Overview
Last week, this column covered courts that have stopped warning lawyers about AI hallucinations and started suspending them and throwing them off cases. The lesson was that the duty to verify a citation runs to the lawyer who signs the filing, not to the tool that drafted it. Days ago, Thomson Reuters made the same point from the vendor side. It opened early access to a ground-up rebuild of CoCounsel, its legal AI assistant, and made verification the selling point. General availability is planned for later this summer.
Thomson Reuters calls the result “fiduciary-grade” AI. By its own definition, that means every output is grounded in authoritative, verified content from the first step of the reasoning, every citation is traceable to a linked source the lawyer can open and read, and the reasoning is visible enough to inspect rather than just approve.
That design helps. The lawyer still answers for the result. A link shows that a source exists. It does not show that the source supports the sentence, is still good law, and fits the record. The point of a linked source is to click it and read it. As the Ninth Circuit put it this month in LNU v. Blanche, a competent and diligent attorney “must also read and reason.”
None of this requires a dedicated legal-AI product. Off-the-shelf frontier models are capable legal researchers on their own. Most statutes, rules, regulations, and reported decisions are available for free somewhere outside Westlaw, and models with browsing or retrieval access can often find those sources and pull them together.
The discipline is in what you demand next. Make the model retrieve the full text of every source it cites, then have it check every quotation, citation, and representation against those documents. Hand that verification to a second model, and the review becomes genuinely adversarial. Given the source text and an instruction to check everything against it, these models are good at catching their own errors and stripping out hallucinations.
Without that source text and instruction, the same models can do the opposite. In Noland v. Land of the Free, a lawyer used ChatGPT to enhance his brief, then ran it through Claude, Gemini, and Grok to check for errors. The fabrications survived. A California Court of Appeal sanctioned him $10,000 after finding that 21 of 23 case quotations were invented.
CoCounsel promises source-grounded checking automatically, so you never have to build that workflow yourself. That convenience is the pitch. The tradeoff is that a general-purpose frontier model may be smarter, or allowed to work harder, than the model tier CoCounsel runs under the hood. And what runs under the hood is the vendor’s call, not yours.
CoCounsel began on OpenAI’s GPT-4 and now runs on Anthropic’s Claude. The debut installment, on Claude Fable 5, warned against letting anything critical ride on a single tool or model. A tool can switch models, as CoCounsel did, and a model can vanish by government order. Keep a backup that does not depend on the same model or vendor.
With the new CoCounsel, patent litigators may have a better tool. The lawyer’s duty to check its work stays exactly where it was.
The Weekly Inference is a recurring Step Into IP feature. Nothing here is legal advice.
