Turns out feeding a law firm's least favorite rulebook to a chatbot does not make the rules disappear.
Researchers built a benchmark of 2,058 Bluebook citation queries and tested frontier language models on formatting them correctly from scratch. The average model produced a fully compliant citation only 42.6% of the time in a zero-shot setting. When a "reasoning" model went up against real law students competing for editor slots at five top law reviews, it scored well below the humans. Simply handing the model the citation rules as reference text barely moved the needle, which is a problem for anyone betting that retrieval-augmented generation alone can enforce strict rule-following.
That's a specific, testable limit on a much-hyped promise: automating the "drudgery" of legal work. Citation formatting is exactly the kind of repetitive, rule-bound task AI was supposed to eat for breakfast, and on its own it mostly can't. The fix that worked wasn't a bigger model - it was a smaller one paired with an old-fashioned rule engine. Using a model to parse citations into structured elements, then handing formatting to deterministic code, raised accuracy by 32.4 percentage points, to as high as 85.5%.
Any legal-tech vendor selling a pure "AI-powered" citation checker might want to explain what's doing the real work under the hood.