Legal: California votes AI duties on lawyers as tools come bundled

Legal AI, the software lawyers use to research, draft and review, changed in two places in August 2026: the rulebook and the products. On 31 August California's legislature passed the first US bill putting AI duties on lawyers, including personally checking citations; it awaits the Governor. In the same month legal AI products began arriving inside the systems a firm already pays for rather than as separate tools. The two moves point the same way. The checking stays with the lawyer, and the software moves to where the files are.

What changed since we last looked

When Era Haus last looked at this field in Legal: AI use went mainstream as courts drew a global line, courts on three continents were punishing invented citations. Four weeks later a legislature went further. California's Senate Bill 574 passed both chambers on 31 August 2026, per the Legislature's record, and still needs the Governor's signature.

The duties, in the text as amended on 21 August: verify what a generative AI tool produces, personally check every citation in a filing, correct any wrong output, tell the court when AI was used in a filing, and keep client confidences and personal data out of AI systems without access controls. The bill also bars a lawyer from delegating "the practice of law" itself to an AI, a phrase California has never defined tightly, so that clause will be the one argued over. This summer's court sanctions would become the minimum standard for every lawyer in the most populous US state, whether or not a judge catches the error.

The EU AI Act's transparency rules applied from 2 August, including a machine-readable mark on AI-generated content, while its high-risk rules slipped to 2 December 2027 (European Commission, 31 July). The marking rule has a practical twin. Anthropic, maker of the Claude models, has embedded an invisible watermark in its newer models' text since 2 August; a detection service exists only in a limited private preview (Anthropic, 11 August; Artificial Lawyer, 13 and 17 August). Heavy rewriting removes the mark, so it proves little alone, but a court or opposing party can now ask whether a brief was machine-drafted, and reused templates carry the marks.

The tools moved to where your files are

The product side changed in the three weeks around ILTACON, the legal technology industry's annual conference, which drew 5,782 registrants in late August 2026 (Above the Law, 1 September). The largest vendors all released new versions then, converging on one idea. The AI sits inside the system that holds the firm's documents, remembers how each lawyer works, and shows who used it on which matter.

Google set the pattern on 25 August with a legal edition of its enterprise AI platform, built to plug into the document and research systems a firm already runs (Google Cloud release). Four large firms previewed it, among them the London firm Freshfields.

Harvey, a legal AI company selling mainly to large firms, released a new version on 18 August with a memory that learns each lawyer's drafting habits, per Law.com. Thomson Reuters, the legal research publisher, followed on 20 August with a new CoCounsel, its AI research assistant, US only for now (company release).

For a small firm the relevant launch came from Clio, the practice-management system many small firms in the US and other English-speaking markets run. On 25 August it put research, drafting and monitoring of new US court filings inside its workspace (Clio release).

Legora, the Swedish legal AI company, is opening offices in three continental European capitals (Legora newsroom, 31 August). Harvey's August product notes add Spanish as a working language, with sources for Spanish and Mexican law. Early movers have stopped piloting. The German firm HEUKING extended Harvey to every professional group, including non-lawyer staff, on 5 August (Harvey newsroom). None of the launches published a price; a small firm cannot compare them without a sales call, which is itself a finding.

What it means for you

The statute, if signed, makes checking the AI's output the lawyer's personal duty, and the largest products now record who ran what on which matter, behind which access wall. That record is your defense in a hearing and your answer to the client. Proof that the tools pay at small-firm scale did not arrive with them. Clio's US Legal Trends survey of solo and small firms, published in May 2026, found fewer than a third had increased revenue with AI, against nearly 60% of large firms.

Pricing, which the last piece told you to decide, has a working example, launched on 18 August (Artificial Lawyer). Athena, a new US firm backed by the venture investor Andreessen Horowitz, does deals and real estate work on fixed fees agreed before work starts.

What to do about it

First, adopt the California duties as your own process, wherever you practice. A named person checks every citation against the source before filing, confidential material stays out of any tool not contracted for it, and the file shows where AI was used. Second, before buying a legal AI tool, ask whether it sits inside the system that holds your files, and ask to see the audit view; the largest vendors now include one. Third, assume AI-drafted templates may carry a mark. That is harmless if the check was done, and a reason to keep the record.

Watch rather than act on the detection service and the Governor's decision; both may be settled before we look again.

The pattern underneath

New this month is that the duty and the record arrived together. A legislature wrote the obligation to check into a bill, and in the same four weeks the vendors built the proof of who checked into the software, neither waiting for the other. A small firm that keeps that record has met the duty before the bill is signed; one that does not will hear the question from a client or judge sooner than it expects.