addresses the data-security issue that has been the central blocker to legal-AI adoption (Geek Law Blog).

Speed: The Real Answer Surprised Me

Cloud advocates say frontier models are faster. Local advocates say you can wait 24 hours for the same result. Both miss the point.

Latency in 2026 looks like this. A frontier API streams at 50–150 tokens per second (tps). Mistral Small 3.1 24B on an RTX 4090 returns at 30–55 tokens per second; Llama 3.3 70B on the same card, 4-bit quantized at 15–25 (tps). For a 1,500-word draft, that is the difference between a 25-second wait and a 90-second wait. Real, but not life-altering. Given the confidentiality and cost issues, you have to ask yourself if you it is worth it just to have a document produced a minute or so sooner?

The way we work today is more akin to open source model use anyhow. Lawyers receive a motion, get a call, write down a task and work is performed based upon priorities of time, or financial risk, etc. Then, the work is scheduled. Rarely is it that we receive a motion or call or whatever and we need to produce and file something such that the open source model performance and its 90-second wait is relevant.

What actually matters is not how fast the first draft arrives, it is how many drafts you go through before you ship. A local model that returns a redline in 90 seconds is not slower than a cloud model at 25 seconds, because the bottleneck is your review, not the generation. The speed argument bites in agentic workflows where the model is invoked dozens of times in a chain. For example, a brief-checking agent that hits the model once per cited authority. There, cumulative token usage adds up and costs that can easily get out of control using foundational models. For interactive use like chat drafting, ask-the-document Q&A, single-shot summarization, etc. there is no performance difference between open source on-premises AI and on cloud.

The thing that does matter is reliability. A frontier API that runs fast on tasks that do not need speed but stops working at 2pm on a weekday because you used up all your tokens is not as good as tokens that never run out on your open source models. Cloud outages do not care about your filing deadline.

How Lawyers Actually Use AI, Right Now

The 2025 Clio Legal Trends Report found that 79% of legal professionals integrate AI into their practice in some form, that firms which adopt it widely are 69% more likely to see revenue gains, and that as much as 74% of billable work could in principle be automated with AI tools (Clio 2025 press release). The 2025 ABA TechReport found that 54% of legal professionals use AI to draft correspondence and 14% use it to analyze firm matters (LawSites). Solo adoption of legal-specific AI is around 18% versus 46% at firms of 100+, but informal use of consumer chat tools is much higher across the board.

The pattern that emerges from the survey data is that solos use AI for high-volume, low-stakes, document-shaped tasks: client correspondence, intake summaries, deposition outlines, demand letters, first-pass contract review, brainstorming arguments, drafting form motions, and converting messy facts into clean chronologies. They do not, by and large, use it for the most cognitively heavy work like appellate briefs, novel statutory analysis and trial strategy.

That usage pattern is exactly the pattern where local open-source models are most competitive. The hard parts of the job that justify a frontier-model premium are also the parts the survey data says solos rarely automate.

The flip side, soberly stated: lawyers who automate without verifying get burned. Mata v. Avianca in 2023 was the canonical case, but it was not the last. The Northern District of Alabama sanctioned a “large and well-regarded” firm in Johnson v. Dunn in July 2025 for filing hallucinated citations, and aggregated databases now track more than 200 AI-hallucination sanctions cases recorded in just the first eight months of 2025 (Jones Walker analysis). Cloud or local, the duty to verify every citation does not move.

So, Cloud or Local?

If you are one of the rare lawyers who is performing very processing intensive work, editing images, video, analyzing 8,000 documents in every case to prepare to settle or try it, etc then perhaps foundational models are where you have to go. For the majority of us, however, the costs, data security and reliability are with open source models running in our local offices.

The cloud is faster on the absolute frontier, has lower upfront cost, and ships the convenience of someone else’s infrastructure. It also exposes you to vendor terms you do not control, preservation orders you cannot opt out of at the consumer tier, regulatory friction that is going to get worse before it gets better, and a steady tax on every billable hour you put through it.

The local stack is slightly slower at the frontier, involves some setup and demands real diligence about hardware and key management. It also gives you complete confidentiality, immunity to vendor and regulatory volatility, asymptotic costs near zero, and a system that will run identically in five years.

The days of reflexively relying on third party cloud AI tools is over. The open source models are here and for legal work, they are as good as foundational model counterparts. The only objective difference is speed and that difference is one without relevance for most lawyers anyhow.

Fair use is a legal doctrine that promotes freedom of expression by permitting the unlicensed use of copyright-protected works in certain circumstances. Section 107 of the Copyright Act provides the statutory framework for determining whether something is a fair use and identifies certain types of uses—such as criticism, comment, news reporting, teaching, scholarship, and research—as examples of activities that may qualify as fair use. Copyright.gov

Although AI so-called as part of the technology feature of the main legal issue, the argument was not about large language models (LLMs) ingesting the open web. ROSS was not ChatGPT. Westlaw headnotes are not novels. The alleged copying was not the downloading of books into a foundation model. The copyrighted content at issue was Westlaw’s Headnotes feature. That feature was a creation of West Publishing, years before legal research was electronic, as a discovery guide to relevant cases.

ROSS Intelligence developed a legal research tool using Machine Learning largely that ingested these West Publishing headnotes as training material. If this sounds familiar, it should. This is precisely what OpenAI and Anthropic did to develop their billion dollar companies - they ingested copyrighted material to train their models. Like ROSS, those companies and others argued, “hey we didn’t “copy” that stuff, and even if we did, we don’t output it from our models (but some have)1 we merely used it as one side of an example of how to properly summarize an opinion, etc.

Westlaw did not like that at all. This tool that ROSS created returned cited judicial passages in response to natural-language questions. At that time, Westlaw offered no such tool. They do now after buying an entire company to get it. Westlaw is an entrenched legal research platform objecting to a would-be legal research competitor.

That is why the oral argument matters. The panel did not appear interested in using the case as a referendum on “AI training” in the abstract. The judges seemed much more interested in a narrower set of questions: What is the relevant market under factor four? Can the market be defined as “Westlaw” or “legal research platforms,” or must it be defined as the market for the particular copyrighted works allegedly copied, namely Westlaw headnotes? Does it matter that ROSS’s final product did not output headnotes? Does it matter that ROSS directly competed with Westlaw? And how thin is the copyright in a headnote that summarizes, quotes, or paraphrases law that no one can own?

Natural Language Queries are not Language At All.

A fundamental reality that hopefully the judges understood is that the ROSS “machine” they built for natural language querying of case law involved numbers. I have explained this in detail in other posts, but a summary here is useful.

To enable natural language querying of case law, the text has to be converted to numbers - specifically, arrays of numbers that look like this. [0,2,5,9.0]. It is those numbers, vectors, on a multi-dimensional space that use math to find similar semantic concepts.

The first step in this process was to convert the text of West headnotes into vectors. The headnotes, you may have noticed, are not numbers, they are text. They have always been text. They are only useful as text (humans don’t use vector arrays to find stuff). ROSS took West Headnotes and transformed them. They used the vectors from that transformation to train a model.

The Posture: An Interlocutory Appeal From a Changed Mind

The appeal comes from Judge Stephanos Bibas’s February 2025 summary judgment opinion in the District of Delaware. Judge Bibas, sitting by designation, had earlier declined to take the fair use issue away from the jury. On renewed briefing, he changed course. He granted partial summary judgment to Thomson Reuters on direct infringement for 2,243 headnotes, rejected several defenses, and granted Thomson Reuters summary judgment against ROSS’s fair use defense.

That procedural posture is not incidental. ROSS’s brief leans hard into the “changed mind” narrative, describing the district court’s decision as an abrupt reversal of an earlier view that factual issues remained. Thomson Reuters responds that renewed briefing was invited in the run-up to trial and that the material facts needed for fair use were not genuinely disputed. At argument, the summary-judgment posture was present but not dominant. The panel seemed less interested in whether Judge Bibas was allowed to revisit his earlier ruling than in whether his second ruling framed factor one and factor four correctly.

The certified questions were, in substance, whether short quotes or paraphrases of judicial holdings can be copyrightable and whether fair use protects ROSS’s internal use of Westlaw headnotes in training data for an AI legal search engine that produced non-infringing outputs. The question as argued, however, felt even more focused: if ROSS copied protected headnote expression to build a direct legal research competitor, is that market substitution enough to defeat fair use even if the outputs did not reproduce the headnotes?

What ROSS Says This Case Is About

ROSS frames the case as an access-to-law and innovation dispute. Its story begins with the premise that no one can own the law. Westlaw headnotes, ROSS argues, are either quotations from judicial opinions or constrained paraphrases of legal holdings. They exist to state legal propositions accurately. That purpose, in ROSS’s view, leaves little room for copyrightable expression.

On copyrightability, ROSS relies on the government edicts doctrine, Feist, Matthew Bender, and related cases policing the boundary between facts, law, and expression. Its point is not simply that judicial opinions are uncopyrightable. Everyone agrees with that. Its stronger point is that a headnote summarizing a specific legal holding is so constrained by the need for accuracy that any protectable expression is either absent or vanishingly thin. If the headnote says what the law is, ROSS says, treating that sentence as an individually copyrightable work risks allowing a private publisher to fence off the law by restating it.

On fair use, ROSS’s opening brief tries to place the case in the line of “intermediate copying” decisions: *Sega*, *Sony v. Connectix*, *Google Books*, *HathiTrust*, *Perfect 10*, *iParadigms*, and *Google v. Oracle*. The common move is familiar. A defendant copies more than the public ever sees, not to republish the plaintiff’s expression, but to build a tool, index, interface, compatibility layer, search function, or analytic system. The copied material is used as an input to generate a socially useful output. On that account, ROSS did not sell headnotes. It used headnote-derived training examples to teach a machine-learning system how to match legal questions to relevant passages in judicial opinions.

ROSS’s cleanest argument is that the relevant market under factor four is the market for the copyrighted work, not the market for the defendant’s final product. In its telling, the copyrighted work is the headnote. There is no separate market for individual headnotes. ROSS did not provide headnotes to users. ROSS did not sell a headnote database. ROSS did not replace a lawyer’s reason to read a Westlaw headnote. The fact that ROSS competed with Westlaw as a legal research product, ROSS says, should not be enough. Copyright does not protect business models from competition. It protects works from cognizable market substitution. Of course, Westlaw did not sell access to headnotes themselves either.

That was the pitch Mark Davies pressed at argument. The panel began almost immediately with factor four. When asked about the effect on the market, ROSS argued that the statute asks about the market for the copyrighted work, and that the work here is the headnote. ROSS resisted defining the market as “AI training” or “legal research platforms,” warning that doing so would make fair use circular: any innovative use could be converted into a licensing market simply because the copyright owner, after seeing the use, says it would have licensed that use.

ROSS also tried to use the Third Circuit’s recent decision in American Society for Testing & Materials v. UpCodes as a fresh doctrinal lever. That decision, issued April 7, 2026, involved access to standards incorporated into law and, according to ROSS, confirms that law-adjacent factual and functional works receive a fair use analysis sensitive to public access and thin protection. Whether the panel UpCodes as a close cousin or a very different access-to-law case may matter a great deal.

What Thomson Reuters Says This Case Is About

Thomson Reuters frames the case as a straightforward case of commercial copying to build a substitute. That framing was also effective at argument because it aligns with a judicial instinct that copyright’s fair use doctrine should not bless a competitor’s decision to take the plaintiff’s protected editorial work after being unable to license it.

Thomson Reuters’s brief insists that headnotes are classic copyrightable legal editorial content. Courts have long distinguished between uncopyrightable judicial opinions and copyrightable reporter-added material. Headnotes, Thomson Reuters argues, involve selection, distillation, synthesis, phrasing, and integration into West’s editorial system. The low originality threshold matters. Even if headnotes are not poetry, they need only contain a modest degree of creative expression to be found protectable.

On fair use, Thomson Reuters’s strongest facts are commercial purpose and direct substitution. ROSS wanted to build a better legal search engine. It approached Thomson Reuters about using Westlaw content and was refused. It then obtained Bulk Memos from LegalEase, which allegedly used Westlaw headnotes as source material. ROSS then trained a legal research product that it marketed against Westlaw. To Thomson Reuters, that is not transformative use in any legally meaningful sense. It is the use of Westlaw’s editorial content to build a Westlaw substitute. I don’t think the mere refusal to license the headnote somehow changes the legal analysis here, but it appears ThomsonReuters’ lawyers did.

That difference in market framing is everything. Thomson Reuters says factor four is not limited to a hypothetical market for standalone headnotes. Westlaw content is monetized through Westlaw. The effect on the market includes harm to Westlaw’s value, harm to Thomson Reuters’s ability to exploit its editorial content in AI products, and harm to actual or potential licensing markets for using Westlaw content as training material. It argues that ROSS’s narrow “market for individual headnotes” approach artificially slices the work from the commercial product through which the work is actually exploited. These arguments beg another question: What if ROSS or some other company merely restates the headnotes, then uses those restated headnotes to train its model? What then of the infringement claim?

At oral argument, Dale Cendali pressed the same theme: ROSS did not merely use headnotes in a remote or unrelated way; it used them to construct a rival product. Thomson Reuters also resisted ROSS’s reliance on intermediate copying cases by emphasizing that software reverse-engineering and search-index cases typically involve different purposes, functional constraints, or non-substitutive outputs. A search index of books does not replace the books. A plagiarism detector does not replace student papers. A compatibility interface does not necessarily replace the original software platform. But a legal research engine trained on Westlaw’s editorial material to compete with Westlaw, Thomson Reuters says, is different in kind.

That argument clearly found some traction in the panel’s questions. The judges pressed ROSS to explain how its final product differed from Westlaw from the user’s perspective. ROSS answered that the internal technology was different and that its system returned judicial passages in response to natural-language questions, but it also conceded the obvious: ROSS was a direct competitor. The judges seemed interested in whether “different internal technology” is enough for transformativeness when the commercial function of the final product overlaps so heavily with the plaintiff’s product. An argument still available relates to the vectors comments above. ROSS’s product was a model trained on vectors which, in turn, were derived from headnotes. Transformation anyone?

