One Year In, I Can No Longer Keep Up
A blog that began by asking how legal education should respond to fast-improving AI now covers a field no one person can track — and that, more than any milestone, is the story of year one.
Two years ago I could keep up with AI and law. I read the academic papers, watched the YouTube channels, tried the tools, and held opinions that stayed current for months at a time. One year ago, the field had grown sufficiently that I could barely keep up. Today I can't even pretend to keep up. That is neither a complaint nor an admission against interest. It is the single most important fact about the field this blog covers, and it shapes what the blog does next.
Legaled.ai began a year ago today with a short welcome and a large question: how should legal education respond to artificial intelligence that keeps getting better? The urgency of that question has grown over these 12 months as law students certainly understand and enlightened legal academics are coming to realize. The energy surrounding Berkeley’s anti-AI default policy is Exhibit A.
What the year produced
The site published 62 posts in its first year, about 195,000 words, and grew to 420 members against 50,000+ views. And, yes, I am proud of the numbers, but they alone are not the point. I’m proud of the content.
One post fed Claude a connection to CourtListener and watched it pull a real case, state the holding, and reason from it — research that a year earlier would have required a Westlaw seat and a trained 2L. "Claude Meets Westlaw and Lexis" became the most-read post on the site. Another built a working law museum out of AI-generated exhibits to explore whether students learn doctrine better when they can walk through it. A third produced enough of a grounded law review article to make "could AI write one?" stop being rhetorical.
The second-most-read post was not a tool demonstration at all. "Law schools should not adopt Berkeley's awful new anti-AI policy" argued against a specific institutional choice, by name. That it landed near the top tells me the audience is at least as hungry for argument about what schools should do as for demos of what the tools can do. The institutional questions — take-home assessments, AI grading, confidentiality, the duty to prepare students for an AI-shaped profession — will get significant billing in year two.
How fast "fast" has gotten
Maybe the most salient development over the year in which this blog has existed has been the relentless pace of legal AI and legal tech. Exponential growth is no longer a metaphor about overgrowing lily ponds. It is what I see every day.
Take one benchmark. Harvey's Legal Agent Benchmark hands frontier models realistic legal assignments and counts a task as passed only if the output clears every line of an expert-written rubric. No partial credit. Two weeks ago the leading model completed 7.1% of tasks under that standard, with the rest of the frontier in the low single digits; the model released this week completes 13.3%. I know. Thirteen percent of finished, end-to-end legal work is not a rate that should induce Cravath to close up shop, but when the rate of A+ achievement doubles while I am drafting this post, that is a noteworthy event. Moreover, the all-pass standard hides how close the misses run: in the months following the inauguration of this blog Vals AI's legal research benchmark found that general-purpose models now match or beat practicing lawyers on accuracy, and in a blind contract-drafting study the best models edged out the best human lawyer on first-draft reliability. A widely heralded Stanford study released earlier this month shows that AI responses to hypothetical student contracts questions are now preferred to those from expert faculty.
Take model releases. As I write in June 2026, Claude Opus 4.8 is just a few weeks old and already overtaken by Claude Fable 5, the first of a new tier above Opus, released this week; Gemini 3.5 Pro was announced at I/O three weeks ago and is due any day; prediction markets put a GPT-5.6 release this month at better than 80%; and what is being styled NotebookLM 2.0 from Google with new agentic capabilities and yet more output capabilities begins rolling out over the next few weeks. Honestly, at this point the new releases arrive faster than I can finish writing about the last.
Take money. Legal tech raised $6 billion in 2025; the first quarter of 2026 added $2.34 billion across 103 deals. The money tells two stories at once. At the top, concentration: Harvey reached an $11 billion valuation in March, Legora hit $5.6 billion in April with Nvidia aboard, and three companies took nearly two-thirds of the quarter's total. The proliferation of LegalTech companies has accelerated dramatically over the past year, driven by generative AI adoption across law firms and corporate legal departments. According to Tracxn, 148 new LegalTech startups were founded in 2025, contributing to a global total exceeding 10,800 companies, of which 1,892 are funded and have collectively raised $26.3 billion. A legaltech vendor (Spellbook) created a $1 million initiative for legal education, including $25,000 fellowships for students who demonstrate that they will use technology brilliantly. Money does not move that way around a technology it expects to plateau.
