The International Legal Technology Association conference is, as always, the big one on the legal tech calendar. It’s the Davos of doc review, the Super Bowl of software integrations, and the Oscars of people who unironically use “lawyers” and “technology” in the same breath. This year, 4600+ legal tech professionals and vendors gathered in Maryland for a weeklong salute to all technology… but mostly AI.
National Harbor exists not so much a town as a synthetic terrarium meticulously constructed to drag conference business out of DC. A simulacrum of a city center, with themed restaurants and bars ideally suited to host client dinners or vendor happy hours. As a conference site, it’s a perfect, enclosed economy running on lanyard-strung badges and drink tickets. There are people who complain about this venue and I will fight every one of them. Anyone whining about National Harbor as a venue should be sentenced to ninth circle of hell (just past the Magnolia Rooms in the Gaylord Nashville — which is, while we’re on the subject, the location of this event next year).
This trade show Narnia served as an eerie counterpoint to the tragi-buffoonery unfolding across the Potomac. A peaceful, walkable village contrasted by a city under siege. Neither reflects reality much. National Harbor is basically a soundstage while D.C. is the safest it’s been in decades. The Humvees cruising the drop-off lane of Union Station stood in for the piano bars in National Harbor, but both are just props backing up an illusion. For National Harbor it’s to provide a comfy conference host that doesn’t feel like it’s been plopped in the middle of nowhere. For D.C. it’s to distract everyone from how many times Donald Trump might show up in the Epstein files. One of these projects is more successful than the other.
But the key to appreciating National Harbor is recognizing it as a custom-built conference center. It’s not a real city, it’s a simulation designed for dispensing knowledge and sterno-warmed chicken dumplings. It’s ruthlessly efficient at delivering both, but you can’t lose sight of the fact that it exists to make attendees feel good about the experience. Not unlike the AI products dominating the conversation. Since arriving on the scene, artificial intelligence has inspired legal tech vendors to design all manner of products delivering on the promise of increased productivity. But most of what we hear about are the lawyers who forgot that the AI experience is an illusion of its own.
“Hallucinations” aren’t the problem, lawyers failing to check their work before sending it out the door is the problem. But that’s also a bit reductionist. There’s a psychological dimension to the chatbot interface. Lawyers wouldn’t trust a summer associate’s brief on face, so why are they trusting ChatGPT? The Christine Lemmer-Webber description of it as Mansplaining as a Service gets a lot of the way there — as a tool, AI delivers results with supreme confidence no matter how wrong it might be. Though that’s not the whole story, because mansplaining takes a condescending tone while the problem with these bots continues to be their overzealous compulsion to give the user the results they want. Standing in the Gaylord’s Belvedere Lounge last week, I explained that it’s more like the guy who’s become convinced that the stripper is in love with him.
The medium is the message, as McLuhan would say. Legal tech vendors expend massive resources to make sure AI products deliver more reliable results, but they’re fighting a constant battle against a public AI sales pitch telling the world that AI isn’t just finding evidence, it’s finding answers. Even though “the answer” is often what the lawyer is hired to get around. But that’s going to be the sales pitch, because no one gets megarich promising cautious improvement, they have to drive revolutionary change… whether it’s warranted or not.
Which becomes its own hallucination. Do we have a word for industries built on shared illusions?
The official theme at ILTACON was… pirates. Attendees dressed up like pirates, trading plastic doubloons for free drinks and snapping pictures around the impressive Crow’s Nest at the end of the exhibit hall. Nothing says “NOT A BUBBLE” like dressing up as the romantic ideal of people showering themselves in stolen wealth.
To be clear, this isn’t to say the legal technology sector — or more specifically the AI component of it — is a bubble. Vendors outdid themselves this year in developing new and more interesting ways to deploy AI to improve the legal workflow. Definely launched its Cascade product, using AI technology to track first, second, and third-order knock-on effects from contract changes to combat negotiation whack-a-mole. Everlaw showed off a new deep dive tool to allow more senior attorneys to interrogate their document sets at every stage of the litigation. Both NetDocuments and iManage continue finding new ways to automate the process of making the firm’s own data be more useful.
Legal AI providers may not be a bubble, but they could well be one of those rainbow swirls shimmering beautifully on the surface of an underlying AI bubble.
Most of the AI on display at the show is still “building off” other products — the “foundational models” to use the parlance of the trade. At the end of the day, a lot of this stuff rests on the energy guzzling backs of OpenAI, Anthropic, Gemini, and MechaHitler (or whatever Grok is calling itself now). Fawning media coverage and half-trillion-dollar valuations suggest this is a gravy train extending decades into the future, but can this really hold up? To quote tech industry analyst Ed Zitron, “by the end of 2025, Meta, Amazon, Microsoft, Google, and Tesla will have spent over $560 billion in capital expenditures on AI in the last two years, all to make around $35 billion.” What happens when one — or more — of the companies behind these models runs out of cash to pay the bills? The fact that legal tech providers are building stuff on top that can actually pay the bills won’t matter much if the foundational bot goes dark.