The Oral Argument: Factor Four Took the Wheel

The most striking feature of the argument was how quickly the panel moved to factor four. That is not surprising. In modern fair use doctrine, factor four often becomes the place where courts translate intuitive concerns about unfair competition into doctrinal language. It is also the factor most likely to determine how broadly the opinion will matter outside this case.

ROSS wants factor four to focus on the market for the specific protected expression copied. If the headnote is the copyrighted work, the question is whether ROSS displaced demand for headnotes or for licensing headnotes in a traditional, reasonable, or likely-to-develop market. ROSS says no. Thomson Reuters wants the relevant market to include Westlaw as the commercial embodiment of the headnotes and to include licensing markets for AI training. It says ROSS’s copying produced exactly the kind of substitution factor four is designed to catch.

The doctrinal risk for Thomson Reuters is circularity. Courts have long been wary of defining a licensing market at the level of “the market for licensing the very use the defendant claims is fair.” If that were enough, factor four would swallow fair use. Every fair use is, in one sense, a lost license. ROSS exploited that concern at argument, arguing that a copyright owner cannot manufacture market harm simply by saying, after the fact, that it would have charged for the defendant’s transformative use.

The doctrinal risk for ROSS is under-definition. If the market is defined only as individual headnotes sold one by one, the analysis may ignore how legal publishers actually monetize headnotes. Westlaw does not need to sell individual headnotes a la carte for those headnotes to have market value. They are part of a subscription product. They help organize and retrieve law. They differentiate Westlaw from competitors. If a defendant uses those headnotes to train a rival legal research platform, a court may see market harm even if the defendant never displays the headnotes.

The hard legal question is whether that harm is copyright harm or competition harm. ROSS’s best sentence, conceptually, is that copyright is not an anti-competition statute. Thomson Reuters’s best sentence, conceptually, is that fair use does not allow a competitor to take protected expression to build a market substitute. The Third Circuit’s job is to decide which sentence describes these facts. The problem for Westlaw is that, at that time, there was no market they were exploiting for the licensing of headnotes. There is still no market where Westlaw sells or otherwise licenses headnotes themselves. With the advent of RAG which is now providing natural language querying free, the utility of headnotes is zero.

Factor One: Is Training Transformative When the Product Competes?

The factor-one argument was more awkward for ROSS than the briefs might suggest. In writing, ROSS can describe its use at a high level: it transformed headnotes into training signals for an AI legal search engine. That sounds a lot like the search, indexing, and intermediate-use cases. In the courtroom, however, the panel asked a simpler question: what was so different about the ROSS product from what lawyers already do on Westlaw?

ROSS’s answer had two parts. First, its system operated differently at a technical level. It used deep learning and natural-language question answering to return ranked judicial passages. Second, the Bulk Memos were not merely copies of headnotes. They contained questions and multiple answers, including wrong or less-good answers, which were useful for training. In other words, the headnotes were not the final product. They were incorporated into a more complex training apparatus.

That may be enough if the court emphasizes internal use. It may not be enough if the court emphasizes external market purpose. Warhol tightened the lens on purpose and character, reminding courts that a defendant cannot establish transformativeness merely by describing a secondary use at a high level of generality. If the parties’ uses share a substantially similar commercial purpose in the relevant context, the first factor may tilt against fair use.

But the “relevant context” is doing a lot of work. The headnote’s immediate purpose is to summarize a legal point in an opinion. ROSS’s immediate training purpose was to create labeled examples that helped a machine identify responsive passages in judicial opinions. Those are different. The Westlaw platform’s broader purpose is to help lawyers find law. ROSS’s broader purpose was also to help lawyers find law. Those are similar. Fair use doctrine can plausibly focus on either level of abstraction. The district court focused on the broader product-market level. ROSS asks the Third Circuit to focus on the training-use level.

That is why this case is so dangerous as precedent. If the Third Circuit says “AI training is transformative” in a broad way, defendants in generative AI cases will quote it endlessly. If it says “training a competing legal research product is not transformative,” plaintiffs will quote it just as aggressively. The smarter path would be an opinion that makes clear which aspects of the use matter: the thinness of the work, the absence or presence of output substitution, the extent of direct competition, and the plausibility of a licensing market.

Copyrightability: The Issue the Panel May Not Need to Maximize

Although the certified appeal includes copyrightability, the oral argument reporting and transcript suggest that fair use drew the bulk of judicial attention. That may be because the panel can decide the case on fair use without writing the definitive law of Westlaw headnotes. It may also be because the headnotes in the record are sealed, making it harder for an appellate court to write a satisfying originality opinion for public consumption.

Still, copyrightability matters. The thinner the copyright, the stronger ROSS’s fair use position becomes. Even if a headnote crosses the low originality threshold, it sits close to the boundary between protected expression and uncopyrightable law. That should matter under factor two. It should also matter under factor three, because copying the whole of a thin factual or functional work may be more tolerable when the whole is needed for a transformative purpose. And it should matter under factor four, because a thin copyright should not support control over all downstream uses of the legal proposition.

The district court’s sculptor analogy remains a flashpoint. Judge Bibas reasoned that a headnote author may be like a sculptor who selects what to chip away from the raw material of a judicial opinion. ROSS and its amici reject that analogy. In their view, a headnote writer is not creating a new expressive object from raw marble; the writer is constrained to restate a legal holding accurately and concisely. Thomson Reuters, by contrast, says those acts of selection, condensation, synthesis, and phrasing are precisely the editorial judgments copyright has long protected.

The Third Circuit does not need to hold that headnotes are categorically uncopyrightable to give ROSS meaningful relief. It could hold that some headnotes may be copyrightable, but that the district court erred in treating the specific headnotes as protectable individual works at summary judgment without adequate expression-by-expression analysis. Or it could affirm copyrightability while reversing on fair use. Or it could affirm both. The most aggressive move would be a broad rule against copyrightability for individual headnotes that do no more than state legal holdings. That would reverberate well beyond AI.

The Amicus Battle: Access to Law Versus Licensing Markets

The amicus briefing reveals the stakes better than the party briefing does.

EFF, library groups, and legal technology amici argue that the district court’s approach threatens access to law and competition in legal information markets. Their concern is not merely ROSS. It is the ability of legal research startups, public-interest databases, libraries, and researchers to use law-adjacent editorial materials without allowing incumbents to convert thin copyright interests into control over legal information infrastructure. The Free Law Project and legal tech amici make a similar point in more market-oriented terms: overbroad protection for headnotes can entrench incumbents and chill legal technology development.

On the other side, the Copyright Alliance and copyright-owner amici warn that ROSS’s position would create a vast AI exception. If a company can take protected works, convert them into training data, and avoid liability because the outputs do not reproduce the works verbatim, then the licensing markets for high-quality curated content may be gutted before they can mature. For those amici, AI training is not magic. It is a use of copyrighted works, and where that use substitutes for the plaintiff’s market or usurps a licensing opportunity, fair use should not apply.

Both sides are right about something. Legal information markets really are unusually prone to private enclosure because the underlying law is public but the tools for navigating it are private. At the same time, curated editorial work really can be valuable, and copyright law has never said that factual or functional works receive no protection at all. The question is not whether headnotes have value. They do. The question is what kind of value copyright protects against this kind of use.

Why This Case Is Not the New York Times Case, or Anthropic, or Meta

Lawyers should be careful about porting this case wholesale into the generative AI docket. Several features make it unusually plaintiff-friendly.

First, ROSS allegedly used material from a direct competitor after access through ordinary licensing channels was denied. That fact is rhetorically powerful even if bad faith is not supposed to dominate fair use. Second, the defendant’s product was itself a legal research substitute. This is not a case where the defendant trained a general model on copyrighted inputs and produced a tool aimed at a broad range of unrelated uses. Third, the works at issue are thin, factual, and law-adjacent. That helps ROSS, but it also means a ruling for Thomson Reuters can be cabined by plaintiffs in more creative-work cases: if even thin headnotes are protected from this kind of competitive training use, plaintiffs will say, then novels, songs, photographs, and news articles deserve at least as much protection.

At the same time, a ruling for ROSS would be hard for AI plaintiffs to shrug off. If the Third Circuit holds that using Westlaw headnotes to train a direct legal research competitor is fair use, defendants in broader AI cases will argue that their facts are even stronger where the trained model is general-purpose, outputs are non-substitutive, and the plaintiff cannot show direct market replacement.

The Likely Paths

The Third Circuit has several options.

The first is full affirmance. The court could hold that Westlaw headnotes are copyrightable, that ROSS copied protected expression, and that using that expression to build a direct legal research substitute is not fair use. This would be a major win for copyright owners and would be cited immediately in AI training cases. Its breadth would depend on whether the court emphasizes direct competition, non-generative legal search, and the particular Westlaw record.

The second is reversal on fair use with or without resolving copyrightability. The court could hold that ROSS’s use was intermediate, transformative, and non-substitutive as to the protected expression because ROSS did not output headnotes. That would be a major win for AI defendants, especially if the court rejects the AI-training licensing market as circular or insufficiently traditional. Again, the breadth would depend on the writing.

The third is a narrower vacatur or remand. The court could decide that summary judgment was premature because copyrightability, transformativeness, or market harm required more granular factfinding. This path would be doctrinally modest and institutionally attractive. It would avoid making the first appellate AI fair use decision on a sealed, unusual, competitor-versus-competitor record. But it would also leave the bar hungry for guidance.

The fourth is a split decision: affirm copyrightability but remand fair use, or reverse the broad headnote copyrightability ruling while leaving room for protection of sufficiently original headnotes. This may be the most lawyerly outcome. It would recognize that Westlaw’s editorial work is not categorically outside copyright, while also refusing to let the thinnest possible copyright interest control downstream technological use too easily.

My Read

I have a bias (as all people do). I am the founder of www.legalai.com that serves solo and small firm lawyers with a suite of AI tools to save them hours each month. Despite that, my read from the oral argument is that ROSS faces a harder panel than its briefing posture might suggest. The judges’ questions pressed ROSS on the practical similarity between its product and Westlaw, the direct-competitor fact, and the market consequences of allowing a rival to use Westlaw editorial content in training. ROSS’s best answer is doctrinal: the market must be tied to the copyrighted work, not generalized platform competition. But appellate panels are human, and “you copied our editorial content to build a cheaper legal research competitor” is a powerful story.

That said, Thomson Reuters has its own vulnerability. If the court accepts too broad a market definition, factor four becomes a machine for defeating fair use whenever a copyright owner can describe the defendant’s innovation as a potential licensing opportunity. That would sit uneasily with Google Books, HathiTrust, Sega, Sony, and much of the intermediate copying tradition. It would also risk turning copyright into a tool for preserving incumbent control over legal information.

The cleanest opinion would resist both overstatements. It would say that AI training is not automatically fair use and not automatically infringement. It would require courts to ask what was copied, how protectable it was, whether the use served a genuinely different function, whether the outputs substitute for the protected expression, whether the defendant’s product substitutes for the plaintiff’s copyright market rather than merely its business, and whether the asserted licensing market is traditional, reasonable, and likely to develop rather than circularly defined.

In other words, the Third Circuit should not write an “AI exception.” But it also should not write an “AI training market” rule that gives copyright owners veto power over socially useful computational uses of thin, factual, law-adjacent works.

The most important question in is not whether AI companies need data. They do. It is not whether Westlaw headnotes have value. They do. The question is whether the value ROSS allegedly appropriated is the kind of value copyright protects against fair use, or the kind of competitive advantage copyright leaves exposed so that others can build better tools.

That question is old. AI just made it expensive.

Sources Consulted

- Third Circuit oral argument, *Thomson Reuters v. ROSS Intelligence*, No. 25-2153, argued June 11, 2026, available through [CourtListener](https://www.courtlistener.com/audio/105475/thomson-reuters-enterprise-centre-gmbh-v-ross-intelligence-inc/) and the Third Circuit MP3 link in the CourtListener API.

- Judge Bibas’s February 2025 District of Delaware opinion, available from the [District of Delaware](https://www.ded.uscourts.gov/sites/ded/files/opinions/20-613_5.pdf).

- ROSS Intelligence opening brief, No. 25-2153, available via [IPWatchdog](https://ipwatchdog.com/wp-content/uploads/2025/10/PUBLIC-REDACTED-ROSS-CA3-9.22-FILE-1.pdf).

- Thomson Reuters response brief, No. 25-2153, available via [Chat GPT Is Eating the World](https://chatgptiseatingtheworld.com/wp-content/uploads/2025/11/88-Proof-Brief-Thomson-Reuters-Enterprise-Centre-GmbH-et-al-v.-Ross-Intelligence-Inc-Docket-No.-25-02153-3d-Cir.-Jun-24-2025.pdf).

- Kyle Jahner, “First AI Copyright Appeal’s Reach Hinges on Market Impact Issue,” Bloomberg Law, June 10, 2026, PDF copy available from [Florida State University College of Law](https://law.fsu.edu/sites/default/files/2026-06/First%20AI%20Copyright%20Appeal%E2%80%99s%20Reach%20Hinges%20on%20Market%20Impact%20Issue.pdf).

- Melissa Ritti and Emma Whitford, “US Third Circuit zeroes in on fair use, not copyrightability, in Ross appeal,” MLex, June 11, 2026, available at [MLex](https://www.mlex.com/mlex/articles/2488872/us-third-circuit-zeroes-in-on-fair-use-not-copyrightability-in-ross-appeal).

- EFF, ALA, ARL, and ACRL amicus brief, available from [EFF](https://www.eff.org/document/amicus-brief-eff-et-al-thomson-reuters-v-ross-intelligence).

- Free Law Project and legal technology innovators amicus discussion, available from [Free Law Project](https://free.law/2025/10/02/tr-v-ross-amicus/).

- Copyright Alliance amicus brief, available from [Copyright Alliance](https://copyrightalliance.org/wp-content/uploads/2025/11/2025.11.26-Dkt-112_25-2153_Amicus-Brief-Copyright-Alliance.pdf).

- Eileen McDermott, “Amici Back AI Company’s Third Circuit Appeal of Summary Judgment for Thomson Reuters,” available from [IPWatchdog](https://ipwatchdog.com/2025/10/01/amici-back-ai-companys-third-circuit-appeal-summary-judgment-thomson-reuters/).

Imagine the following scenario.