And take what the funding charts miss: the open-source stack — of which the models are just one component. There are now open weight models from DeepSeek and Alibaba (Qwen) that perform as well as the best frontier closed weight models did a year ago. There are even models such as Gemma4 from Google that run on a (high end but not crazy) personal computer that compete well with the leading versions of ChatGPT or Gemini from a year ago. But the harness around the model has come to matter as much as the model: in May an ex-Latham associate shipped MikeOSS, a free, open-source replica of Harvey and Legora's core product, built in two weeks, and from Helsinki came Lavern, an open multi-agent system that sets sixty-seven specialist agents arguing over a matter through evidence-backed debate and can run entirely on a local machine. I built a decent cert petition using it.
Open weight models inside open harnesses are important not only for what they can do but for what they can keep safe from prying eyes. Among other things, they make on-premises, confidentiality-respecting deployment thinkable for law school clinics and unpopular public interest organizations. I would tell you how they perform on legal tasks, except that I mostly cannot: Texas, my employer, has placed DeepSeek and most other Chinese models on its prohibited-technologies list, which binds public-university employees except under limited conditions. The models best positioned to solve legal education's privacy problem are unfortunately the ones I am least free to test. (See request for guest bloggers below!)
Raising the bar, and a request
There are two honest responses to a beat that has outgrown its reporter, and neither is quitting the part I like.
The first is to raise the bar, not switch beats. I am not giving up tool demonstrations; they are too interesting and too much fun. But a demo now must either be super fun (as determined by me) or clear a higher hurdle than "look what it can do." The year-two test is whether the capability should change what can occur in legal education or, sometimes, legal practice — its assessments, its curriculum, its confidentiality rules, its hiring, its production of legal fodder. One theme I expect to press on the practice side: access to justice. AI-assisted arbitration and mediation look to me like the most plausible route yet for getting real dispute resolution to the people who now go without it, and I am hoping to develop and test some tools in those arenas. I concede that this will constitute a line extension for legaled.ai — dispute resolution is practice, not pedagogy. But legal practice and legal education are entwined, and should be: a profession that resolves disputes differently will need schools that teach differently. And an app that both conducts mediation and coaches students on how to improve performance could be extraordinarily valuable.
The second response is to stop doing this almost entirely alone. So here is the request. If you teach law, practice it, study it, or build the tools, and you have tested something worth showing or an argument worth making, the blog is now open to guest contributors. Write me. The standard is the standard the blog has tried to hold itself to — show the thing, concede the other side, and say what it means for education. I am not looking for press releases or vendor pitches.
What year two is for
Some of the drafts underway are institutional: a degree in legal engineering and whether any school should offer one; whether faculty should write fewer law review articles and build more reusable teaching skills; how AI might tell us which schools beat their predicted bar-passage rates and why.
Others are demonstrations I cannot wait to run, because voice has gotten good. Speech-to-text now runs well on a local machine using free apps such as TypeWhisper you can download. Speech-to-speech models such as gpt-realtime from OpenAI or Gemini Live from Google answer in less than a second and may help where input is oral rather than written. And affordable text-to-speech is beginning to exist that is both affordable and can hold your attention for whole minutes before the computer voice starts to grate. Kokoro is an example that runs on your machine. OpenAI TTS-1 swiftly produces even better voices for a moderate charge. These advances make many things buildable today. One can readily imagine an app that transcribes the professor while a sidebar supplies grounded commentary on the cases – at last there is competition with Etsy for student attention. One can imagine an automated lecture creator that turns a syllabus section into text, audio, and slides – maybe even video in a year or so. Live AI feedback in advocacy classes may not be far behind. Plus, law school is more than its classrooms. Admissions, institutional research, even the marketing office run on documents and prediction, and AI is or certainly should be seeping into all three. I want to write about that too.
Ordinarily, I'd be a little embarrassed if a good fraction of what I wrote during the preceding year was already a period piece. But AI is moving so fast that this half life is no cause for shame. Moreover, the posts keep a record of how legal education met, or failed to meet, a technological shift while it was happening — worth keeping even when any single entry is obsolete by the next release.
Thank you to everyone who read, subscribed, shared, pushed back, or tried one of the ideas here. The second year will bring developments I cannot now picture — which is the reason to keep testing, measuring, arguing, and writing. Please join me for the continuing journey.