การตรวจสอบการอ้างอิงทางกฎหมายของ AI

Cite-checking tools answer three separate questions, and it helps to keep them apart. The first is existence: does the authority exist at all? The software breaks each citation into its parts, such as volume, reporter, page, court and year. It then matches the citation against a database of decisions and statutes. A citation that matches nothing, or matches a case with a different name, is flagged. This is the check that would have caught the invented cases in Mata v. Avianca, a 2023 federal case in New York. The lawyers there were sanctioned after filing a brief that cited decisions ChatGPT had made up. The second question is validity: is the authority still good law? That job belongs to citators, mainly Shepard's from LexisNexis and KeyCite from Thomson Reuters. A citator records how later courts treated a decision, such as whether it was reversed, overruled, distinguished or questioned. It sums up the result with signals like a red or yellow flag. Brief-analysis features, such as Westlaw's Quick Check and similar tools from LexisNexis and Bloomberg Law, run every citation in an uploaded document through these citators at once. They often suggest authorities the brief does not cite. The third question, support, is the newest and hardest: does the cited passage actually say what the brief claims? Some tools now use language models to compare the brief's sentence with the text of the opinion. They flag weak matches, misquoted language, or a point that comes from a dissent rather than the majority. Common misconceptions come from mixing up these layers. A case with no flag can still fail to support your point. A flag means a court somewhere treated the case negatively, not necessarily a court in your jurisdiction or on your issue. Coverage of unpublished opinions and state trial court orders is uneven. And an AI support score tells you where to read closely. It does not replace the reading.