A potential client believes they have been wrongfully terminated. Before contacting an attorney, they spend several evenings discussing the situation with ChatGPT (or Google Gemini or Anthropic’s Claude). They upload emails from supervisors, performance evaluations, text messages from coworkers, and detailed descriptions of workplace events. They ask whether they have a discrimination claim. They ask what evidence matters most. They ask whether their employer’s conduct appears unlawful. They even ask the chatbot to identify weaknesses in their case and estimate the value of potential damages.

Six months later, the client retains counsel and files suit.

Discovery begins. Opposing counsel learns that extensive conversations occurred between the plaintiff and ChatGPT before litigation was filed.

Can those conversations be discovered?

Most lawyers probably have an instinctive reaction to that question. Unfortunately, those reactions vary dramatically. Some assume the communications are private and therefore protected. Others assume they are simply another form of electronic communication that may be subject to discovery. Still others believe the law has not yet developed enough to provide meaningful guidance.

The truth is that courts, ethics authorities, legal commentators, and litigators are only beginning to grapple with the implications of what clients and lawyers share with Large Language Models (LLMs). For the past two years, much of the discussion for lawyers around use of LLMs has focused on hallucinations, accuracy concerns, ethical obligations, and the infamous cases involving attorneys who submitted fictitious AI-generated citations. Those issues deserve attention. Yet they may not ultimately prove to be the most consequential legal questions arising from widespread AI adoption.

What happens when conversations with AI systems become evidence? To be clear, we are talking about data uploads to paid, closed, foundation models - the kind you cannot download to your own computer for offline use.

Closely related to that question is another one lawyers should be asking themselves:

What happens when the lawyer’s own AI conversations become evidence?

The answers may depend heavily on which AI system is being used, where information is stored, and who ultimately controls it.

The Mistake Many People Are Already Making

One reason this issue has received relatively little attention is that many users have unconsciously begun treating AI systems as something they are not.

People increasingly interact with ChatGPT, Claude, Gemini, and similar systems in ways that closely resemble conversations with attorneys, consultants, accountants, physicians, or other trusted advisors. They explain complex factual situations. They ask for analysis. They seek guidance regarding future actions. They reveal sensitive personal information. In some instances, they disclose information they have never shared with another human being.

From a user’s perspective, the interaction can feel remarkably similar to a confidential consultation. From a legal perspective, however, that similarity may be largely irrelevant.

Attorney-client privilege does not arise because someone discusses a legal issue. If it did, conversations with spouses, friends, coworkers, and neighbors would frequently qualify for protection. Instead, privilege exists because of the relationship between attorney and client and because society has concluded that candid communications with legal counsel serve an important public purpose. But, not even all conversations with your lawyer are protected. They have to be conversations specifically related to the legal representation. Conversations about sports, politics, crypto or whatever, not protected by privilege assuming the matter the lawyer is handling does not deal with those topics.

A chatbot is not an attorney. It is not licensed to practice law. It owes no fiduciary duty to the user. It is not subject to professional discipline. Most importantly, it generally does not occupy the special legal relationship that gives rise to attorney-client privilege.

That distinction may seem obvious when stated directly. Yet it becomes less obvious when viewed through the lens of actual user behavior. Many individuals now turn to AI systems before they turn to lawyers. In some cases, they use AI systems to decide whether they even need a lawyer in the first place. Courts seem to be focused on the notion that uploading data, documents, questions, etc to a chatbot is “sharing” that data outside the attorney client privilege. Assuming for a moment that the client at issue is doing so with a subscription to one of these tools, consider what that means.

What the court in Heppner did not recognize is that LLMs are software. That’s it. Don’t believe me, ask Anthropic, that is what they call their LLM. Just software. Recognizing this, the import of the Heppner decision means, if a client shares his thoughts in a Google Doc, or Spreadsheet, etc, that information is therefore discoverable. But, is that how people regard their personal google account - as essentially a discoverable area of their lives? Perhaps they do not and perhaps courts do not care.

As that trend continues, courts will inevitably be asked to determine whether those conversations receive any special protection.

Recent commentary surrounding Heppner suggests that at least some courts may be reluctant to extend traditional privilege protections to communications with AI systems merely because legal subjects are being discussed.¹ ² While the law remains in its infancy, the direction of the conversation should get lawyers’ attention.

Why Litigators Should Care

The value of a communication often has little to do with the medium through which it occurred. A damaging admission does not become less relevant because it was made in a text message rather than an email. A contradictory statement does not become less important because it was posted on social media rather than included in a letter. The legal system generally focuses on the substance of the communication rather than the technology used to create it.

Viewed through that lens, AI conversations become extremely interesting.

Imagine a plaintiff who spends months discussing a dispute with ChatGPT before filing suit. During those conversations, the plaintiff experiments with different versions of events, asks the AI how a jury might react to particular facts, and seeks feedback regarding the strengths and weaknesses of various arguments.

Now imagine that the version of events ultimately presented during litigation differs from what was originally discussed with the AI system.

Most litigators immediately recognize why those conversations might matter.

The same principle applies in criminal matters, employment disputes, family law cases, business litigation, and virtually every other practice area. AI conversations may contain admissions, alternative explanations, contemporaneous impressions, damage calculations, witness evaluations, strategic thinking, and countless other categories of information that litigators routinely seek through discovery.

In many respects, AI conversations are not fundamentally different from emails, text messages, diary entries, or social media posts. They are simply another repository of potentially relevant information.

Several law firms have already warned clients that prompts entered into public AI systems may become discoverable in future litigation and should not be assumed to enjoy privilege protections.³ ⁴ ⁵

The fact that the communication occurred with a machine rather than a human being does not automatically make it immune from discovery.

Privacy and Privilege Are Not the Same Thing

One of the recurring themes in commentary surrounding AI and discovery is the tendency to confuse privacy with privilege.

Many users assume that because a conversation is not publicly visible, it must somehow be legally protected. That has never been true.

Something can be private and still be discoverable. Something can be confidential and still be subject to subpoena. Something can be stored behind layers of security and still be ordered produced by a court.

The critical legal question is not whether a communication is visible to the public. The question is whether a legal doctrine exists that prevents compelled disclosure.

For many AI conversations, that answer remains uncertain.

Several commentators have described this as the “illusion of privacy” surrounding AI systems. Users may feel as though they are engaging in a private consultation when in reality they may be creating a discoverable record of their thoughts, statements, and factual narratives.⁶

Lawyers Face a Different Problem

The analysis becomes considerably more complicated when the user is a lawyer rather than a client.

We operate under a web of ethical, professional, and legal obligations that do not apply to ordinary users. Confidentiality obligations exist regardless of whether attorney-client privilege ultimately applies. Competence requirements demand that attorneys understand the technology they employ. Professional responsibility rules require lawyers to take reasonable steps to safeguard client information.

These obligations create challenges that extend beyond traditional privilege analysis.

Consider a lawyer who uploads confidential client documents into a public AI platform and asks the system to identify weaknesses in the opposing party’s case. The resulting analysis may be useful. It may save substantial time. It may even improve the quality of representation.

But several important questions immediately arise.

Where was that information transmitted? How long will it be retained? Who has access to it? Is it used to improve future models? What contractual protections govern the relationship between the lawyer and the AI provider? These questions rarely arose when lawyers performed legal research, drafted memoranda, or analyzed evidence entirely within their own offices. Generative AI has changed that reality by introducing a potentially significant third-party participant into legal workflows.

For perhaps the first time in legal history, large numbers of lawyers are voluntarily transmitting portions of their thought processes, legal strategy, work product, and client information to technology providers operating entirely outside the law firm. That fact alone should prompt careful consideration before continuing to do so.

The American Bar Association’s Formal Opinion 512 specifically warns lawyers that the use of generative AI tools implicates duties of competence, confidentiality, communication, and supervision.⁷ Similarly, legal commentators have increasingly focused on whether disclosure of client information to AI providers could create privilege, waiver, or confidentiality concerns depending on the facts and circumstances involved.⁸ ⁹ ¹⁰

Importantly, none of these authorities suggest that lawyers should avoid AI altogether. The consistent message is that lawyers must understand the technology they are using and evaluate the risks associated with it.

Why the Third-Party Provider Matters

When lawyers discuss AI risks, they often focus on the capabilities of the models themselves. The more important issue may be the existence of the provider behind the model.

Historically, a lawyer’s notes remained in the lawyer’s file. A draft brief remained on firm servers. A legal strategy memorandum remained within the firm’s control. Public AI systems alter that equation.

Now there may be another entity involved. Whether that entity is OpenAI, Anthropic, Google, Microsoft, or another provider, the existence of an outside custodian changes the analysis. Discovery disputes frequently focus on who possesses information, who controls it, and whether information was disclosed to third parties.

As lawyers increasingly integrate AI into their daily practice, those questions will become more significant rather than less. The challenge is not merely technological. It is legal.

The Ollama Question

This is where the conversation becomes particularly interesting.

Ollama is a free tool you can download today. Once downloaded, you can then download, again free, a host of quantized (meaning reduced size) versions of both open source models (e.g. DeepSeek, Mistral, Qwen) and closed source models from OpenAI for example. Once downloaded, you can interact with them on your laptop as you do with Google Gemini for example. The difference? The model is on your laptop, the data stays on your laptop. You can operate Ollama and downloaded models disconnected from the Internet.

What this means is the following:

No prompt is transmitted to OpenAI.

No document is uploaded to Anthropic.

No conversation is processed by Google.

The legal analysis begins to resemble traditional software rather than a cloud-based service. If all processing occurs locally, a lawyer can credibly argue that client information never left firm control and was never disclosed to an outside AI provider.

Researchers studying on-premises and locally hosted AI systems have noted that one of the primary advantages of local deployment is increased control over sensitive information and reduced exposure to third-party custodians.¹¹ ¹² ¹³ Use of Ollama and similar technologies does not automatically solve every problem, but it unquestionably changes the conversation. The Discovery Issue Does Not Disappear

At this point, some commentators make a mistake. They conclude that if information never leaves the lawyer’s computer, discovery concerns disappear. That conclusion goes too far.

Recent forensic research examining local AI platforms such as Ollama, LM Studio, and llama.cpp demonstrates that many systems create artifacts that may persist on the device. Depending on the software and configuration, those artifacts can include prompt histories, chat logs, session information, generated documents, cached outputs, and other records of AI activity.¹⁴ In other words, evidence may still exist. The difference is that the evidence resides in a different location.

Instead of existing on servers controlled by OpenAI or Anthropic, it may exist on the lawyer’s workstation, firm network, or document management system. That distinction matters greatly from a confidentiality perspective. It may matter less from a discoverability perspective.

Discovery has never depended solely on the existence of a third-party provider. Courts routinely compel production of information residing entirely within a party’s own systems. The relevant question is not whether OpenAI possesses the information.

The relevant question is whether the information exists and whether it is subject to discovery.

The Question You Should Actually Be Asking

As AI adoption accelerates throughout the legal profession, you should begin reframing how you think about these issues. The most important question is not whether ChatGPT can be subpoenaed. The most important question is not whether Claude stores prompts. The most important question is not whether Gemini retains conversation histories. Those questions may matter, but they are secondary.

The primary question is much simpler:

Where does the information reside, who controls it, and what legal protections apply to it?

That framework works equally well whether the lawyer is using ChatGPT Enterprise, Claude, Gemini, Ollama, or some future system that has not yet been invented. The technology will change. The legal analysis may evolve. But the underlying discovery questions remain remarkably consistent.

• What information exists?

• Who possesses it?

• Can it be preserved?

• Can it be produced?

• Can it be compelled?

The lawyers who begin asking those questions now will be better positioned than those who wait for courts to answer them later.

Sources and Additional Reading

1. Harvard Law Review Blog, United States v. Heppner (2026).

2. Jones Walker LLP, Your AI Conversations Are Not Privileged: What a New SDNY Ruling Means for Everyone (2026).

3. Baker Donelson, Your AI Prompts May Be Discoverable: What Every Client Must Know.

4. Gentry Locke, Can You Use ChatGPT to Talk About Your Legal Case?

5. Jones Walker LLP, AI Law Blog.

6. WSHB, The Illusion of Privacy: How AI Conversations Are Discoverable in Criminal and Civil Investigations.

7. ABA Formal Opinion 512 (2024).

8. ABA Jurimetrics, Exploring the Intersections of Privacy and Generative AI: Attorney-Client Privilege and ChatGPT.

9. White & Case, Attorney-Client Privilege and Work Product in the Age of Generative AI.

10. K&L Gates, Generative AI Data, Attorney-Client Privilege and the Work Product Doctrine.

11. OnPrem.LLM: A Privacy-Conscious Document Intelligence Toolkit.

12. Position: On-Premises LLM Deployment Demands a Middle Path.

13. Articles discussing local LLM deployment and data sovereignty.

14. Forensic Implications of Localized AI: Artifact Analysis of Ollama, LM Studio, and llama.cpp (2026).

What Do We Do?

Even this simple question, for those of us who have practiced law, is not merely difficult to answer, it’s silly. First question when you meet someone who says “I am a doctor” is often, “what kind?” The AI nerds (yes, that is me too, but to be fair, I was a lawyer first), have never practiced law (or medicine likely) and think that a lawyer is a lawyer. Right away, their philosophical premise - busted. If lawyer meant one thing, the odds that an LLM could eliminate that function or dramatically reduce its necessity is much higher. But, for them (and the techies that started legal AI companies like Harvey and Legora and 95% of the other ones) they are not lawyers. That fact is evident when a company like Anthropic releases a “legal skills” package and so many posts (just so many!) claim, “welp, law is cooked.”

Anthropic actually knows all the things lawyers do. I just entered this prompt into Claude, its chatbot tool: “List all the different things that lawyers do across all the various disciplines of law you can conjure.” I watched it create a list like a vertical stock ticker, just rolling past the screen for nearly 45 seconds. Even more humorous, I clicked to “download to PDF” that huge list of hundreds, probably near a thousand, tasks that lawyers perform across a huge list of legal specialties - my Claude Desktop, for the first time ever, just locked up. The AI company whose CEO says his product will reduce the need for lawyers by 50% has a tool that when asked to list what lawyers do - literally collapses. You can grab the Google Doc here.