“Pirates, Be Ye Warned,” that we’re a lackluster NVIDIA earnings call from a precipitous drop.
Seriously, how does OpenAI pay off a $500+ billion valuation? How is this revenue supposed to arrive? Tokens are the coin of the realm in AI, and unlike groceries, they’re actually getting less expensive. A massive surge in users isn’t likely barring the imposition of year-round school — a recent study shows AI subscriptions plummet when students don’t need it to write papers for them over the summer — it’s hard to imagine where the AI world expects to get all enough volume to make back their money. It sounds crazy to suggest a half-trillion would just disappear, but it sounded crazy to suggest Lehman Brothers would disappear until it did.
Few tech observers are ready to embrace a notion this grim. But several at the show seemed willing to acknowledge this risk, even if they wouldn’t admit it out loud. When I’d suggest the possibility of a foundational model provider going belly up, one industry insider said directly, “I think that’s very possible.”
The show arrived against the backdrop of GPT-5 arriving with a general “meh,” which certainly didn’t help promote the sense that we were all riding AI to the moon. The newest OpenAI model didn’t come up much throughout the week, surprising for an announcement with so much hype, and it’s prompting the broader computer science community to ask “What If A.I. Doesn’t Get Much Better Than This?” (which, as a title, certainly sounds familiar…). While the most audience-friendly take on that question deals with the much ballyhooed “displacement” and whether or not it will actually place every human in a Matrix-style incubator by 2027, it’s worth realizing that without an exponential step improvement for AI, the only displaced humans will be the ones working for the foundational model providers.
Though most kept a positive outlook while signaling caution. Lexis, showing off their new AI-powered legal research tool stressed that they’re constantly evaluating the foundational models and suggested they could quickly swap to another model as necessary. In fact, one of the most singularly practical features of the new Protégé offering is the option of using general AI models of the lawyer’s choice (4o, 3o, 5, or Claude Sonnet) from within the Lexis tool. Questions that lawyers might otherwise plug into these consumer facing models can be posed within the secure Lexis environment. The purpose was empowering choice, but it’s also useful for keeping options open. Of note, GPT-5 failed to sufficiently impress Lexis into updating the melange of algorithms they use for different purposes within its Legal AI tool — another interesting dig at the newest model heard at the conference.
Thomson Reuters CEO Steve Hasker underscored the company’s commitment to the large language models we all know and love/hate, but interestingly took the opportunity to heavily tout the company’s acquisition of Safe Sign Technologies. Hasker said he believed the Safe Sign scientists “were in the process of building the best, small language models for the legal profession,” and that Thomson Reuters is investing significantly in their work. As a strategic matter, the decision to devote resources to building, for lack of a better analogy, an American DeepSeek seems like the powerful hedge. Even if none of the big providers collapses under their own bloated valuations and paltry revenues — and I still think they will — Thomson Reuters will have a model on-hand that can be thrown at problems far more cheaply and, continuing the DeepSeek analogy, more accurately based on a cleaner training regimen. Some firms, he said, are already asking about bringing instances of these model behind their firewalls.
Maybe the lasting accomplishment of large language models will be the small language models we make along the way.
Hang around National Harbor long enough and the cracks become apparent. Illusions only last so long, after all. The bar immediately across from the Gaylord sat abandoned this time around, a particularly jarring development considering it had a built-in customer base of sloppy conventioneers. It certainly seemed unthinkable that one of these curated entertainment experiences could fail. Is it so absurd to imagine an artificial intelligence behemoth will go under like that bar? National Harbor carried on without this watering hole, of course, but the absence served as a reminder: all gin joints are fleeting. AI is too powerful to disappear from the legal workflow, but strolling the conference this year, it seemed as though providers would do well to consider the possibility that one or more of the models underlying all this progress might implode. For that matter, customers need to make sure they’ll be covered. Until then, the band plays on, the rum flows, and everyone convinces themselves they’ll be holding the doubloons when the lights come on.
Joe Patrice is a senior editor at Above the Law and co-host of Thinking Like A Lawyer. Feel free to email any tips, questions, or comments. Follow him on Twitter or Bluesky if you’re interested in law, politics, and a healthy dose of college sports news. Joe also serves as a Managing Director at RPN Executive Search.
The post Dispatches From The AI Bubble: ILTACON 2025 appeared first on Above the Law.