California's legislature sent Newsom the first state statute writing AI verification duties into lawyers' obligations, the same week the DOJ told a court that training AI on copyrighted text is fair use. Precisely launched a CLM platform that keeps AI out of legal risk decisions, Wilson Sonsini licensed a litigation drafting tool, and Harvey and Legora turned ILTACON into legal tech's biggest marketing spend yet.
On August 31, the California Senate concurred in Assembly amendments to SB 574 by a 39-0 vote, sending the bill to Governor Newsom. If he signs it before his October deadline, California becomes the first state with a statute, not an advisory opinion, that tells lawyers and arbitrators exactly what using AI requires of them. Where the New York City Bar's policy paper last week said AI can't substitute for judgment, SB 574 tries to say what exercising that judgment actually looks like in a filing.
Sponsored by state senator Tom Umberg, who chairs the Senate's judiciary committee, the bill requires lawyers to disclose AI use in court filings, take reasonable steps to verify AI-generated content including citations, and correct anything false before it reaches a judge. It bars entering confidential client information into public generative AI systems and explicitly prohibits delegating the practice of law, or an arbitrator's decision-making, to AI outright. The bill carries no independent penalty regime of its own; violations would be enforced the way AI-related failures already are, through court sanctions and State Bar discipline.
SB 574 codifies a duty every legal team already owes its clients: verify before it ships, and know who's accountable if it doesn't. What the statute doesn't provide is the mechanism for meeting that duty at volume. Reading and verifying every AI-generated citation is itself billable, expensive work, and if it lands on the same lawyer who's already stretched thin on everything else, the statute just adds a compliance step to inexpensive work still being done by expensive resources. A supervised routing layer, agents that draft against a team's own templates with a human check built into the workflow before anything leaves the building, is what turns "verify before filing" from a duty a lawyer hopes to remember into a step that happens by default.
On September 1, the Department of Justice filed a Statement of Interest in the multidistrict litigation against OpenAI and Microsoft, urging the Southern District of New York to rule that training large language models on The New York Times' and other publishers' copyrighted works is fair use. It's the first time the federal government has staked out a formal position in the wave of copyright suits against AI labs. The filing addresses what an AI model can be built from. It says nothing about what a legal team is responsible for once that model's output reaches a client.
DOJ's brief argues the litigation risk facing AI labs threatens US competitiveness and national security, and that courts should weigh the public benefit of LLMs assisting writers, researchers, and even government officials as part of the fair-use analysis. The New York Times has objected publicly to the administration weighing in on a case where the outcome could benefit a company it has separately discussed taking a financial stake in. Whatever the SDNY judge decides, the Statement of Interest doesn't touch a separate and, for enterprise legal teams, more immediate question: a model can be legally trained and still generate a fabricated citation, and someone still has to catch it before it reaches a filing.
If the DOJ's position holds, the legal risk around how a model was trained shrinks for every vendor building on frontier models, including the ones your team already uses. That doesn't reduce your own exposure by a single filing. The fair-use fight is about the input. Your liability, and California's SB 574, is about the output. Has your team's review process kept up with how much of that output is now AI-generated, or is it still sized for a world where a person drafted the first pass?
Precisely launched Lexnus on September 1, a contract lifecycle management platform built on what the company calls an uncomfortable premise: most CLM tools start in the wrong place, the document, when the actual risk decision is upstream of any document. Lexnus separates the two: AI can draft or analyze a contract in seconds, but a company's own approved legal policy, built through playbooks, decides what risk that company will accept, and AI doesn't get a vote.
The platform runs on a choice of large language models, from Claude to Mistral, rather than locking a customer into one vendor's model, and it's hosted on EU-sovereign infrastructure. Precisely says a team can see first results in under an hour, with no traditional implementation project and no implementation fee, a direct pitch at the CLM category's reputation for multi-month rollouts. The architecture is the notable part: policy sits above generation as a fixed layer the AI has to work within, not a setting the AI can reason its way around.
