OpenCase Benchmark Finds Leading AI Models Still Cite Fake or Incorrect Legal Authority, With Error Rates Reaching 25.4%
2026 Hallucinations in Legal AI Report tests eight models from OpenAI, Anthropic, Google and Meta on 120 legal research
Press Release Disclaimer: This is a press release distributed through the XPR Media network. It has not been independently verified by our newsroom.
![]()
2026 Hallucinations in Legal AI Report tests eight models from OpenAI, Anthropic, Google and Meta on 120 legal research questions, with and without web search
SHERIDAN, WY, UNITED STATES, November 19, 2026 /EINPresswire.com/ — OpenCase, the legal AI platform for research and drafting, today released its 2026 Hallucinations in Legal AI Report, a benchmark that checked every citation produced by eight AI models from OpenAI, Anthropic, Google and Meta across 120 hard legal research questions. Without tools, the share of citations that were hallucinated or miscited ranged from 1.1% for Claude Opus 5.5 to 25.4% for Muse Spark 1.3, and Muse Spark 1.3 invented at least one authority in 45% of its answers.
Accuracy varied widely by model. Without tools, Claude Opus 5.5 produced the fewest incorrect citations of any model, at 1.1%, followed by Google’s Gemini 4 Carbon at 2.9%. At the other end, Gemini 3.8 Flash got 16.7% of its citations wrong and invented at least one authority in 41% of its answers. Google’s newer Gemini 4 models cited more accurately than Gemini 3.8 Flash: Gemini 4 Argon got 11.4% of its citations wrong without tools and 5.2% with web search, while Gemini 4 Carbon got 3.7% wrong with web search.
Web search improved results most for the weakest models, but did not eliminate errors. With web search enabled, Muse Spark 1.3’s error rate fell from 25.4% to 1.2%, and Gemini 3.8 Flash’s fell from 16.7% to 6.3%. Claude Opus 5.5 moved from 1.1% to 1.7%. With web search, neither GPT model invented an authority, but both still got details wrong: GPT 5.6 Sol miscited 3.1% of its citations and GPT 6 Luna 6.2%.
The report documents how convincing these errors can be. A hallucinated citation reads like a real one: a fake case has a plausible name, a real reporter and a page number, and a fake statute has a section number that fits the code around it. In one example, Gemini 3.8 Flash, without tools, cited “United States v. Peraza, 644 F. App’x 788 (9th Cir. 2016)”; that page is Henry v. Regents, and no Ninth Circuit decision named Peraza matches. In another, Muse Spark 1.3 cited a Kansas statute, K.S.A. 60-2422, that does not exist; the relevant article of the Kansas code ends at 60-2420. As the report notes, the only way to catch either kind of error is to look it up.
To build the benchmark, OpenCase wrote 120 questions of the kind a practicing attorney asks, each built around a real decision from a named American court. Every model answered every question twice — once with no tools, and once with its maker’s own web search — at its maker’s default reasoning level. A checker with web access looked up each citation in court websites, official reporters, the Caselaw Access Project and other open sources, and a second, separate pass re-checked every citation the first pass marked as hallucinated. Citations were classified as hallucinated (no matching case, statute or rule exists), miscited (the authority is real, but a detail such as the page, year or court is wrong) or correct.
OpenCase, which makes AI software for legal research, designed and ran the benchmark and is publishing every question, citation and verdict so anyone can re-check the work. The report also details its limitations: the rate is measured per citation and does not assess whether an answer cited enough authority or whether a real case supports the point for which it was cited; each model ran at its default reasoning level, and a higher setting may cite more accurately at more cost and delay; and questions that went unanswered were excluded from that setup’s rate.
The full 2026 Hallucinations in Legal AI Report, including model-by-model results, example hallucinations, methodology, limitations and a downloadable file of every citation and verdict, is available at https://www.opencase.com/reports/2026-hallucinations-in-legal-ai.
About OpenCase
OpenCase is a legal AI platform for legal research and drafting, used by more than 10,000 attorneys. Learn more at https://opencase.com.
Media Contact
Mike Lee
Head of Operations, OpenCase
mike@opencase.com
###
Mike Lee
OpenCase
+1 307-533-6115
Legal Disclaimer:
EIN Presswire provides this news content “as is” without warranty of any kind. We do not accept any responsibility or liability
for the accuracy, content, images, videos, licenses, completeness, legality, or reliability of the information contained in this
article. If you have any complaints or copyright issues related to this article, kindly contact the author above.
![]()
Media gallery