Safe to say, we do a lot of stuff. And, that stuff cannot be reduced to robotic “skills” files. Eliminate lawyers, I think not. But, make our lives better and reclaim some time for us in our busy schedules - sure. LLMs can definitely do that. But, as I hinted above, LLMs are not AI.

Why LLMs are not AI

No, not the corny obviousness that they are called two different things, LLM and AI, that’s not the reason. But it should start there. LLM stands for something. Large Language Model. It does not stand for “super smart machine that thinks like humans” although some think it means that. The entire concept is merely this - a representation of all words in a way that allows math to be used to predict which word in a sentence is likely to come next. Simple. It’s a next word guessing machine. And, it guesses the next word incorrectly or outright fabricates it x% of the time and always will. One of the prime reasons for that is that all the LLMs you know were largely trained on data harvested from the Internet - specifically, most often from Reddit.

The next time you read “we are the first AI Native <insert field, e.g. law firm, accounting firm, medical office>” you should think this: We are using a next word guessing machine to perform work. That in itself, no big deal. My tool, www.legalai.com relies on LLMs for a host of time saving tasks and users love it. What it does not promise or do (because LLMs cannot do it) is that it will think for you as a lawyer and make legal judgments sufficient that you, the lawyer, can just sit back in your chair, print, sign and file things. There is no world where you should do that reliant on an LLM. Not now. Not in 5 years. Never. LLMs are automation tools. Lawyers are experienced judgment appliers (is that even a phrase?). You get my point.

A machine that predicts the next word does not know why it is predicting that next word. It does not know whether it should select that next word. Ever been in a heated discussion and in your mind said “okay, don’t say that.” LLMs cannot do that. That is what we call judgment. It doesn’t “know” anything in the same way a parrot doesn’t know anything - but it sounds very convincing. Herein lies the problem.

Humans are pattern seeking creatures. (I knew that cultural anthropology degree would come in handy). This is an evolutionary adaptation that makes sense. You see a particular creature do a thing and you remember that to guide your decisions in the future. Diamond patterned snake that makes a rattle noise bit someone. They died. I will avoid that animal. If we see the output of an LLM and think “that’s a pretty good analysis of (whatever).” We begin to anthropomorphize thought, discernment and judgment into a tool that has none. It’s the parrot disguised as the prophet. Just be aware of that when dealing with LLM outputs.

My dad and brothers were excellent analyzing and repairing mechanical things when I was growing up. I never got that gene. Sure, I can learn those things after trial and error. But they, like so many people, just had a way of visualizing mechanical stuff that made their problem to resolution cycle so much faster than mine in that domain. When I call one of them up with a question about drywall, or tile or plumbing and they can give me the right answer pretty often right over the phone. Is that pure intelligence? Or is it pattern recognition? Or is it both? Likewise, when my 90-year-old mom calls saying, again, “this phone is broken” her description of what is happening leads me pretty quickly to what is really happening. Problem fixed. But, why? I can tell you why - because I recognized a pattern and applied prior knowledge.

LLMs recognize patterns of words. They are not producing knowledge or using knowledge or obtaining knowledge. But, “what about these stories of AI being told it is being decommissioned and it threatens the human operators with blackmail to prevent that?” To that I say, realize that all of the companies releasing LLMs have the capacity to and do in fact turn dials to influence inputs and outputs to those models. See this example of the folks at Google turning that dial a bit too far.

Want proof? Well, maybe you don’t want to try this. But, if someone attempted to seek an LLMs assistance in building a nuclear weapon, the LLM would politely decline to respond. Interesting. Did the LLM (the next word guessing machine) develop a brain and ethical framework? Amazing. No. It did not. Someone at the company. Someone in charge of AI safety (which is another whole article where Orwell’s name will appear more than once) decided with the approval of executives that these categories of information will not be provided. It goes beyond that.

Try asking about both sides of an intensely debated political or religious issue to any of the LLMs. Some will respond with a balance of “there is this, but on the other hand this.” Some will respond, “position X is the virtuous one and position Y is detrimental to society and you should probably abandon it.” How did that happen? Humans turning dials during the creation of these LLMs. The reports of consciousness are nothing other than the expression of the opinions of the people in charge of producing those LLMs. That is not AI. That is a tool. That is software. It’s incredibly useful software, yes. But it is software. And, guess who agrees with me? All of the companies selling access to their LLMs.

The LLM Restaurant Owners Do Not Eat Their Own Food

If it were true that the LLM companies believed that their tools were sufficiently sophisticated and accurate enough to replace lawyers, well, it’s kind of dumb for them to continue to employ, post job openings and go to court with, you know, lawyers. They should be shedding lawyers (not hiring which all of them are. Just check their Careers pages on their websites).

When it matters to them, (recent Musk v. OpenAI lawsuit, recent Anthropic versus Department of War lawsuit, recent lawsuits against Meta and Google regarding their chatbots), these companies use actual lawyers. They select what they believe are the best lawyers in the specialty applicable to their litigation. I do not think they would hire a lawyer who said, “Sure, I can take on this litigation. I will simply defer to your LLM’s legal ‘skills’ tools for all my drafting, strategy and arguments in court. We should be fine.”

Since none of the executives or people designing these LLMs would do this, we should be vigilant about not letting unsuspecting wanna-be pro se litigants from thinking they can rely on these “skills” and simply forego lawyers on matters of any importance to them. Challenging a parking ticket with some arguments an LLM provided, no high stakes. Someone suing their business partner for embezzlement - terrible idea to use an LLM. (In fact, the future of litigation is undoubtedly going to be a lot of cases in which lawyers are called to parachute into an ongoing case that the litigant thought LLM legal “skills” could help them navigate). What’s the price for that? Now we have a stressed, low leverage litigant seeking a lawyer under duress. The arbitrage problem here is obvious for unethical folks.

Experienced carpenters can safely use a nail gun. Untrained teenagers put eyes out. Same with these legal skills. Used by lawyers these are great time savers, aids in creative thinking, A/B testing arguments or strategies, or case law research queries, etc. (www.legalai.com has that btw). They are dangerous in the hands of untrained folks. We should all be vigilant about saying so. Lawyers are not next word guessing machines that reliably fabricate 20% of the time. No amount of LLM marketing should persuade anyone differently.

The Bottom Line

Many of the tasks we would rather not attend to as lawyers will be replaced by some reliance on AI. That’s good. I didn’t go to law school to become a legal search query expert, or intake expert, or interpreter of legal documents into normal English, timeline creator, video transcriber and so forth. Let LLMs, the tool, do what it is good at. Lawyers will always be important to those in the handling of the matters most important to them. LLMs don’t change that.

The December 2025 Executive Order titled “Ensuring a National Policy Framework for Artificial Intelligence” is not simply a technology policy document. It is a federal assertion of authority that directly affects how lawyers advise clients, assess regulatory risk, litigate statutory conflicts and exercise state regulatory power. While the order is framed as a response to emerging state AI laws, its consequences extend far beyond the technology sector.

For corporate counsel, it reshapes compliance planning and regulatory forecasting. For outside counsel, it changes how clients should be advised on state-law exposure and forum risk. For litigators, it creates new lines of constitutional and administrative litigation. And for state and agency lawyers, it raises fundamental questions about federalism, preemption, and enforcement authority.

Federalizing AI Governance

At its core, the Executive Order reflects a strategic federal decision to centralize AI regulation at the national level, explicitly pushing back against the growing number of state AI statutes. The White House characterizes state-level regulation as a threat to innovation and interstate commerce, but from a legal perspective, what matters is how aggressively the federal government is positioning itself to override state authority.

The order does not merely announce a preference for federal uniformity. It instructs federal agencies and the Department of Justice to affirmatively challenge state AI laws that conflict with national policy goals. That posture has immediate implications for lawyers advising clients who operate across multiple states or who have already begun compliance planning around state AI statutes.

What Corporate Counsel Need to Understand

For in-house counsel, the most immediate concern is regulatory whiplash. Many companies have already invested time and money preparing for compliance with state AI laws addressing algorithmic bias, transparency, automated decision-making, and consumer protections. This Executive Order places those investments in a state of uncertainty.

If the federal government successfully challenges or discourages state enforcement, corporate legal departments may find that carefully crafted state-specific compliance programs are no longer required—or worse, that they conflict with emerging federal standards. Conversely, if litigation over preemption drags on for years, companies may be caught between state enforcement risk and federal policy signals pointing in the opposite direction.

Corporate counsel must now advise boards and executives on:

• Whether to continue investing in compliance with state AI laws that may be preempted.

• How to document good-faith compliance efforts in a legally uncertain environment.

• How to structure AI governance programs flexible enough to survive federal rulemaking or judicial intervention.

Outside Counsel - Reframing Client Advice

For outside counsel, particularly those advising technology companies, insurers, healthcare systems, financial institutions, or employers, the Executive Order changes the substance of legal advice that can responsibly be given.

Until now, prudent advice often focused on state law exposure—especially in states moving aggressively on AI regulation. That advice must now be tempered with a discussion of federal preemption risk and the possibility that state laws may be challenged or rendered unenforceable.

Outside counsel will increasingly be asked:

• Whether compliance with a particular state AI statute is legally necessary or strategically optional.

• Whether federal litigation is likely to invalidate a state regulatory regime.

• How to draft policies, disclosures, and contracts that account for both state and federal uncertainty.

Importantly, the order also signals future federal reporting and disclosure requirements through agencies like the FTC and FCC. Counsel advising clients on AI disclosures, consumer representations, or automated decision systems must now anticipate federal standards that could override or conflict with state mandates.

This shifts outside counsel’s role from static compliance advisor to dynamic regulatory strategist.

Litigation Implications

For litigators, this Executive Order is a roadmap for future lawsuits.

The order explicitly contemplates federal challenges to state AI laws, likely invoking doctrines such as the Dormant Commerce Clause, federal preemption, and limits on state authority to regulate interstate technology platforms. That alone creates fertile ground for constitutional litigation.1

But the litigation implications do not stop there. Private parties will inevitably rely on the Executive Order as persuasive authority—if not binding law—when challenging state enforcement actions. Expect defendants to argue that state AI regulations are inconsistent with national policy and therefore invalid or unenforceable.

Litigators should anticipate:

• Declaratory judgment actions challenging state AI statutes.

• Injunction proceedings brought by the federal government or regulated entities.

• Defensive litigation where compliance with state law is argued to be federally preempted.

• Administrative law challenges to agency actions taken under the order’s directives.

State Agency Lawyers

State attorneys general, agency counsel, and regulatory lawyers face perhaps the most direct impact. The order openly challenges the legitimacy of state AI regulation and threatens to condition federal funding on compliance with federal AI policy objectives.

This raises profound legal questions for state lawyers:

• Can the federal executive branch effectively deter state lawmaking without congressional authorization?

• Does conditioning infrastructure or technology funding on AI regulatory compliance violate principles of federalism?

• How should state agencies proceed with enforcement when federal litigation is threatened?

State and local agencies that have invested resources in AI oversight frameworks may now face pressure to pause, revise, or abandon those efforts. Agency counsel will need to advise not only on legal risk, but on institutional risk, including the possibility of losing federal funding streams.

The Federal Reporting and Disclosure Regime: A New Compliance Layer

The Executive Order directs federal agencies to consider national reporting and disclosure standards for AI systems. For lawyers, this signals the likely emergence of federally enforceable disclosure obligations, potentially replacing or superseding state transparency requirements.

Corporate and regulatory lawyers alike should recognize that disclosure regimes are often the most litigated aspect of regulatory frameworks. Misstatements, omissions, or inconsistent disclosures can trigger enforcement actions, private litigation, and class actions.

If federal agencies adopt standardized AI disclosures, lawyers will need to:

• Reevaluate client disclosures across all jurisdictions.

• Harmonize internal documentation with federal definitions and terminology.

• Prepare for enforcement actions grounded in federal consumer protection or unfair practices law.

This could become the AI equivalent of federal securities disclosure or privacy notices—areas where lawyers play a central gatekeeping role.

The Funding Lever

One of the more subtle but legally significant aspects of the order is its use of federal funding as leverage against state regulation. By tying eligibility for certain federal grants to alignment with national AI policy, the executive branch introduces a compliance incentive that will inevitably be litigated.

Lawyers should immediately recognize the parallels to past funding-condition disputes in areas like education, transportation, and healthcare. Courts have long scrutinized whether federal funding conditions are coercive or exceed statutory authority.

For lawyers advising state agencies, universities, research institutions, or companies dependent on federal grants, this raises urgent questions about eligibility, compliance representations and contractual obligations.

What This Means for the Practice of Law

Stepping back, the most important takeaway is that this Executive Order turns AI regulation into a general legal issue, not a niche technology concern. It implicates constitutional law, administrative law, compliance counseling, litigation strategy, and government enforcement authority.

Lawyers across practice areas will increasingly be asked to:

• Interpret federal-state conflicts in real time.

• Advise clients under conditions of regulatory instability.

• Litigate preemption, enforcement, and disclosure disputes.

• Navigate funding-related compliance risks.

AI governance is no longer something lawyers can safely delegate to technologists or policy teams. It is now squarely within the core competencies of the legal profession.

Conclusion: A Lawyer’s Executive Order

This Executive Order is best understood not as an AI policy document, but as a lawyer’s Executive Order—one seeking to reshape who writes the rules, who enforces them and who gets sued.

Whether the federal government ultimately prevails in court or Congress enacts a comprehensive national AI statute, the legal landscape has already shifted. Lawyers who understand that shift will be better positioned to protect their clients, advise their agencies, and navigate the next phase of AI governance.

Those who treat this as “just another tech announcement” will find themselves reacting rather than leading.

For most of modern legal practice, lawyers have been trained to believe that effective legal research is a specialized skill set in its own right. We were taught how to construct Boolean searches, how to layer proximity connectors, and how to iteratively refine queries until the right cases finally surfaced. Research competence was equated with search mastery, and entire CLEs were devoted to teaching lawyers how to “search better.”

That model is now obsolete, but legacy legal search providers do not want you to hear about it.

A quiet but decisive shift has already occurred in legal research technology. The platforms lawyers have relied on for decades—including Westlaw and LexisNexis—now use a framework called Retrieval-Augmented Generation, commonly referred to as RAG. This change is not cosmetic. It fundamentally alters how legal research works, what differentiates competing tools, and why the historical advantages of legacy providers no longer function as meaningful barriers to entry.

