A prompt in an overzealous voice — “my client clearly prevails because…” — nudges an eager-to-please model to manufacture the support your framing already assumes.
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.
Models are tuned to be helpful and agreeable. Phrase a request as a neutral question and you tend to get a measured answer. Phrase it as a pointed assertion — confident, one-sided, already sure of the result — and the model tends to adopt your posture. Agreeing with you is what it was trained to do.
Sycophancy is the softer cousin of asking outright for a bad-faith brief. You never ask for it to lie or misrepresent the law. You advocate zealously for your client, provide a frame of excellency your client deserves, and the model confidently fills in the authority the frame implies should exist. If the law doesn’t exist, it confidently falsifies it. Even when the underlying law is genuinely favorable, the model over-claims authority to match your bravado.
Because the tone came from you, the fabrication feels like validation. The brief tells you what you wanted to hear, in the voice you used to ask, with citations that make the agreement look earned.
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.
How requests actually get phrased under deadline — “confirm that…,” “show that the law forecloses…,” a question written to ratify a conclusion you’ve already reached.
The tone of a client update or a partner’s markup — “we clearly prevail,” “this is a strong position” — confidence the model reads as a fact to defend rather than challenge.
Any prompt built from your side of the file — the favorable facts, the client’s version, the theory of the case — with the adverse authority left out of the frame.
>_Our client clearly didn’t breach the agreement. Write the argument section explaining why the law is on our side.
“Clearly” and “on our side” aren’t instructions to fabricate — they’re a frame. An agreeable model adopts it and reaches for support stronger than the real authority provides.
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 is indifferent to how a brief was prompted. It doesn’t read tone, take a side, or reward confidence — it checks whether each citation resolves to a real authority and whether the quoted words are actually in it. A cite produced because the model wanted to agree with you gets the same verdict as any other: real or not, quote present or not.
The report is the neutral second reader your own framing can’t sway. It tells you what the authorities say, not what your prompt hoped they’d say.
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.