Google and Anthropic Sell AI to Law Firms as Courts Penalize Fake Citations
Google placed Gemini Enterprise for Legal into preview in late August with four firms participating: Gottlieb, Freshfields, Weil and Williams & Connolly. The preview arrived about three months after Anthropic released Claude Legal Solutions. Engadget
Two makers of large AI models are now selling directly for court and deal work. These are not general chatbots that lawyers adapt on their own. They are products packaged for legal tasks. Anthropic shipped first. Google followed with an enterprise preview built with large-firm partners.
The designs are similar. Google sends questions through connectors, which are links to outside legal databases, including Everlaw and NetDocuments. Anthropic lists 20 connectors to legal platforms and 12 pre-built plugins, small add-ons for drafting, review and filing preparation. Engadget
The method will be familiar to technical readers. The AI retrieves answers from customer files and third-party stores, then uses task-specific add-ons. Courts have already issued penalty orders for failures of this method.
A Stanford HAI study found that legal AI models invent information in 1 out of 6 or more test questions. Stanford HAI The National Center for State Courts describes those hallucinations as fake case citations, misdescribed rulings, or false procedural details. NCSC
Grounding by connector
Researcher Damien Charlotin's database has counted nearly 2,000 cases worldwide with AI-made legal errors, involving more than 800 lawyers and 1,100 people representing themselves. Engadget
Canada provides numbers over time. Reported cases with invented citations rose from 7 in 2024 to 86 in 2025, with 39 more in the first quarter of 2026. In the United States, penalties over false AI briefs were still rising as of April 2026. NPR
The cases cut across firm size and court. Sullivan & Cromwell filed an emergency letter with the U.S. Southern District of New York in April after a bankruptcy motion to Chief Judge Martin Glenn was found to contain 42 AI-made falsehoods. A federal judge in Mississippi fined four lawyers after finding both sides used AI in the same lawsuit. NYT Lawyers for a former Los Angeles homeowner found AI-made errors in State Farm's court filings. CalMatters
Smaller dockets show the same pattern. A U.S. appeals court ordered a lawyer to pay $2,500 over AI errors in a brief. Reuters A Massachusetts lawyer was penalized for citing fake AI-made cases in court papers. Connecticut's top court warned lawyers about AI risks after fake citations in a ruling that was the court's first on penalties for AI-made errors, with attorney Ian Gottlieb expressing regret in connection with the case. Reuters
Ontario added a costs decision with an exact figure. The Law Society Tribunal ordered lawyer Shahryar Mazaheri to pay $31,150 in a June costs decision for filings with AI-invented citations. Mazaheri had used an earlier version of Grok to draft a factum, a written legal argument, for a client's appeal that contained invented citations and invented legal rules. Engadget
A sanctions record that keeps growing
This is not a new risk. DoNotPay dropped plans in early 2023 to have an AI robot lawyer argue a case in court after state bar associations threatened criminal prosecution. Scale is different now. General models put drafting help in every browser, and legal-specific products put it inside document systems where lawyers already work.
Connectors narrow the set of documents the AI searches, but they do not shift professional responsibility. Grounding in Everlaw or NetDocuments can reduce open-web invention. It still leaves misdescribed rulings, misquoted summaries and false procedural claims, the exact types courts keep finding. Checking remains human work, citation by citation, and penalty orders make clear that courts will treat unchecked model output as the lawyer's own work.
In my view, that strict responsibility is useful. I have watched adoption cycles where convenience ran ahead of controls, from early web research to cloud file sharing. The correction usually came through audit and liability rather than better demos. The tools that win will build checking into the workflow, with pinned sources, exact quotes that can be compared and refusal when the store lacks support, not just more plugins.
The broader context here is hopeful, though demanding. Legal work is text heavy, bound by past cases and built around audit, so retrieval-grounded help should fit well once the citation layer can be trusted. The current fines are the cost of skipping that layer. Firms in the Google preview and users of Claude Legal Solutions now have a chance to define careful supervision in practice, and to make the next count lower than the last.