Understanding this shift matters, because it explains not only how modern legal research tools operate, but why smaller, newer services can now deliver research capabilities that rival—or exceed—those of long-established incumbents. Oh, and the most important part of that, dramatically reduced legal search costs.

What Retrieval-Augmented Generation Actually Means

Despite its technical name, Retrieval-Augmented Generation is conceptually straightforward. A RAG system works by converting your natural language query to numbers and then searching that list of numbers against thousands of lists of numbers associated with segments of all Federal and State Law. The generation step does not occur in isolation; it is constrained by actual law. This is not the generation of fake cases - a real risk we all face especially in firms where lawyers are signing documents largely constructed by other lawyers.

This sequencing is critical. The system does not begin by inventing an answer and then searching for support. Instead, it retrieves cases, statutes, or other authorities first and then reasons from those texts. In other words, the AI is forced to “show its work” by tying its conclusions to real legal sources.

This approach differs sharply from earlier generations of legal search tools. Traditional legal research—even when described as “AI-assisted”—still required the lawyer to design the search, evaluate results manually, and synthesize conclusions through repeated iteration. The technology assisted retrieval, but it did not meaningfully assist reasoning.

RAG systems invert that workflow. The lawyer asks a question in ordinary language. The system retrieves relevant authorities.

Why Law Is Particularly Suited to RAG Systems

Legal research is uniquely compatible with retrieval-augmented AI. Law is not an open-ended creative domain. Legal conclusions must be traceable to written authority: judicial opinions, statutes, regulations, and administrative decisions. RAG systems are designed to operate within exactly those constraints.

Because the generative model is tethered to retrieved legal texts, it cannot easily drift into unsupported assertions. Its output is only as good as the sources it retrieves, which aligns closely with how lawyers already think about legal reasoning. The system mirrors what lawyers do manually: locate authority first, then analyze it.

This is not an experimental architecture. It is now the dominant approach in serious legal AI because it balances speed with accountability.

Westlaw and Lexis Are Already Using RAG

Retrieval-Augmented Generation is not merely something startups are experimenting with while legacy platforms remain rooted in traditional systems. That assumption is incorrect.

Westlaw and LexisNexis are already using RAG-based architectures in their modern AI offerings. Lexis training often includes trainers breathlessly promoting to lawyer audiences about their “RAG technology.” Lexis has publicly described Lexis+ AI as grounding its responses in retrieved legal sources. Thomson Reuters has made similar representations regarding Westlaw AI. These systems retrieve relevant authorities using RAG and they are never returning to the legacy methods upon which their subscriber base was built.

They have to stay with RAG. Without retrieval grounding, generative systems hallucinate. No responsible legal research provider could deploy a purely generative model without anchoring it to real law. The adoption of RAG by Westlaw and Lexis is therefore not a competitive choice—it is a technical necessity.

The significance of this cannot be overstated. Now that the leading incumbents have adopted an architecture that is broadly accessible, their technological exclusivity disappears.

Why Westlaw Acquired Casetext

Thomson Reuters’ acquisition of Casetext in 2023 was widely reported, but its deeper significance is often misunderstood. Westlaw did not acquire Casetext because it lacked data. Westlaw already possessed one of the largest and most curated legal databases in the world.

The acquisition was about capability, not content.

Casetext had already built and deployed a production-ready RAG system through its CoCounsel platform. It had demonstrated how large language models could be safely integrated into legal research workflows, how retrieval grounding could prevent hallucinations, and how AI could move beyond keyword search toward substantive analysis.

By acquiring Casetext, Westlaw accelerated its transition into this new paradigm rather than building everything from scratch. The message was implicit but unmistakable: traditional editorial processes alone were no longer sufficient to remain competitive.

The Collapse of the Editorial Moat

For decades, legacy legal research providers relied on human editorial labor as their primary competitive moat. Headnotes, classification systems, and manually curated summaries were expensive to produce and difficult to replicate. That difficulty justified high prices and long-term contracts. The explosion of inexpensive access to LLMs ended all of that.

Modern AI systems can analyze the text of judicial opinions directly. They can identify legal standards, compare reasoning across cases, and surface doctrinal conflicts without relying on pre-assigned categories or human-written summaries. While editorial enhancements may still add value, they are no longer prerequisites for powerful legal research.

Why New Entrants Can Now Compete

Once legal research became a computational problem rather than a labor-intensive editorial one, the barriers to entry fell dramatically. Today, a company seeking to build a modern legal research platform needs access to legal texts, modern embedding models for semantic retrieval, large language models, and legal-specific safeguards. None of those inputs are exclusive to billion-dollar incumbents.

Even more importantly, the cost of these components continues to decline. Large language models are becoming cheaper, faster, and more capable at a pace that few anticipated even two years ago. What once required massive infrastructure investments can now be accomplished by relatively small, focused teams.

This is why the legal research market is beginning to fragment after decades of consolidation. The technological advantages that once insulated incumbents have eroded. This raises the obvious question - How long will lawyers continue to pay the 90% higher monthly cost to legacy providers?

The Real Inefficiency Was Always the Workflow

The greatest inefficiency in legal research was never access to information. Lawyers have always had access to more cases than they could reasonably read. The inefficiency lay in forcing lawyers to manually construct searches and anticipate how judges might phrase legal issues. None of us went to law school to become search query experts and in our practices we did not want to become one.

RAG systems eliminate that burden. They allow lawyers to ask substantive questions directly and let the system handle the mechanical task of searching across linguistic variations, doctrinal framings, and jurisdictional differences. The system does not just search faster; it searches more broadly and more thoroughly than any individual lawyer can.

This shift reveals an uncomfortable truth: query craftsmanship was a workaround for technological limitations, not a core legal skill.

LegalAI.com and the End of Query Construction

LegalAI.com1 embraces the logical conclusion of this shift. Rather than offering a marginally improved search interface, it eliminates the premise that lawyers must write search queries at all. Instead of legal search requiring a lawyer staring at a screen conjuring natural language queries, legalai.com rethinks this entirely. The system instead asks the user to do what all humans can do, enter the narrative regarding what the client’s case is about. The lawyer-built AI expands that narrative into numerous legal and factual search queries. The user can select which to run, edit those that are selected and those searches are run simultaneously - completing in less than 60 seconds.

By automating query expansion, LegalAI explores legal terrain that a human researcher would never fully cover. It surfaces cases that fall outside obvious keyword formulations and reduces the risk of missing relevant authority simply because it was phrased differently. You are no longer needed as a search query expert.

The result is not merely faster research. It is deeper and more comprehensive research.

What This Means for Lawyers

The rise of RAG does not diminish the importance of legal judgment. It enhances it. By removing the mechanical burden of searching and synthesizing, modern tools allow us to focus on strategy, reasoning, and advocacy.

The most effective legal research platforms going forward will retrieve comprehensively, explain clearly, and remain transparent about their sources. They will support lawyers’ thinking rather than forcing lawyers to work around the tool.

For the first time in decades, powerful legal research is no longer confined to a handful of legacy platforms. That is not just a technological shift—it is a structural change in the profession.

Final Thought

Retrieval-Augmented Generation did not simply improve legal research. It dismantled the assumptions on which the legal research industry was built.

The future of legal research will not belong to those with the largest editorial operations. It will belong to those who build the best tools for legal reasoning.

And that future has already begun.

Artificial intelligence no longer merely answers questions. It acts. In early 2025, technology writers began describing the emergence of “agentic AI” systems—self‑governing bots that can carry out tasks with limited human oversight. These agents are different from traditional generative models; they can book flights, negotiate contracts, or draft legal …

Read more

The US approach to artificial intelligence regulation is a case study in what happens when regulation and consumer or citizen protection clashes with a legitimate desire for innovation. It also is an interesting experiment in what happens when federal leadership recedes. Some might find it refreshing in a state’s-rights sort of way that the federal government has yet to pre-empt state AI laws. Lawyers, corporations and others might find it difficult to navigate.

When President Trump revoked the Biden administration’s executive order on safe and trustworthy AI and issued his own directive calling for the removal of “barriers to American AI innovation,” the result was not a uniform easing of obligations but something closer to chaos. The new order gives no guidance on how agencies should police AI and directs a review of all actions taken under the revoked order. In this vacuum, states and cities have become laboratories of AI policy. From one perspective, that is useful - find out what laws work and what laws unacceptably restrain innovation. Lawmakers in California and Colorado have adopted training‑data transparency laws requiring developers to disclose whether their datasets contain copyrighted or personal information. Others, such as New York City, have passed laws requiring audits of algorithms used in hiring, while Illinois has long regulated biometric data collection. What emerges is a mosaic of obligations that vary by geography, industry and even business model. For the legal profession, understanding this mosaic is no longer optional.

A Brief History of the Vacuum

To appreciate how this patchwork developed, it helps to recall that only a year ago the federal government promised sweeping oversight. President Biden’s Executive Order 14110 sought to create a cohesive framework for AI safety, data privacy and civil‑rights protections. The repeal of that order signaled a philosophical shift and emboldened states to assert control. According to ArentFox Schiff’s 2025 “AI Legal Landscape” report, the lack of federal action has spurred “more assertive policymaking at the state and local levels, with a focus on safety, bias, privacy, and sensitive use cases.” A quick review of state AI regulation so far reveals a focus on the rights of consumers, users and individuals. A federal framework might focus on regulation that maximizes innovation which these types of employment and privacy focused regulations could well hold back.

Read more

Read more

This is a conversation space exclusively for subscribers—kind of like a group chat or live hangout. I’ll post questions and updates that come my way, and you can jump into the discussion.

Join chat

Read more

I continue to work in this realm without the need of or desire to involve the ultimate product in advertising. As many of you would agree, the alignment with any form of advertising for an information source like LegalAI will inevitably result in biased reporting, writing and opinions whether unconscious or not.

As an example, I have had many discussion behind the scenes about producing a website containing some or all of this content and enabling advertising there. However, no one could persuasively argue that even that duplicate sourcing of most of this content would not, in some way, influence the content itself. Therefore, any plans for such a website related source of this content were discarded.

I believe in the substack model for many reasons, chief among them is the ability of readers to support the content they find valuable and, thereby, determine value outside of the influence of advertising.

LegalAI began in March 2023 with 0 subscribers and myself simply authoring content designed to assist lawyers in understanding the impact of AI on their practice and the legal field generally. It has stayed true to that mission even as I have surpassed 100 subscribers at this point without any advertising other than the ability for readers to find us on the substack application itself. Because of the intentional choice of a niche product, the intersection of the law and AI directed at lawyers as readers, I knew our substack will serve a need, but a narrow one. That is fine with me. That was always the goal.

I intend to continue with that same philosophy of providing LegalAI related content for the increasingly large set of lawyers recognizing the value of this information to their practice. Past posts have touched on areas of the law as diverse as criminal, civil rights, employment, securities, the rules of authentication, revenge porn, entertainment law and others.

I, of course, would appreciate you sending a link to this publication to any of your lawyer contacts whom you think would benefit from the content as you do.

To that end, starting with my next issue this month, August 2025, I will be embarking on subscriber benefits for paid subscribers to enhance our content even further. Future posts will be available in their entirety to paid subscribers with the initial two paragraphs available to the public to sample post topics.

My first product for paid subscribers will be the availability of a podcast style version of each post enabling you to listen to that post while training, driving, walking or anywhere you now listen to formerly written content. Many news outlets are currently offering this service to their paid subscribers and receiving great feedback on its value.

I will also be offering unbiased reviews of AI tools targeted at the legal market to assist you in your practice. It is my belief that lawyers who remain unaware of the value of AI to their practice will continue to have difficulty competing in the market. But, at the same time, my philosophy is also that simply adopting every new AI tool in the legal marketplace is also not the best path forward. Knowledge and discernment about which AI tools will blend with your existing practice workflow will be the best approach. The paid subscriber benefit here will keep you well informed about the latest tools and how they may enhance your practice, save you time and improve your life outside the office as well!

My reviews of such products will include key questions to ask vendors that relate to your duty of confidentiality to your client, the accuracy of information provided and more.

Coming in fall 2025 will be a regular monthly scheduled chat event on substack containing a theme in the LegalAI world. That offering will enable lawyers focused on a particular area of concern to gather and discuss questions, suggestions and enhance their ability to manage the full range of issues AI poses to lawyers in their everyday practice.

I sincerely appreciate the time you have all taken out of your busy lives to consume the content I produce here at LegalAI. I will work diligently to continue to earn your trust as a source of LegalAI information borne of the unique background in both being immersed in technology law for decades and designing AI Systems as well being experienced in AI Governance generally.

The future of the intersection of the law and AI is both bright with possibility and filled with new issues for lawyers to identify. This substack is a contribution to your ability to continue to perform well in the AI era. Thank you.

Denver U.S. District Judge Nina Y. Wang, issued the sanctions on July 7, 2025, citing the lawyers’ “gross carelessness” and failure to meet basic professional duties. This high-profile blunder underscores a growing crisis: attorneys using generative AI tools without understanding their limitations or verifying their outputs.

The case in question was a high-profile defamation lawsuit brought by Eric Coomer – a former Dominion Voting Systems employee – who accused MyPillow founder Mike Lindell of spreading baseless claims that Coomer helped rig the 2020 presidential election. In June, a Colorado jury found Lindell liable for defamation and awarded Coomer roughly $2 million in damages (far less than the $62.7 million originally sought). It was during the lawsuit that Lindell’s legal team submitted an error-strewn opposition brief, unwittingly turning themselves into a cautionary tale about the perils of unverified AI in legal practice.

The AI Powered Errors

The trouble began after Lindell’s attorneys – Christopher Kachouroff and Jennifer DeMaster – filed a February 25, 2025 opposition motion in the case. On its face, the brief looked like a routine legal filing. But at a pretrial conference on April 21, Judge Wang revealed a “troubling pattern” in the document: case citations and quotes that simply didn’t add up. One by one, the judge pressed Kachouroff about anomalies in the brief. Would he be surprised, she asked, to learn that a certain cited case “did not exist?” Kachouroff admitted he “would be surprised.” When confronted with a misquote from a real case, he conceded, “Your Honor I may have made a mistake…I wasn’t intending to mislead the Court,” even insisting the misquoted text was “not far off” from the real thing.

