Law firms trying Google's and Anthropic's legal AI still answer to judges for every citation

Shahryar Mazaheri used an earlier version of Grok to draft a factum, the written argument for a client's appeal. The Ontario Law Society found last December that it produced fabricated citations and invented legal principles ^[6]. Adjudicators called the materials "gibberish" ^[7]. A language model, they wrote, "does not appreciate nuance or exercise judgment or use a moral compass" ^[8]. Google and Anthropic both present their legal products as built for verification. Gemini Enterprise for Legal routes queries through connectors to legal databases such as Everlaw and NetDocuments ^[11]. Claude Legal Solutions has a reported 20 connectors to legal platforms and 12 pre-built plugins ^[13]. According to Engadget, the difference from the Mazaheri case is that these tools pull citations from external, verified sources instead of generating them from memory ^[15]. Google Cloud CEO Thomas Kurian said at launch that "ensuring every aspect of these agentic workflows is accurate, factual, and grounded in legal authority is of critical importance" ^[12]. Teams signing tend to assume the connector handles accuracy, so review can shrink. The only independent test of purpose-built legal tools is Stanford RegLab's 2024 study of platforms from LexisNexis and Westlaw. It found hallucination rates of 17 percent and 33 percent ^[19]. That works out to about one response in six from one platform and one in three from the other ^[1]. For Anthropic, firms have the lab's own claim of 90.9 percent on the BigLaw Bench legal reasoning benchmark for Opus 4.7 ^[14]. The figure is a vendor-reported reasoning score, not a measured rate of invented citations ^[14]. No comparable independent audit of Google's or Anthropic's legal tools has been done ^[20]. Users file what the tool hands them. Researcher Damien Charlotin's database counts more than 800 lawyers among its cases, alongside 1,100 self-represented litigants ^[4]. Mazaheri's filings came from a general Grok model, and the reporting does not link any sanction to Google's or Anthropic's legal products ^[6]. Reported Canadian cases involving fabricated citations went from seven in 2024 to 86 in 2025, with 39 more in the first quarter of 2026 ^[10]. At that quarterly pace, 2026 would finish near 156, about 1.8 times the 2025 count ^[2]. SpaceXAI's legal pitch has less behind it. A Cursor blog post said Grok 4.5 was suited for "finance, legal work, or anything else you do on a computer," but SpaceXAI's own announcement for the model focused on coding, agentic tasks and knowledge work and did not mention legal applications ^[17]. SpaceX acquired Cursor for $60 billion on August 15, and Cursor's legal marketing now sits under the Grok name ^[18]. I'd judge any of these tools on time to a verified citation: the minutes between a draft and a citation a lawyer has opened and confirmed. Two tests sort the options into a 2x2. The first is whether the tool links each citation to the primary source it came from. Google and Anthropic describe their products that way ^[15]. SpaceXAI's page does not disclose how it traces citations to primary sources ^[16]. The second test is whether the firm's own procedure requires a person to open that source before filing. Linked and checked is the one box where the product saves time, mostly on search. Linked but unchecked sends an error rate like Stanford's into the court record ^[19]. Unlinked with a required check means the lawyer redoes the research by hand. Unlinked and unchecked is the Mazaheri case again ^[6]. In my view, firms should pilot only in the linked-and-checked box and budget checking hours as if the error rate sat in Stanford's range until an independent audit says otherwise. The cost is that verification eats much of the time the tool saves, so an honest pilot may show a modest gain.