Hearsay at scale

Hearsay is one of those legal words that sounds like a technicality and is in fact a theory of knowledge. In ordinary courts it names a simple thing: a statement offered to prove its own truth, made by someone who is not present to be questioned about it. The rule against admitting such statements is old, much older than any modern court, and it has survived centuries of attempts to refine it away because it rests on something durable. The reliability of a claim is not a property of the claim. It is a property of the relationship between the claim and a source that can be examined. Remove the examinable source and you have not removed a formality. You have removed the only thing that let you tell a fact from a fluent sentence.

We are now generating fluent sentences at a volume the framers of any evidence rule could not have imagined. A system can produce, in the time it takes to read this paragraph, more confident assertions than a human witness could make in a lifetime. Each one arrives polished, grammatical, internally coherent, and entirely detached from any source you could put under oath. This is not a new category of problem. It is the oldest problem in evidence, arriving at industrial scale.

Call it hearsay at scale. A single machine claim, considered alone, is recognizably hearsay: a statement whose author cannot be cross-examined, offered to prove the thing it says. What changes when you produce a billion of them is not the nature of the claim but the cost of the instinct that used to protect us. When unsourced assertions were scarce and slow, skepticism was cheap; you could afford to stop and ask where each one came from. When they are infinite and instantaneous, that same skepticism is overwhelmed by volume. The defense erodes not because anyone defeated it but because nobody can apply it fast enough.

The rule against hearsay was never about distrusting words. It was about insisting that words point back to something a person could be made to answer for.

Why fluency is the wrong signal

The human mind treats fluency as a proxy for truth. A claim delivered smoothly, in the right register, with the cadence of expertise, reads as more credible than the same claim delivered haltingly — and for most of history that heuristic was roughly serviceable, because fluency was expensive. It took knowledge, or at least practice, to sound authoritative. Machine generation severs that link completely. Fluency is now the cheapest thing in the system, available in unlimited quantity at no marginal cost, decoupled from whether anything behind the sentence is so. The signal we evolved to trust is precisely the signal that has been industrialized first.

This is why the hearsay instinct matters more now, not less. It was never an instinct against words; it was an instinct against accepting words on the strength of how they sounded rather than what stood behind them. The rule forces a question that fluency is designed to suppress: who is the source, and can that source be examined? A machine answer that arrives without provenance fails this test in exactly the way an absent witness's secondhand report fails it. The eloquence is not evidence of reliability. Under these conditions it is barely even correlated with it.

From exclusion to provenance

The courts' answer to hearsay was never simply to discard everything secondhand. The doctrine is full of exceptions, and what those exceptions have in common is instructive. A statement becomes admissible when something about its circumstances restores the examinability that ordinary hearsay lacks — a record kept in the regular course of business, a declaration made under conditions that make fabrication unlikely, an admission against one's own interest. In each case the exception works by supplying a substitute for the missing cross-examination: a verifiable chain back to the conditions under which the statement was made. The deep principle is not exclusion. It is that a claim earns admissibility by carrying its provenance with it.

That is the move available to us now, and it is the only one that scales. We cannot examine a billion sources one at a time, but we do not have to, any more than a court examines every entry in a ledger. What we can require is that a consequential claim arrive attached to the record of where it came from — the inputs that produced it, the evidence actually consulted, the state that would let someone reproduce it. A claim that carries that provenance is no longer hearsay; it has, in effect, brought its witness with it. A Decision Receipt is one name for that demand made concrete: not a request that you trust the output, but a structure that lets the output be examined the way an examinable source could be. Provenance is what converts an assertion into something you can stand on, or knock down.

The framing that helps me most is to stop asking whether a given machine output is true and start asking whether it is admissible — whether it carries what it would take to be tested by someone who did not want to believe it. Truth is a property the sentence claims for itself. Admissibility is a property the record either has or lacks, visible from the outside, gradeable without trusting the speaker. The first is what fluency counterfeits. The second is what fluency cannot fake, because it is not made of words.

None of this requires deciding that machine output is worthless. The point is narrower and older than that. A statement detached from any examinable source has always been hearsay, and producing it faster does not make it more reliable; it only makes the old discipline harder to keep and more necessary to insist on. The world is about to be saturated with confident, sourceless sentences. The centuries-long habit of refusing to let an absent witness decide a question is not a relic to be apologized for. It is the one piece of inherited wisdom that was built, almost prophetically, for exactly this.

— Dispatches · Summit Cognitive