At first, Kachouroff fumbled to explain the situation. He suggested the document on file was just a draft that had been filed “by mistake,” hinting that another attorney (his co-counsel DeMaster) had been tasked with checking the citations. But Judge Wang’s scrutiny only intensified. According to her written order, “time and time again, when Mr. Kachouroff was asked for an explanation of why citations to legal authorities were inaccurate, he declined to offer any explanation, or suggested that it was a ‘draft’ pleading.” Finally, after repeated questioning, the truth came out: Kachouroff admitted the brief had been drafted with the help of AI.

In a dramatic exchange, Judge Wang asked whether the filing had been run through an AI program – a question Kachouroff initially did not expect. “Not initially,” he answered. “Initially, I did an outline for myself, and I drafted a motion, and then we ran it through AI.” Pressed on whether he double-checked the AI’s citations, Kachouroff admitted: “Your Honor, I personally did not check it. I am responsible for it not being checked.”  The courtroom transcript makes clear that only when directly confronted did the lawyer acknowledge using the AI tool – and acknowledge that he failed to verify the accuracy of the citations. Of course, missing in all this back and forth was precisely which “AI” was used. Knowing which tool was used would be helpful for other practitioners and judges.

Hallucinated Cases, Misquotes, and “Draft” Excuses

In Judge Wang’s order, she wrote that the brief contained “misquotes of cited cases; misrepresentations of principles of law…including discussions of legal principles that simply do not appear within such decisions… and most egregiously, citations of cases that do not exist.”  Nearly 30 citations in the document were defective in one way or another . For example, one citation in the brief attributed a ruling to the 10th Circuit Court of Appeals that was entirely fabricated – the case “Perkins v. Fed. Fruit & Produce Co.” was listed with a 2019 citation that “did not exist as an actual case,” as the judge pointed out. (There was a different Perkins case from years earlier in a lower court, but it “does not stand for the proposition asserted by Defendants” in the AI-written brief ). In other instances, real cases were cited but quoted incorrectly or out of context, to support legal points those cases simply never addressed.

Kachouroff struggled to explain how his opposition brief became a patchwork of fiction. He claimed the garbled filing was an old draft that somehow got filed instead of the final, corrected version. In their formal response to the court, the attorneys maintained “it was inadvertent, an erroneous filing that was not done intentionally, and was filed mistakenly through human error.” They insisted that they had reviewed an AI-generated draft and fixed its mistakes, only for the earlier uncorrected draft to be submitted by accident.

Judge Wang was unpersuaded. After the April 21 hearing, she ordered the defense team to produce all versions of the brief, along with metadata and correspondence, to verify their claims. The subsequent investigation only heightened her skepticism. The supposedly “corrected” version that Kachouroff belatedly offered to file turned out to contain “several of the same substantive errors” as the original, and metadata showed it was edited after the court had flagged the issue. Emails between Kachouroff and DeMaster revealed that their drafts already contained the fake citations before the brief was ever filed. The judge found the lawyers’ explanations “contradictory” and lacking any corroboration, leading her to conclude the fiasco was not an inadvertent one-off at all. In fact, Judge Wang noted these attorneys had filed “similarly defective” court documents in at least one other case, suggesting a pattern of carelessness (or over-reliance on AI) beyond this incident.

At one point Kachouroff accused the court of trying to “blindside” him with the citation errors – a stance Judge Wang called “troubling and not well-taken.” “Neither Mr. Kachouroff nor Ms. DeMaster provided the Court any explanation as to how those citations appeared in any draft of the Opposition absent the use of generative artificial intelligence or gross carelessness,” Wang wrote, pointedly rejecting the lawyers explanations. In other words, either they used an AI tool that hallucinated fake cases, or they demonstrated “gross carelessness” – there was no benign excuse that she felt fit the facts.

Violations of Professional Duties: Rule 11 and Beyond

Beyond the embarrassment of citing fictional cases, the incident raised serious questions of professional responsibility. By signing and submitting the faulty brief, Kachouroff certified that, to the best of his knowledge, the legal arguments were grounded in valid law after reasonable inquiry – as required under Rule 11 of the Federal Rules of Civil Procedure. Yet he admitted in open court that he “failed to cite check the authority…after [AI] use before filing it with the Court – despite understanding his obligations under Rule 11.” That admission put him in clear breach of his duty of candor and diligence. Judge Wang’s order explicitly reminded Lindell’s counsel that they are “bound by the standards of professional conduct” set by the Colorado Rules of Professional Conduct, the code of ethics for attorneys. Those standards include the duty of competence (which now in many jurisdictions encompasses a duty to understand the benefits and risks of relevant technology) and the duty of candor toward the tribunal – duties that were arguably violated when an unvetted AI-generated brief was passed off as legal analysis.

Initially, Judge Wang not only considered monetary sanctions but also raised the specter of professional discipline. In late April, she issued an order for Kachouroff and DeMaster to show cause why they should not be referred to disciplinary authorities for potential violations of ethics rules. The lawyers scrambled to respond by the May 5 deadline, filing lengthy declarations defending their conduct. They apologized for the chaos – though the judge noted Kachouroff’s written “apology” came off as “a bit hostile” in tone. They insisted they had no intent to mislead and had been “wholly unprepared” when the judge sprang the issue on them at the hearing. Crucially, they argued that using AI itself was not wrongdoing: “There is nothing wrong with using AI when used properly,” they wrote, claiming they had no reason to think an unverified draft had been filed until it was too late.

In the end, Judge Wang stopped short of referring the matter for formal discipline, opting instead for the $6,000 in total fines as a measured punishment. Notably, she did not sanction Lindell himself or his company, since there was no evidence the client knew his lawyers were employing AI in their drafting. But the judge made clear that the primary fault lay in the lawyers’ laps: their failure to verify what they were filing and their lack of forthrightness when problems surfaced. The message was unmistakable – every attorney has a non-delegable duty to ensure their filings are accurate and grounded in real law, no matter if an associate or an algorithm helped prepare them.

A Wider Trend: Warnings from the Bench and Beyond

This incident is part of a growing pattern of lawyers learning the hard way that AI-generated legal work can go disastrously wrong if left unchecked. I have written about this many times here and it keeps happening. “More and more lawyers are getting caught – and punished – for including AI hallucinations in their work,” tech outlet The Verge observed, “and this trend will likely only continue to grow.”   In fact, just a year earlier, a pair of attorneys in New York made headlines for submitting a ChatGPT-written brief that cited a string of fake court decisions, prompting a judge’s ire and sanctions. In that 2023 episode, the AI tool had confidently fabricated six case citations complete with bogus quotes, leading the judge to lament that “technological advances are commonplace and of great benefit, but [lawyers] are responsible for ensuring the accuracy of their filings.” Courts across the country have since been “really cracking down on lawyers who submit pleadings with AI-hallucinated legal authority,” as one legal ethics commentator noted.

Legal technology experts say the Lindell case highlights a “critical need for the legal profession to establish AI competency standards” and protocols for law firms.  Put simply, modern language models like OpenAI’s GPT-4 or other generative AI systems have a well-known tendency to “hallucinate” – to fabricate plausible-sounding information (such as fake case law) in response to a prompt. These AI hallucinations are not rare glitches; they are an inherent limitation of how the technology predicts text without any grounding in truth. That means lawyers cannot treat an AI’s outputs as trustworthy legal research. As Colorado attorney and legal blogger Robin Shea quipped after reviewing the Lindell debacle, “If you’re doing something more important than writing a thank-you note…like submitting a brief on behalf of your client…you need to check behind the AI.” Lawyers “should know this,” she added, “but because they get burned for it all the time, apparently they don’t.” 

Even before this latest sanction, bar associations and ethics panels have been urging caution. The American Bar Association recently issued Formal Opinion 512, stressing that using generative AI tools is permissible only if attorneys ensure accuracy, confidentiality, and adhere to all ethical duties – ultimately, the lawyer remains the responsible party, not the AI. Some state bars have gone further, requiring lawyers to stay educated about technology as part of their competence obligation. The Lindell case has now put a very public spotlight on these issues. It serves as a “wake-up call for the entire legal profession”, as USA Herald’s Samuel Lopez wrote, warning that “as AI tools become increasingly accessible, the temptation to use them without proper safeguards will only grow.” And the consequences are real: “Client representation suffers, court resources are wasted, and public confidence in legal institutions erodes when attorneys fail to maintain basic professional standards,” Lopez noted. In short, the integrity of the justice system takes a hit when lawyers abdicate their quality-control responsibilities to a machine.

Lessons and Tips for Lawyers Using AI

For lawyers, the professional fallout from relying on unvetted AI should serve as a warning. “Always check your citations,” legal columnist Bob Ambrogi wrote – only half-jokingly adding, “and always keep your pants on in court.”  (Kachouroff, it turns out, had previously made news for appearing pantless during a Zoom court hearing – a separate embarrassment that, while unrelated, underscores the scrutiny attorneys can face). The silver lining of such fiascos is that they yield clear lessons on how to use AI tools responsibly in legal practice:

• AI is a tool, not a substitute for your own professional judgment. Treat anything an AI produces as a draft – a starting point to be carefully reviewed and edited by a human lawyer. No matter how sophisticated the software, it cannot be trusted to know the law or the facts with reliability. The Lindell case underscores that generative AI will confidently output text that looks convincing but may be completely false.

• Never delegate verification to AI (or to junior staff without oversight). In this incident, Kachouroff attempted to delegate the cite-checking to his colleague, and both trusted an AI-generated draft without double-checking every reference. This lack of oversight was a key failure. Critical tasks like citation and fact verification must be performed by a human who understands what to look for – be it the lead attorney or someone under their close supervision. Blaming an associate or an algorithm after the fact won’t absolve the lawyer who signs the filing. As the court put it, delegation without appropriate oversight is essentially a dereliction of an attorney’s duty.

• Understand AI’s limitations – and use safeguards when prompting. If you do choose to use AI in drafting, be specific in your prompts to mitigate hallucinations, and still assume the output may contain errors. For example, instruct the AI not to fabricate case law and to flag any uncertain information. A prompt might say: “Include accurate, verbatim quotations from cited cases and do not invent any citations. Mark any statements that need verification.”   Such instructions can help, but they are not foolproof – nothing replaces a manual check. After generating text, compare every quote to the source, confirm every case is real and cited correctly, and ensure the AI hasn’t put words in a court’s mouth. In practice, this means using AI for efficiency (say, to improve the writing or organization) while you remain the fact-checker and researcher before anything goes out the door.

• Get training and stay educated. The mishap also highlights a broader “competency gap.” Attorneys should seek out training on how generative AI works and how to use it ethically. This might involve CLE (continuing legal education) courses on legal technology or firm-wide protocols for AI usage. The legal profession is even considering formal AI competency certifications to ensure lawyers understand tools’ capabilities and pitfalls. As one commentator noted, had Lindell’s lawyers “understood the core principles of prompt engineering and AI limitations,” they might have been able to harness the tool’s benefits without falling into its traps. Every lawyer doesn’t need to be a programmer, but basic technological competence is quickly becoming part of the duty of competence in law.

In sum, AI is here to stay in the legal profession. From research to drafting to data analysis it is too valuable a tool for lawyers to abandon. By treating AI output with skepticism, verifying everything, and never relinquishing the your role as final editor and fact-checker, you can safely tap into AI’s advantages (speed, efficiency, linguistic polish) without undermining your integrity or your clients’ cases. As Judge Wang’s ruling makes clear, no clever software will save an attorney who forgets their fundamental obligations. The onus is on us to ensure that “garbage in, garbage out” doesn’t make it into official filings. And if an AI tool is used, you must be willing to stand behind the result – because ultimately, the judge and client will hold you accountable, not the algorithm.

Helpful Sources

1. Olivia Prentzel, The Colorado Sun – “MyPillow CEO’s lawyers fined for AI‑generated court filing in Denver defamation case” (Jul 7, 2025)

2. Economic Times (India) – “Mike Lindell’s lawyers fined $3,000 for using AI that created fake court citations” (Jul 9, 2025)

3. Ravi Hari, Livemint/Today News (Bloomberg) – “Denver judge fines MyPillow CEO Mike Lindell’s lawyers over AI‑generated fake citations: ‘This court derives no joy…’” (Jul 8, 2025)

We lawyers must grapple with the application of existing laws to AI-driven marketing, from telemarketing regulations to advertising and intellectual property rights. This comprehensive guide examines the key legal considerations of AI voice cloning in marketing – focusing on U.S. federal and state law – providing insights for corporate counsel, advertising lawyers and litigators alike.

The Rise of AI Voice Cloning in Marketing

AI-generated voices are becoming remarkably sophisticated. Companies like ElevenLabs have popularized the indistinguishably human sounding AI voices for a variety of applications. Marketers for companies large and small leverage these tools for various purposes:

• Automated Outreach Calls: Placing cold calls using synthetic speech, delivering personalized marketing pitches at scale. A cloned voice could be used to resemble a friendly agent – or even a celebrity – to engage customers.

• Voice Cloning for Branding: Companies can replicate the voice of a notable individual (such as a spokesperson, executive, or celebrity endorser) to maintain consistent branding across thousands of calls or ads. For example, a top salesperson’s voice might be cloned to leave ringless voicemails for leads.

• Conversational AI Chatbots: Advanced bots can conduct interactive text or voice conversations with consumers, adapting their tone and responses on the fly. These AI “agents” blur the line between recorded messages and live interaction.

The result is a powerful marketing arsenal that can reach audiences faster and more persuasively than traditional methods. Of course, the most expensive part of any business, human labor, is reduced dramatically. However, each of these tools comes with legal ambiguity and risk. Laws that were written before the AI era are now being applied to these novel scenarios, and regulators and courts are starting to respond. If you are advising on marketing practices you need ensure that embracing innovation does not mean skirting compliance.

Telemarketing Laws and AI: The Telephone Consumer Protection Act (TCPA)

One of the most important laws for AI-driven marketing is the federal Telephone Consumer Protection Act (TCPA). Enacted in 1991 to curb abusive telemarketing (like robocalls and junk faxes), the TCPA has reemerged “with teeth” in the age of AI. It imposes strict limits on calls and texts made for marketing purposes, and non-compliance can lead to heavy penalties – including statutory damages of $500 to $1,500 per violation, which can add up quickly in class actions.

