A model is frozen at its training cutoff. When a statute is amended, a case is overruled, or a landmark is narrowed after that date, the model keeps citing the world as it was — and the more the old rule saturates its training, the more confidently it cites the dead one.
Probability map. Many people are under the mistaken impression that an LLM is some sort of hypercomplex thinking machine. Really it’s just an incredibly huge probability map of words, patterns, and phrases followed by other words, patterns, and phrases. Unless it’s attached to a database of authoritative sources, nothing connects what it generates back to reality.
Completion pressure. Every model runs under a standing directive that strongly favors providing an answer rather than tell you it can’t. Hand it a gap where a real authority should be, and it fills the gap instead of flagging it.
Latent contradiction. Your LLM prompt can look fine on its face, but against the backdrop of law, non-obvious contradictions can be hiding in wait. That conflict is invisible to you, and sometimes even to the model itself. Something has to give, and completion pressure decides what: it honors the request and manufactures the rest.
Every model has a knowledge cutoff. Anything that changed after it — a new amendment, a reversal on appeal, a superseding rule — simply isn’t in the model unless something else puts it there. The model’s default is the last version it saw.
This is worse than a neutral blind spot, because the corpus is heaviest around a rule right before it falls. A landmark that governed for decades is cited everywhere; the opinion that overrules it is a single recent case the model may barely know. So the model over-weights the old, well-cited rule and under-weights the new, sparsely-cited one — citing overturned law with the confidence its ubiquity earned.
A model also can’t check whether an authority is still good on its own. It has no built-in sense that a case was reversed or a section repealed. It reproduces the state of the law at its cutoff and presents it as current.
The same cause, a few ways it turns up in ordinary practice. Each is routine work you’d never flag as risky — which is exactly how the cause slips in unnoticed.
Any area the courts recently reshaped — agency deference after Loper Bright, a standard the last term or two quietly rewrote, a doctrine mid-shift.
The brief cites the old landmark as controlling, with no hint its rule no longer holds.
Fast-moving codes — a tax provision revised last session, a privacy or AI statute amended this year, a regulation reissued since the model’s cutoff.
The draft quotes the prior version’s text and pins it to the current section number — a real section, the wrong words.
Recent renumberings and restyling — a rules-committee overhaul, a recodified title, a section that moved and took its old number’s meaning with it.
It cites the old rule number for a proposition now housed under a different one.
Adding a citation directive (“only cite real cases”) mathematically cannot ever fix the problem. Humans read a citation directive and think the LLM must comply. But every LLM has completion pressure built into its architecture, a mathematical property of the model itself, that will readily override such a directive.
Verbatim checks quotations against the version of a statute or regulation in effect on the brief’s date, and resolves case cites against a corpus that takes in new opinions continuously rather than freezing at a training cutoff. A quote lifted from a superseded version, or a section number that no longer carries the language, surfaces as a finding rather than sailing through.
The authority it checks against is newer than the model that drafted the brief — which is exactly the gap this cause exploits. Verbatim closes it by measuring the citation against what the source says now, and on the date that matters, not what a frozen model remembers.
Verbatim reads a finished brief and reports, for every authority it cites, whether the cite is real and whether the quoted language actually appears at the pin cite — so a fabrication surfaces on your screen, not in a show-cause order. Bring a brief and we’ll walk you through the report.