- Source
- arXiv
- Published
- Runtime
- 0:00
- Snippets
- 4
§02
Snippets
-
LLMs fail to detect citations pointing to real cases that don't support the claimed proposition, catching only 37-61% of wrong-pinpoint errors in court opinions.
This failure mode bypasses standard citation-checking tools and could mislead courts with plausible-sounding but unsupported legal authority.
-
Models confuse topical relevance with page-level support: they recognize a case discusses the right legal topic but fail to verify the proposition appears at the cited pinpoint.
Topic matching is not enough for legal reasoning; precision down to specific pages within cases is needed to validate arguments.
-
Even GPT-4 with extended reasoning and page-verification prompts still misses 40% of pinpoint mismatches in court opinions and 18% in briefs.
Brute-force prompting alone cannot solve proposition-level citation verification; the gap suggests a deeper architectural limitation in current models.
-
Prompting models to verify page-level support improves detection of wrong pinpoints but increases false positives, creating a precision-recall tradeoff.
There is no simple prompt that solves the problem; legal citation verification requires more robust checking than current models can reliably provide.
§03
Synthesis
The Problem: Citations That Sound Right But Aren't
When a lawyer cites a case to support a legal argument, the case must actually say what the lawyer claims it says—and ideally at the specific page cited. The 2023 Mata v. Avianca debacle involved fabricated cases outright, which are easy to catch with a database lookup. But here's the harder failure mode: what if a cited case is real, exists in the legal database, and discusses related topics, but doesn't actually support the specific proposition on the page the lawyer pointed to? This paper shows that large language models (LLMs) routinely fail at this verification task, and the research community hasn't been measuring it.
How the Study Works
The authors took real citations from two legal corpora (court opinions and legal briefs) and deliberately corrupted them in two ways:
- Wrong-case corruption: Replace the cited case with a different real case on a similar topic.
- Wrong-pinpoint corruption: Keep the same case but change the page number to a different page within that case.
The first corruption tests whether a model truly knows which case supports a claim. The second tests whether a model can verify that the specific page actually contains the evidence—a more granular check that mimics real citation mistakes.
They then asked fourteen different model configurations (varying by model size, reasoning capability, and prompting strategy) to determine whether corrupted citations were valid.
The Results: Scale Helps, But Not Enough
Models performed well on wrong-case corruptions, catching 93–100% of them. This suggests they can recognize topical fit reasonably well.
On wrong-pinpoint corruptions—the harder task—performance collapsed: 37–61% catch rate on court opinions, 52–83% on legal briefs. Even GPT-4 with extended reasoning effort missed 40% of pinpoint mismatches in opinions and 18% in briefs. When models failed, they justified their acceptance by invoking topical overlap: "Yes, this case discusses contract law, and your proposition is about contracts, so the citation looks good."
A simple fix—explicitly prompting models to verify support at the cited page—improved recall. But this came at a cost: false positives rose sharply. Models began incorrectly validating more corrupted citations, trading one error for another.
Why It Matters
The research reveals a fundamental gap between two distinct skills. Current LLMs can recognize that a case falls within the right legal domain, but they cannot reliably verify that a particular passage supports a particular claim. For legal work, this distinction is critical: topical relevance and propositional support are not the same thing.
This finding has immediate practical consequences. Lawyers using LLMs for legal research or brief drafting face a subtle risk: the model will confidently cite cases that sound topically appropriate but don't actually support the argument at the cited page. Database lookups won't catch this. Only human review of the actual cited material will.
The paper argues that evaluations of LLMs for legal tasks need to include this failure mode. Current benchmarks largely overlook it, leaving practitioners and regulators with an incomplete picture of where these tools remain unreliable.
Mine your own.
Lode is a workbench, not a feed. Paste a YouTube URL. The model proposes a transcript, a set of quote-grounded snippets, a synthesis essay, and the fan-out. You decide what stays.