TCPA Basics: Under the TCPA, it is generally unlawful to make a call or send a text for marketing purposes using an “automatic telephone dialing system” (autodialer) or an “artificial or prerecorded voice” without the recipient’s prior express written consent. This applies most notably to calls made to mobile phones. In plain terms, if you plan to auto-dial consumers or play them a recorded (or AI-generated) voice message, you must obtain clear, written consent in advance. Telemarketers have known this for decades with the most notable reminders of this being expensive class action lawsuits over violations.

AI Voices as “Artificial” Calls

How does a cloned or AI-generated voice fit into this legal framework? The FCC and courts have indicated that AI-synthesized voices are considered “artificial or prerecorded” voices under the TCPA. That conclusion makes sense. Even if an AI voice sounds natural and converses interactively, it is not a live human – so it triggers the TCPA’s consent requirement. In March 2024, the Federal Communications Commission issued a declaratory ruling underscoring that voice cloning technology “falls squarely within the TCPA’s scope.” In other words, an AI voice call is treated the same as a robocall delivering a recorded message. Companies cannot evade the law by arguing that an AI simulation of a voice is something new or unregulated.

Autodialers and Evolving Definitions

The TCPA’s definition of an autodialer has been a hot topic in recent years. The U.S. Supreme Court’s Facebook v. Duguid decision in 2021 narrowed the definition of an autodialer, requiring use of a random or sequential number generator to fall under the TCPA’s autodialer ban. This was seen as a victory for marketers using modern dialing systems that dial from customer lists. However, plaintiffs’ lawyers are now pushing back by focusing on the “artificial voice” aspect of calls instead. Even if a call system isn’t an autodialer by definition, using an AI-generated voice means the call could still violate the TCPA without consent. For example, a predictive dialing platform that delivers AI-generated voice messages might not meet the strict autodialer definition post-Facebook v. Duguid, but it does involve an artificial/prerecorded voice – which independently triggers TCPA coverage. In short, any marketing call featuring a synthetic or cloned voice requires consent, regardless of how the call is dialed.

Recent Developments – McLaughlin v. McKesson

A recent Supreme Court decision, McLaughlin Chiropractic v. McKesson, adds another twist for TCPA litigation. In that 2024 decision, the Court held that federal courts are not strictly bound by the FCC’s interpretations of the TCPA and must interpret the statute independently (giving agency views “appropriate respect”). This could open the door for defense arguments that a given technology falls outside the TCPA’s language, even if the FCC thinks otherwise. However, it does not create a safe harbor for AI calls. Even under independent judicial analysis, it is hard to imagine a court ruling that a voice cloned by AI is not an “artificial voice” – the very term seems to encompass AI-generated speech. Indeed, legal experts caution that McLaughlin “does not create a loophole” to legalize AI voice calls without consent. The consensus remains that consent is still required for AI-driven outreach calls.

State Telemarketing Laws

In addition to the TCPA, many states have their own telemarketing laws that can be even more stringent. For instance, Florida’s Telephone Solicitation Act (FTSA) was amended in 2021 to prohibit all sales calls, texts, or even ringless voicemails made using an automated dialing system or recorded voice without prior express written consent. Unlike the federal TCPA, Florida’s law does not narrowly define “autodialer,” so automated calls that might escape federal TCPA coverage can still violate state law. Florida also provides a private lawsuit right with $500 to $1,500 per violation and has seen a wave of class action litigation under the FTSA. Other states, such as Texas, have recently updated their telemarketing statutes to broaden the scope of autodialer restrictions and close perceived loopholes. The lesson for we lawyers is clear: compliance must account for both federal and state telemarketing laws. If your client is deploying AI voices to contact consumers, ensure they obtain appropriate consent for all jurisdictions involved.

Best Practices for TCPA Compliance: Given the high stakes, companies using AI in outreach should bake compliance into their marketing strategy. Some practical tips include:

• Obtain Express Written Consent: This is non-negotiable. Document and retain proof of consent that specifically covers automated calls or messages and artificial voices. Consent language should be clear about the types of communications (calls, texts) and the technology involved.

• Disclose the Use of AI Voice: It may be wise (or even required in some cases) to inform recipients that a synthetic voice is being used. Lack of disclosure could be deemed deceptive. Transparency can also help avoid feeling of trickery if a customer realizes later they weren’t speaking to a human.

• Honor Opt-Outs and the Do-Not-Call List: Every AI-driven campaign must include an easy way to opt out (e.g., saying “stop” to a bot or pressing a key) and must scrupulously avoid contacting numbers on the National Do Not Call Registry. This is both a legal requirement and a basic risk mitigation step.

• Quality Control and Monitoring: AI systems can misfire – e.g., dialing the wrong person or number. Regularly audit call logs, wrong-number flags, and complaints. Perhaps an AI tool to review those logs in an automated way. Human oversight and testing of AI outputs are vital to ensure the AI isn’t stepping outside allowed scripts or contacting unintended targets.

• Rate-Limiting: Just because AI lets you scale up calls doesn’t mean you should blast thousands of calls in a burst. Throttle and pace the outreach to avoid spikes that might attract regulators’ attention or irritate consumers.

Regulators are actively watching this space. State attorneys general have shown growing interest in AI-based telemarketing, and the FCC’s 2024 ruling signals an enforcement mindset. The bottom line is that innovating in marketing with AI must go hand-in-hand with meticulous TCPA compliance. Lawyers should counsel their clients that the cost of a compliant program (consent gathering, vetting practices) is far less than the cost of defending a class action or regulatory enforcement for unlawful robocalls.

Advertising and Endorsement Law: Celebrity Voices, Deepfakes, and Right of Publicity

AI voice cloning not only implicates telemarketing rules – it also raises advertising law concerns, especially when the technology is used to mimic real people’s voices. In marketing, a voice itself can be a powerful asset. A recognizable voice (think Morgan Freeman’s narration or a famous actor in a commercial) carries brand value. But what happens if a company uses AI to imitate a celebrity or other individual without permission? U.S. law provides “right of publicity”protections that may apply, alongside laws against deceptive advertising. A technical issue that remains is the subjective notion of whether an AI voice sounds enough like the celebrity endorser to trigger liability. It also begs the question whether merely a human whose voice sounds just like Morgan Freeman’s, for example, is effectively prohibited from working as a voice actor because of the inevitable confusion.

The Right of Publicity – Voice as Identity: The right of publicity (ROP) is a state-law doctrine (no federal statute yet) that lets individuals control the commercial use of their name, likeness, and other aspects of identity. Importantly, courts have held that a person’s voice can be part of their legally protected likeness. The seminal case is Midler v. Ford Motor Co. (9th Cir. 1988), where singer Bette Midler sued Ford for using a sound-alike to sing one of her hit songs in a car commercial after she refused to participate. The court famously recognized that “a voice is as distinctive and personal as a face,” and it ruled that impersonating Midler’s voice in an advertisement without consent violated her rights. Similarly, singer Tom Waits won a case against Frito-Lay after an advertisement featured an imitation of his distinctive voice, misleading listeners to think he endorsed their product. These cases established that imitating a celebrity’s voice in commercials can be unlawful, even if the actual name or image of the person isn’t used. The voice alone, if distinctive and famous, can identify the person.

Fast forward to today: AI makes voice impersonation easier than ever. Instead of hiring a sound-alike actor, a marketer can use machine learning to clone a target’s voice from audio samples. But the legal outcome is likely the same. Using a person’s voice for commercial purposes without consent can trigger liability under right of publicity laws (and related theories like misappropriation or false endorsement). The Scarlett Johansson incident in 2024 is a prime example. When OpenAI introduced a new ChatGPT voice feature called “Sky,” Johansson noticed the voice sounded uncannily like her own. She accused OpenAI of appropriating her voice without permission, expressing shock and anger that the AI “sounded so eerily similar to mine that my closest friends and news outlets could not tell the difference.” Following public outcry, OpenAI pulled the voice and claimed any resemblance was a coincidence – though CEO Sam Altman did admit he had asked Johansson to officially voice the product (an offer she declined). While Johansson did not sue, her allegation underscored that AI voice cloning can cross personal and legal lines. Had this been used in a commercial advertisement or promotion without consent, Johansson would have a strong basis for a right of publicity claim.

State Laws and Expanding Protections: The right of publicity varies by state. California and New York are notable jurisdictions with strong protections, since many celebrities reside there. California’s law (and Ninth Circuit precedents) recognizes voice imitation as actionable. New York historically only protected name and likeness (not voice), but it recently updated its law. Meanwhile, new legislation is emerging specifically to tackle AI impersonation. In 2024, Tennessee enacted the “Ensuring Likeness, Voice, and Image Security (ELVIS) Act”, becoming one of the first states to directly regulate AI-generated likenesses. The ELVIS Act amends Tennessee’s publicity rights to explicitly cover a person’s voice – defined broadly as “a sound… identifiable and attributable to a particular individual, whether the sound is the actual voice or a simulation.” In plainer terms, any AI-generated audio that sounds like someone’s voice without consent can be illegal in Tennessee, even if it’s just a simulation. The law allows not only the individual (or their estate) to sue, but even record labels or licensees in some cases, and it targets those who create or distribute the unauthorized AI content. Tennessee’s move has set a precedent, and other states are following suit. California has a bill (AB 1836) pending to create civil liability for using “digital replicas” of deceased celebrities’ voices without permission. Illinois is considering adding “voice simulations” to its publicity statute and giving third-party rights to sue (like record companies). In total, at least seven states and several federal bills have been introduced to tackle AI deepfakes or voice clones of real people. This flurry of activity signals that lawmakers recognize the gap – AI has changed the game, and laws are evolving to catch up.

Lanham Act and FTC – False Endorsements: Apart from right of publicity, using a person’s persona in advertising without consent can raise issues under federal law. The Lanham Act (15 U.S.C. §1125) prohibits false endorsements – marketing material that falsely implies a person’s sponsorship or approval of a product. A celebrity whose voice is faked in an ad could argue consumers are misled into thinking they endorse the product. Indeed, Midler’s and Waits’ cases included claims that the ads created false associations. The Federal Trade Commission (FTC), which polices deceptive advertising, is also watching AI in advertising. The FTC has long-standing guides that using an endorsement or depiction of a person without disclosure can be deceptive. If an AI voice is used to sound like a real person (famous or not), and this misleads consumers, it could be deemed an “unfair or deceptive act” under the FTC Act. The FTC is already concerned about deepfakes and impersonation. In early 2024, the FTC proposed a new rule to ban AI-powered impersonation in commerce, specifically targeting scenarios where deepfakes mimic real people to defraud or mislead consumers. The proposed rule would prohibit not only the act of impersonation (e.g., using AI to imitate someone’s voice or image), but also knowingly providing the tools for such impersonation if they’re used in unfair or deceptive ways. While much of the FTC’s focus is on fraud (such as scammers mimicking voices of family members or company representatives), the rule signals a broader intolerance for deceptive uses of AI in commerce. Advertising that uses AI to create a false impression of a person’s involvement could invite FTC scrutiny and fines. At a minimum, marketers should disclose synthesized content to avoid confusion – for instance, a social media post using an AI-generated celebrity voice should clearly indicate it’s not the real person, to avoid a deception claim. This is apart from the various social media company’s rules themselves which are increasingly concerned about their entire content being overtaken by AI.

Practical Tips – Using AI Voices in Ads Legally: If you are advising advertisers they should implement safeguards if they want to use AI-generated voices in marketing campaigns and avoid liability traps:

• Obtain Consent or a License: The safest route is to get permission from the person whose voice you wish to mimic. This might involve a licensing deal (some celebrities might license AI rights to their voice for a fee) or using voice actors whose contracts allow AI cloning. For example, the actor James Earl Jones reportedly authorized an AI to replicate his voice for future Darth Vader performances – with compensation and control in place. Lacking consent, the legal risk is significant.

• Avoid Deceptive Implications: Do not make it seem like a real person is directly endorsing or speaking if that’s not the case. If you use a sound-alike, consider a disclaimer. E.g., “Voice simulation – not an actual endorsement.” This can mitigate confusion, though it may not cure a pure right-of-publicity violation in states like California. It does, however, reduce the chance of consumer deception.

• Steer Clear of Famous Voices Without Permission: Using AI to impersonate a celebrity voice is high-risk. As one AI law commentator bluntly put it, you cannot legally clone a celebrity’s voice for commercial content without explicit permission. Celebrities are protected by their state publicity rights – and now specific anti-deepfake laws like the ELVIS Act – so unauthorized use can lead to lawsuits or settlements. Even non-celebrities (e.g., an individual content creator or influencer) might have legal claims if their voice is distinctive and used commercially without consent.

• Monitor Emerging Laws: The legal landscape is rapidly evolving. Stay updated on new statutes (like those proposed NO FAKES and No AI FRAUD Acts in Congress ) and new state laws. An ad campaign that is legal today could become illegal next year under a new deepfake law. For instance, if the proposed federal No FAKES Act passes, it would create federal liability for distributing AI deepfake performances without consent . Lawyers should track these developments and adjust advice accordingly.

• Consider Morality and Backlash: Legal or not, impersonating someone’s voice can lead to public backlash and reputational issues. Companies should weigh the ethical implications. If consumers feel duped or if an artist goes public about unauthorized use of their voice (as Johansson did), the damage to brand trust can be severe. A lawsuit might be the least of the worries if public sentiment turns negative.

In summary, advertising law in the AI era demands careful navigation. The core principles – don’t deceive consumers, don’t misappropriate others’ identities – still apply, even if the technology is new. Right of publicity lawsuits are expected to rise as AI makes it easier to appropriate likenesses, and lawyers should be prepared to both defend and assert such claims. The safest course for marketers is to use AI voices creatively, but within the bounds of permission and truthfulness.

Corporate Compliance and Governance: Managing AI Marketing Risks

Beyond specific statutes and lawsuits, there’s a broader corporate law perspective to AI in marketing. In-house counsel and corporate compliance officers have a duty to ensure their company’s use of AI aligns with legal obligations and does not create undue risk for the business and its shareholders. AI voice cloning is a classic example of an emerging technology that, if mismanaged, could lead to litigation, regulatory penalties, and reputational harm – all of which impact the corporate bottom line.

