Legal: OpenAI enters legal research as the checking duty spreads

Legal AI, the research and drafting software law firms subscribe to, gained a new kind of seller in September 2026. On 17 September OpenAI began selling legal research directly to large US law firms, through a model built around an index of US case law, and the best-funded legal AI startup bought a safety company. In the same three weeks US courts extended the duty to check AI output beyond a lawyer's own brief, and Spain raised what it costs to breach it. The tools moved toward the biggest firms. The checking duty moved everywhere.

What changed since we last looked

When Era Haus last looked at this field in Legal: California votes AI duties on lawyers as tools come bundled, California had just passed the first US bill putting AI duties on lawyers and the big vendors had moved inside the systems firms already run. The bill still awaits the Governor's signature, and this month's largest move came from a model maker.

On 17 September OpenAI released Astra for Law, a version of its newest model wrapped around a search index of US case law and statutes (OpenAI and LawSites, 17 September). The index holds more than 230 million sources, refreshed daily, built with the Free Law Project, a nonprofit that publishes court records. Early users include several of the largest US firms, among them Latham & Watkins. OpenAI's own test puts it at a 54% pass rate on legal research questions, against 38.7% for the same model with ordinary web search: an assistant whose citations still need a human check.

Three details matter more to a small firm. There is no published price. Access is by invitation, aimed at the 200 largest US firms; the connection for other software is listed as coming soon (Artificial Lawyer, 18 September). The index covers US law only. Of the 26 partner plugins on the launch list, Clio, the practice-management system many small firms in the US and other English-speaking markets run, was the only small-firm system (Artificial Lawyer, 18 September).

Thomson Reuters, which owns Westlaw, the paid case-law research service most US firms subscribe to, chose to plug in rather than fight, previewing a connector of its own (Artificial Lawyer and LawSites, 18 September). Its rival LexisNexis had said nothing by 21 September.

The money went to guardrails, and one firm became a vendor

Harvey, a legal AI company selling mainly to large firms, raised $550 million on 9 September at a valuation of about $15.5 billion, and bought Guardrails AI, whose software polices what an AI agent is allowed to do (Artificial Lawyer, 9 September; ABA Journal). The day before, Harvey shipped a search that reads a firm's whole document store and picks the files a question needs, with ethical walls, the internal barriers that keep one client's matter away from lawyers on the other side, enforced across the product (Harvey newsroom, 8 September). The question at the top of the market is now whether the tool can run across the firm's files without crossing a wall.

Morgan & Morgan, the largest personal-injury firm in the US, said on 14 September it will spend at least $1 billion on legal technology and AI over ten years (Bloomberg Law, 15 September; ABA Journal). It showed MX2, the platform it built for itself, which pulls medical records and drafts demand letters, the pre-lawsuit letters that set out what a claim is worth. It plans to license it to other firms from late 2027. A law firm becoming a software vendor is new.

The checking duty reached experts and clerks

On 31 August the US Fifth Circuit, a federal appeals court, heard argument on whether a Mississippi case should be taken away from a judge whose clerk used Perplexity, an AI search tool, to draft an order with invented parties and citations (ABA Journal, 31 August). No ruling had come by 21 September. In the US state of Washington a federal court threw out an expert whose report carried made-up citations and dismissed the plaintiff's claims, in an order of 18 August reported by the state bar on 31 August. In Spain, a Galician court fined a lawyer €1,800 and opened a disciplinary file over two dozen false citations in July, and on 11 September El Español reported that exposure now runs to €6,000, with a referral to the bar.

Europe produced no dated launch in the window. Latin America produced the one tool a small firm there can buy. Magnar, a Chilean legal AI platform, raised US$8 million on 9 September, says close to 30,000 lawyers in seven countries use it, and is entering Brazil and Mexico (La Nación, 9 September). The Argentine firm Beccar Varela invested, a law firm buying a stake in its own vendor.

What it means for you

If you run a small firm in the US, nothing launched this month that you can buy. The research market is changing above you: the model maker sells the index, the research publisher plugs in, and your practice-management vendor either made the plugin list or did not. It reaches you at your Westlaw or LexisNexis renewal, where a second quote now exists in principle. If you practice in Europe or Latin America, the US index does nothing for you, and the live change is the sanction: personal, growing, and increasingly followed by a bar referral.

The duty to check now attaches to work you did not write. An expert's report and a clerk's draft each carried fabricated citations into court this month, and the answer was a lost case rather than a fine.

What to do about it

Ask every expert you retain, in writing, whether AI was used and every citation checked, and keep the answer in the file. Ask your practice-management vendor whether it connects to the new research tools and at what cost; Clio users already have an answer. Watch rather than buy on Astra for Law until a price arrives, and watch the Governor's decision on the California bill, due within weeks. Anthropic's watermark detection service, which the last piece said to watch, was still in private preview at its 1 September release (The Next Web).