A Colombian-law benchmark posted to arXiv on October 2 found a failure that is easy to miss during a quick source check. In a judged subset covering 5,140 numeric legal citations from 15 models, 1.6% named a legal authority that did not exist. Another 45.2% named a real authority that did not support the model's claim.

The study used an automated rubric judge, then checked its citation flags against a blinded human sample. It tested Colombian law without internet access, retrieval, or external tools. The paper is also a preprint. Its percentages should not be treated as a universal error rate for every AI product or research task.

The useful distinction travels well: a source can be real while the claim attached to it is wrong.

That problem is harder to spot than an invented citation. A working link, familiar publication, or valid case number can feel like confirmation. It only confirms that the document exists. The American Bar Association made the same boundary explicit in Formal Opinion 512, warning that lawyers may violate their duty of competence if they rely on generative-AI output without an appropriate degree of independent verification or review.

Check the claim, not the decoration

When an AI answer matters, open the cited source and run three checks:

  1. Find the exact passage that is supposed to support the sentence.
  2. Read around it for limits, exceptions, and context the answer may have dropped.
  3. Check the date, jurisdiction, and status. A proposal, old rule, dissent, or study in a different setting may be real and still be the wrong support.

This habit is useful beyond legal research. It applies to health claims, financial explanations, schoolwork, policy summaries, and any answer that borrows authority from a link.

Treat the citation as a route to evidence. Follow it far enough to see whether the evidence and the sentence actually match.

Public sources

This article offers a source-checking habit, not legal advice.