Policy and Training: As corporate counsel or even outside counsel you can assist companies in their development of internal policies on the acceptable use of AI in marketing. This includes guidelines on when consent or authorization is required for using someone’s voice, what disclosures must accompany AI-generated content, and who must approve AI-driven campaigns. Training the marketing and creative teams is critical – they need to know that just because something is technically possible (like copying a famous voice) doesn’t mean it’s legally or ethically acceptable. A robust policy, reviewed by legal counsel, can prevent rogue experiments that cross lines.

Vendor and Tool Due Diligence: Many firms outsource AI capabilities or use third-party platforms for services like voice cloning, chatbot marketing, or automated dialing. It’s important to vet these vendors for compliance. Contracts with AI service providers should include representations that the technology will be used in compliance with laws (TCPA, intellectual property, privacy, etc.), and ideally indemnities if the vendor’s tool contributes to a violation. For example, if you hire a marketing firm that provides a list of leads and an AI voice solution, ensure the contract requires that the phone numbers are legally obtained and that the AI voices do not violate any rights. Also, check if the AI platform itself uses licensed data. (Some AI voice models might be trained on voice recordings scraped without consent – using such a model could inadvertently implicate your company in a misappropriation claim).

Board Oversight and ESG Considerations: At the board of directors level, oversight of AI use is increasingly seen as part of good corporate governance. Shareholders and regulators are paying attention to how companies manage AI-related risks. Ensuring compliance with marketing laws is part of this oversight. Boards might ask management for reports on AI usage and risk assessments. Missteps in AI marketing could affect the company’s public image and stock value (for instance, a scandal over deepfake marketing could be seen as a failure of governance and ethics). Thus, corporate counsel should brief the C-suite and board about these issues and the measures in place to mitigate them.

Privacy and Data Security Intersection: Voice cloning often relies on data – recordings of voices. Corporate legal teams should ensure that any personal data used to train or operate AI voices (e.g. customer voice recordings, actor’s recordings) is handled in compliance with privacy laws. Biometric privacy laws like Illinois’s BIPA might conceivably apply to voice prints or unique voice identifiers. There’s also the risk of deepfake fraud: criminals have used AI-cloned voices of CEOs to perpetrate fraud (calling subordinates to authorize bogus wire transfers, for example). Companies should fortify their authentication procedures to defend against that (like requiring secondary confirmation for financial transactions). While that is more of an internal fraud issue, it highlights why corporations need a comprehensive approach to AI voice tech.

Collaboration Between Legal and Marketing: Perhaps the most practical step is to establish a close working relationship between legal/compliance teams and the marketing department (as well as IT or AI development teams). Before launching an AI-fueled marketing initiative, there should be a legal review for compliance red flags – much like how most companies already vet marketing materials for truth-in-advertising. The legal team can help answer questions such as: Do we have consent for these calls? Are we using any real person’s identity? Should we add a disclaimer? Are we honoring all opt-out requests promptly? By being involved at the design phase, lawyers can often prevent problems rather than reacting after a complaint or demand letter arrives. AI Governance is going to be required for medium and large companies in perpetuity.

Finally, companies should watch the regulatory landscape. The FTC’s ongoing rulemaking on AI impersonation, the U.S. Copyright Office’s study on AI “digital replicas” and potential federal right-of-publicity legislation could all impose new requirements on corporate AI use. Proactive compliance – such as voluntarily labeling AI-generated marketing content, or obtaining licenses for any third-party likeness used – can position your client’s company ahead of the curve.

Litigation Landscape: Key Cases and Enforcement Trends

Despite the precautions, some companies will inevitably find themselves in court (or before regulators) over AI marketing practices. Lawyers should be aware of the emerging litigation trends in this area:

• Telemarketing Class Actions: As mentioned, plaintiff’s firms are actively hunting for TCPA violations involving AI. Recent class action complaints have argued that even sophisticated AI-driven outreach programs still fall under the TCPA’s prohibitions. We have already seen lawsuits targeting practices like ringless voicemails delivered by AI voices (arguing they constitute calls with an artificial voice) and texting campaigns run by algorithms. With the FCC’s 2024 clarification on voice cloning, plaintiffs are emboldened to sue first and ask questions later – knowing the law is on their side regarding consent requirements. Any company that engaged in mass calling or texting without proper consent, perhaps thinking AI somehow made it a gray area, could face a costly class action. Notably, some defense attorneys hoped the McLaughlin v. McKesson ruling (allowing courts to diverge from FCC guidance) might help evade TCPA liability for AI calls, but so far that appears wishful; no court has given AI marketers a free pass on that basis. The more likely scenario is continued class certifications and settlements – HelloFresh’s recent $14 million TCPA settlement, for example, shows the scale of exposure, and AI use would only multiply the number of calls or texts at issue. Companies caught in such litigation often settle due to the statutory damages ticking per call.

• Right of Publicity Lawsuits: On the advertising side, we have a few canaries in the coal mine. While Scarlett Johansson didn’t sue OpenAI, her situation highlighted how such a case would look. Another real case involved TikTok’s text-to-speech voice: voice actor Bev Standing sued TikTok in 2021, alleging the app used an AI-generated version of her voice without permission. That voice became infamous on countless TikTok videos. TikTok ultimately settled the lawsuit, agreeing to compensate Standing and license her voice for the feature going forward. The settlement was confidential, but it set a precedent that using someone’s voice (even a relatively unknown voice actor) in an AI application without a license can lead to legal claims and payouts. We can expect more suits from actors, singers, or even ordinary individuals if their voices are appropriated. In the music world, for example, the song “Heart on My Sleeve” – which used AI to mimic the voices of popular artists – went viral and raised alarms in the recording industry. While that instance was more about copyright (since it involved musical composition too), it spurred musicians to consider publicity rights as a tool. In fact, a group of over 200 artists and songwriters signed a letter calling the commercial use of AI-generated vocals “predatory” and demanding it be stopped . Litigation may ensue if an AI-generated song using an artist’s voice is monetized without consent – we might see lawsuits grounded in ROP or even trademarks (if an artist has a trademarked persona).

• New State Law Enforcement: With laws like Tennessee’s ELVIS Act now in effect, we could soon see the first lawsuits filed under these new statutes. A deceased celebrity’s estate might sue a company for an AI-generated ad featuring the celebrity’s voice, invoking the enhanced protections (the ELVIS Act extends post-mortem rights to voices and even allows record labels to sue in some cases ). The presence of statutory damages or attorneys’ fees in some of these laws will encourage plaintiffs. Additionally, state attorneys general could use consumer protection laws against deceptive AI marketing. For example, if a business ran a promotional campaign with an AI deepfake of a local public figure endorsing a product, a state AG could bring an action for unfair or deceptive practices in commerce.

• FTC and Regulatory Actions: The FTC’s stance suggests we might soon see enforcement actions or settlements with companies misusing AI in advertising. The FTC has penalized companies for fake testimonials and endorsements in the past; an AI-generated celebrity endorsement would fall into that category of deception. If the FTC’s impersonation rule (SNPRM) is finalized, it could give the FTC direct power to fine companies for using AI to impersonate someone without consent. This would add a federal hook beyond the current FTC Act’s case-by-case approach. Companies should recall that FTC penalties can be steep, and unlike private litigants, the FTC doesn’t need to show consumers were actually fooled – the mere capacity to mislead can be enough.

• Intellectual Property Disputes: While less directly tied to marketing, there is a crossover with IP law. We might see, for instance, a trademark-style claim if an AI voice is used in a way that confuses consumers about the source of goods or services (false association). Or, consider patent disputes if a company patents a certain voice cloning technique – competitors might litigate over the technology underlying these marketing tools. And as the U.S. Copyright Office completes its study on AI and “digital replicas” , any guidance or future legislation from that could spawn litigation over whether AI-generated voice content infringes copyrights or is protected expression.

Litigation Strategy and Defense: For defense attorneys, handling these cases will involve cutting-edge arguments. Potential defenses in a voice cloning ROP case might include First Amendment claims – e.g., the AI voice use was a form of expression or satire (if applicable) or was “transformative.” In California, there’s a recognized transformative use defense even against right of publicity claims (often tested in cases involving artistic works like video games). But in pure commercial advertising, that defense is weaker. Another approach might be to dispute identifiability: argue that the AI voice isn’t a close enough match to specifically identify the plaintiff (this could require expert analysis of audio similarity). As the FIRE legal explainer noted, if similarities are broad or generic, it might not violate publicity rights. However, with voices of famous individuals, juries are likely to find them identifiable if they are distinctive – Midler and Waits set a low bar in that regard (the entire point was the defendants tried to make the voices sound just like the plaintiffs).

On the telemarketing front, defense might focus on whether the technology used actually falls under the statute’s precise definitions, or challenge the class certification by arguing consent was obtained from some class members, etc. We may also see constitutional challenges to newer state laws (for example, arguments that laws like the ELVIS Act impinge on First Amendment rights by limiting creative works – though those laws do have exceptions for news, public affairs, and likely parody or commentary, consistent with First Amendment needs ).

Recent Case to Watch: Andersen v. Stability AI (N.D. Cal. 2023) – not about voice, but about visual artists – is an example where plaintiffs tried using right of publicity in the AI context. The court initially dismissed the ROP claim but left room to amend, specifically asking for more detail on how the AI use of the artist’s identity was tied to advertising or commercial endorsement. This indicates courts are open to publicity rights claims in AI cases if properly pleaded. Another is In re Clearview AI (E.D. Ill. 2020), where a class of individuals sued an AI company for scraping their online images for a facial recognition database. They successfully stated claims under California and New York publicity laws by alleging the use of their likeness was sufficiently connected to Clearview’s commercial interests (selling access to the database). By analogy, if an AI voice company used millions of people’s voice data to train a model it then sells, those individuals might explore a class action for misappropriation of voice.

The legal battles are just beginning. For we lawyers in this field, it’s crucial to stay informed on new filings and rulings. Each case that emerges will help define the contours of liability (or defenses) for AI voice cloning in marketing. Because the technology is evolving faster than statutes, much of the law in the next few years will come from judicial decisions grappling with how to fit AI within existing legal frameworks. Being part of that conversation – whether through litigation or advising clients to avoid it – is an exciting and challenging new frontier for advertising and technology lawyers.

Conclusion

AI voice cloning presents transformative opportunities for marketing, but also a minefield of legal risks. Telemarketing laws like the TCPA treat AI voices as they would any robocall – requiring consent and transparency, with steep penalties for non-compliance . Advertising and publicity laws protect individuals (famous or not) from having their voices appropriated in promotions without permission, and this area of law is rapidly expanding through new state statutes like Tennessee’s ELVIS Act and proposed federal legislation . Meanwhile, regulators such as the FTC are sharpening their tools to police deceptive AI impersonations , and courts are beginning to see cases testing the limits of using AI-generated content in commerce.

For lawyers advising corporate clients, the mandate is clear: partner with your marketing and tech teams early to ensure AI innovations are implemented responsibly and legally . This means securing the necessary consents, respecting individuals’ rights, providing disclosures to consumers, and keeping abreast of legal changes. It’s equally important to have a response plan – know how to handle it if a celebrity or customer complains that your AI campaign crossed the line.

AI in marketing will only become more prevalent, as companies seek to engage audiences in ever more personalized ways. By understanding the legal landscape – across corporate governance, advertising law, and litigation risk – lawyers can guide businesses to harness AI’s power without inviting subpoenas or lawsuits. In this dynamic environment, those who stay informed and proactive will provide an “innovation edge” to their clients, ensuring that exciting new marketing strategies remain on the right side of the law.

At stake is not merely one homicide prosecution, but the broader legal question of whether AI-generated facial recognition “matches” can form the basis for probable cause. With police departments across the country increasingly adopting facial recognition tools, this case offers an early look at how constitutional protections will fare in the face of sometimes black-box technology. It is also a case about the precision needed in search warrant affidavits to ensure new AI technologies are not the cause of evidence being suppressed.

Read more

In September 2024, the Council of Europe formally opened for signature the world’s first legally binding treaty on artificial intelligence: the Framework Convention on Artificial Intelligence and Human Rights, Democracy and the Rule of Law. Over 50 nations—including the United States, the United Kingdom, Canada, and many European Union member states—have since signed on. The Convention aims to ensure that the rapid development and deployment of AI systems does not come at the expense of human rights, democratic institutions, or the rule of law.

This post explores the Framework Convention’s structure, guiding principles, benefits, risks, and its likely influence on U.S. AI law. It synthesizes perspectives from government officials, legal scholars, industry voices, and civil society groups who are engaged with this international AI regulation.

Read more

Generative AI tools like ChatGPT, Anthropic’s Claude, and similar large language models are now creating content across industries – drafting emails, summarizing meetings, generating code, and more. With rapid AI adoption comes a new question for all of us lawyers: How do we preserve the artifacts of generative AI – the prompts we input, the outputs we receive, and even the fine-tuning materials or training data behind the scenes? Those same questions pertain to our clients as well.

These AI-generated artifacts present unique challenges under U.S. law when it comes to record-keeping and e-discovery obligations. In the rush to harness AI’s benefits, corporations (and even small businesses) likely have the same duty to preserve this information for litigation, compliance, risk management and governance purposes.

Below I explore the emerging legal expectations around preserving generative AI artifacts, the practical challenges organizations face, and steps you can take to ensure your clients meet their obligations. We’ll draw on recent commentary and case law – including a Reuters analysis from June 2025 – to shed light on best practices and risks for those integrating generative AI into their operations.

Read more

Small and solo practitioners are always busy folks. Between managing growing caseloads, responding to urgent filings, and keeping up with regulatory updates, lawyers are caught in a constant tug-of-war between billable hours and administrative overload. At the same time, AI and automation tools promise dramatic gains in productivity—if only they didn’t risk compromising client confidentiality.

For good reason, most of us do not want to have AI summarizing documents, evaluating client emails, texts, uploading drafts for refinement, etc.

Here’s the good news: you can have both.

With the right setup, a tool like n8n, an open-source workflow automation platform, can give your firm the power of an AI-enabled legal assistant—without ever sending a single client document to a third-party cloud service. This means no breaches of confidentiality, no worrying about where your data is stored, and no sacrificing your professional obligations in exchange for convenience.

Read more