The failure mode, mechanically
A large language model does not look up case law when it generates a citation. It predicts what a citation-shaped string of text should look like, based on patterns in whatever it was trained on — and sometimes that prediction is fluent, correctly formatted, and refers to nothing that was ever decided. This isn't a bug that gets patched. It's the base mechanism of how these models generate text, applied to a context — legal citation — where "plausible-sounding" and "true" need to be the same thing and usually aren't guaranteed to be.
Why it's so easy to miss
The citation looks completely normal. Party names that sound like real parties. A reporter, volume, and page number in the right format. The right court, a year that's in range. Nothing about its shape signals that it's wrong — a fabricated citation and a real one are visually and structurally identical. The only way to know the difference is to actually pull the case, which is precisely the step that gets skipped when the citation already looks right and the deadline is close.
What it costs when it gets through
Sanctions exposure. Malpractice risk. A credibility problem with the court and with opposing counsel that doesn't reset after the first instance. And underneath all of that, the more basic cost: the lawyer signed something they didn't actually verify, which was the one job the citation was supposed to prove had been done.
Why "enforced citation" is the actual fix
The fix isn't "don't use AI for legal work." It's a system that cannot produce an answer without a traceable source attached to it — one that searches the firm's actual knowledge base and case files, and returns a citation only when it can point to where that citation came from. That's the difference between a model that's asked to sound right and a system that's structurally unable to make something up: the second one either finds the source or says it can't find one. It never fills the gap with something plausible.
That's what a litigation knowledge engine is built to do — dynamic search across a firm's own knowledge, with enforced citation, so every answer traces back to a source before anyone reads it.