This is the architecture Flank builds on: agents that generate fast, and a policy and escalation layer above them, built from a team's own templates and terms, that decides what's acceptable before anything reaches a human for final sign-off. The model underneath is replaceable. The policy isn't, because the policy is the actual asset a legal team owns. Precisely reaching the same conclusion for contracts is a second vendor, independently, deciding that inexpensive work done by expensive resources isn't fixed by a smarter model. It's fixed by separating what the model does from who's accountable for the outcome.
ILTACON drew a record 5,782 attendees to Nashville in late August, and this week's retrospectives from LawNext and Above the Law focused less on the sessions than on how much Harvey and Legora spent to be seen there. Harvey occupied 24 ten-by-ten booth spaces to Legora's 22, at a reported $10,000 per space before any partnership discount, and both companies threw parties with paid celebrity performers, Lady A for one, Sheryl Crow for the other.
Highway billboards and Uber ads around Nashville carried legal AI branding for the week. A veteran legal tech CEO, quoted by LawNext, called it "peak legal tech." None of this is evidence of anything about either platform's actual accuracy, retention, or supervision model, it's a demonstration of how much venture capital both companies are willing to convert into booth space and a Sheryl Crow performance, estimated by trade press at $150,000 to $500,000 for a private corporate booking. Booth size and party budget are the same kind of signal as a funding headline: a company can afford to spend, which is a fact about its cap table, not about whether its output needs a lawyer's review before a client sees it.
A booth budget and a party lineup tell a buyer which vendor has the most capital to deploy on visibility this year. They tell you nothing about which vendor's supervision model actually catches a fabricated citation before it reaches a client. When the marketing spend is the loudest signal in the room, what's your actual evaluation criteria, and is it based on anything either company disclosed at ILTACON?
Wilson Sonsini moved from evaluating to a paid commercial license for LexText, a litigation drafting and analysis platform founded in San Francisco in 2024, according to LexText chief executive Jobe Danganan. A firm that bills partners and associates at full litigation rates is now paying a separate vendor to automate the drafting those same billable hours used to cover.
Ben Crosson, a Wilson Sonsini partner and co-chair of its nationwide securities litigation practice, said LexText is "built for the way litigators actually work." The platform's guided workflows pull the case-specific factual record from a lawyer's own uploaded filings, discovery, transcripts, and exhibits rather than from the model's training data, which is the detail that separates a drafting tool grounded in a matter's actual record from one that's just generating plausible-sounding text. Neither party disclosed seat counts, which practice groups are using it, or commercial terms.
Wilson Sonsini didn't buy LexText because litigation drafting stopped being valuable, it bought it because that drafting is high-volume and repetitive enough that automating it makes commercial sense even for the firm charging full rates for it. That's the same arithmetic enterprise legal teams should be running on their own high-volume work: NDAs, standard redlines, procurement paper. If a law firm will automate its own commodity drafting to protect margin, an in-house team routing that same commodity work to outside counsel by default is paying full price for exactly the labor the firm itself just decided to make cheaper.
California wrote a verification duty into law. Washington cleared the way for AI labs to keep training on copyrighted text without saying who checks what comes out the other end. Precisely built a product that puts policy above generation, and an Am Law 100 firm bought a tool to automate its own drafting. Every story this week assumes a check exists somewhere in the chain. Only one of them, Precisely's, actually built where that check sits into the architecture.
Take the week's threads together: two government actions defined duties and cleared inputs without touching who checks outputs, one vendor built the check into its product, one conference proved capital isn't the same as capability, and one law firm quietly admitted its own commodity work is worth automating. None of that is routing infrastructure a legal team owns for itself.
Outsource legal work to supervised agents and the checkpoint stops being a statutory duty a lawyer has to remember to discharge under time pressure. It becomes a mechanism: agents that draft against a team's own templates, terms, and escalation rules, with a human reviewing the output before anything leaves the building. Inexpensive work stops being done by expensive, unsupervised resources, whether that resource is an unchecked model or a firm billing full rates for work it's already trying to automate itself.