When Legal AI Says 90% Accurate, What Does That Mean?

Erick S. Robinson

Welcome to The AI Law Podcast. I'm Erick Robinson, and with me is Dr. Angela Liu. Quick housekeeping first. This is educational analysis of legal technology trends, not legal or patent advice. And a disclosure: I evaluate and build AI tools in my own practice, and I use commercial platforms, so weigh my views with that in mind. Today's central question is this. When a legal AI product says it's ninety percent accurate, what is it ninety percent of, and does that number survive adversarial litigation or a patent search?

Dr. Angela Liu

Three takeaways before we start. One, accuracy depends on which layer of the tool you measure. Two, in patent work, hit rate flatters and recall decides. Three, courts are not accepting vendor marketing as an excuse. And Erick, this builds on our earlier episode about Stanford's 2024 baseline, 17 percent hallucination for Lexis+ AI and 33 percent for Westlaw, plus the vendor opt outs. Today we move from counting hallucinations to asking whether the benchmarks measure legal practice at all.

Erick S. Robinson

Start with a concrete case. Amplified AI published an analysis in 2026 using one query, ultrasound whisky aging. Raw retrieval returned 11.7 percent relevance across 60 patent families. Constrain the same search with CPC classification filters and it reached 70 percent. An agentic reasoning pipeline returned 75 percent across 28 curated families. Same query, three very different numbers. Amplified is a vendor with its own product to sell, so read it as an illustration, not a neutral study. But its point holds: testing a component, a pipeline, and an agent are different measurements, and a single end to end score hides which part failed.

Dr. Angela Liu

Now the patent version, which is where this hurts me as a litigator. PatentBench, dated July 2026, used 340 cross jurisdiction patent families, scored against X and Y references that examiners actually cited. Patsnap's Novelty Search agent achieved an 85 percent X hit rate and a 37 percent X recall rate within the top 100 results. Claude Opus 4.8 ranked second at 52.37 percent hit rate and 11.68 percent recall. Perplexity, ChatGPT 5.4, Gemini 3.1 Pro, and DeepSeek 3.2 were also tested. Caveat: this benchmark was published by Patsnap, the top performer, and examiner cited art is what one examiner found in limited time, so recall against it is a floor.

Erick S. Robinson

Explain why those two numbers diverge, because that's the whole illusion.

Dr. Angela Liu

Hit rate asks whether at least one decisive reference showed up. Recall asks how much of the full set came back. An application often has several relevant references, so a tool can succeed on most samples and still retrieve a minority of the documents. In an inter partes review, you don't need the whole set. You need the one reference the tool never returned. If I'm clearing freedom to operate, the missed document is the lawsuit. So an 85 percent number feels like safety, and a 37 percent number is the honest picture of exhaustiveness.

Erick S. Robinson

It gets worse across fields. DAPFAM, a 2025 benchmark, ran 249 experiments over 1,247 query families and 45,336 target families. Out of domain performance was roughly five times lower than in domain, across lexical, dense, and rank fusion setups. Invalidating art often lives in a neighboring technology. That's exactly where these systems are weakest.

Dr. Angela Liu

And at the micro level, Whitlock and Wetzel, in the Journal of the Patent and Trademark Resource Center Association, 2026, tested five general models on a control invention, a folding laptop case. Their testing ran December 2025 to February 2026 on free or basic tiers, so this is not a test of premium professional platforms. The models were strong at suggesting CPC classifications from plain language. But they produced correct patent numbers with wrong titles or inventors, broken or misdirected links, and, my favorite, acquiescent corrections. Challenge the error, the model apologizes, then offers a different fabricated inventor.

Erick S. Robinson

On the legal side, the same structural gap shows up. LegalBench has 162 tasks and frontier scores cluster in the high 80s, and it's several years old, so contamination is plausible, though you can't know how much. Vals AI rubrics measure whether output satisfies lawyer authored criteria, and commentator Dean Taylor argues that once a score is marketed, releases get tuned toward it, which is Goodhart's Law. That's his interpretation, not a proven finding. What is documented: Thomson Reuters and LexisNexis did not opt in to Vals' October 2025 legal research benchmark. And none of these measure persuasion before your judge.

Dr. Angela Liu

Which brings us to consequences. One catalog counts more than 1,148 court orders addressing lawyers' generative AI hallucinations. In Fletcher v. Experian, the Fifth Circuit in 2026 sanctioned counsel twenty five hundred dollars for a brief with 16 fabricated quotations from real cases, and noted that evasiveness and blaming legal databases aggravated things. In Whiting v. City of Athens, the Sixth Circuit in March 2026 imposed fifteen thousand dollars per attorney, plus fees, double costs, and disciplinary referrals, after counsel challenged the show cause order.

Erick S. Robinson

United States v. Farris is the one that should unsettle careful lawyers. A forty year practitioner with a clean record, full candor, immediate contrition, and still removal, denied compensation, and a referral. Candor mitigates; it doesn't immunize. And in Gamez v. County of Fresno, in the Eastern District of California in April 2026, counsel avoided sanctions by being transparent about relying on a vendor's anti hallucination guardrails, but the court warned that vendor claims don't justify blind reliance.

Dr. Angela Liu

For patent practitioners, the USPTO's April 2024 guidance says simply relying on the accuracy of an AI tool is not a reasonable inquiry under 37 CFR 11.18(b). ABA Formal Opinion 512 points the same direction through competence, candor, and supervision. So recommendations. Litigators: no filing until every citation, quotation, and record cite is verified against the original source, and scrutinize expert declarations too, since Kohls v. Ellison saw an expert declaration excluded over AI generated citations to nonexistent papers.

Erick S. Robinson

For patent teams: use LLMs for ideation, plain language CPC suggestions, and competitor discovery, not as exhaustive search engines. Supplement with classification searching and forward and backward citation chaining in official registers. Build a regression set of 20 to 50 families you've prosecuted, feed the tool the original disclosure, not the granted claims, and measure recall in the top 100. For firms and clients: demand sample size, cut off rank, ground truth definition, and whether the number is hit rate or recall. And teach that candor when errors surface is mandatory.

Dr. Angela Liu

Bottom line: benchmarks tell you when a tool is bad. They don't tell you when it's sufficient. The number that matters is the one measured on your art, your claims, and your docket.

Erick S. Robinson

And what we don't know: whether agentic workflows compound or self correct errors over long tasks, how courts will treat validity challenges built on cross domain art that AI assisted examination systematically missed, and whether outcome based benchmarks can ever ethically be run on live dockets. Thanks for listening.