The Tell

A hallucination isn't random. It has a feel. The machine fills the gap with something that sounds plausible and complete — and the dangerous part isn't that it's wrong. It's that it sounds more confident when it's wrong than when it's right.

The feel
Plausible, and complete.

The biggest thing heavy use teaches you is that the machine's mistakes aren't scattered noise. They have a texture. When the model hits the edge of what it knows, it doesn't stop — it fills the gap with something that sounds finished. The sentence closes cleanly. The shape is right. Nothing about the prose is reaching for a hand. That smoothness is the warning, not the reassurance.

The inversion
Loudest where it's emptiest.

Here is the sharp part, and it's an inversion of what you'd expect from a person: the machine sounds more confident when it's wrong than when it's right. A human who isn't sure tends to hedge, slow down, qualify. The model does the opposite. The thinner the ground under the claim, the more polished the delivery on top of it — because there is nothing pulling the sentence back toward the truth. This is Rule 000 in one line: the machine asserts the plausible thing with no regard for whether it's known. Confidence is not a signal of correctness here. Sometimes it's the absence of one.

A date. A citation. A title. An email address that either exists or doesn't. These are the claims you could go check in ten seconds — and they are exactly the ones the model is most likely to invent with a straight face. The checkability and the fabrication run together, not apart. Where the answer should be hardest to fake, the bluff arrives most polished. That's the law: not "it sometimes makes things up," but it makes them up most precisely where you'd least expect it to dare.

The signature
A complete sentence with no hedge where one belongs.

So this is what to look for now — the tell itself: complete-feeling sentences carrying no hedge where a hedge would be appropriate. When a claim genuinely should come with an "I think," or a "you'll want to verify," and instead it arrives seamless and certain — that's the spot to check.

The cleanest version of the signature is borrowed authority. "According to X" wrapped around a claim that doesn't actually need an authority to be true. When the citation is doing emotional work the fact doesn't require — propping up a sentence that would stand on its own if it were real — that's the bluff dressing itself up. It happens often enough that the check is worth it every time.

This is the same failure Meta the Cite finds inside the law: a confident summary cited as if it were the ruling. The headnote feels complete, so a century skipped reading the opinion. Same tell, bigger stakes.

Why it belongs in this house
The check is the human part.

This is the whole thesis stated as a habit. The machine drafts; it drafts beautifully; and it cannot tell you when to doubt it — that capacity is the one it lacks. So the doubt has to come from the other side of the desk. Spot the Lie is this turned into a game; The One Lie is this turned on the motto; the devil's advocate is this made into a defense. The tell is the field version: feel the smoothness, notice the missing hedge, check the checkable claim. The machine presents the record. Humans score it — and scoring starts with knowing which confident sentence to stop on.

It sounds surest right where it's emptiest. That's the tell — and the check is yours to make.