What replaces "is it true"

For most of human history, "is it true" was a question you could actually answer, and the way you answered it was rarely by independently establishing the truth. You answered it by knowing who said the thing, by reading how confident they seemed when they said it, and by estimating how much trouble it would be to check. Those three signals — the identity of the speaker, the texture of their confidence, the cost of verification — did most of the work. They were proxies, and most of the time they were good enough proxies, because they were correlated with the thing you actually cared about. A person who knew what they were talking about usually sounded like it, usually had a name and a reputation attached to it, and usually could not produce fluent detail at scale without having done the underlying work. The shortcut held because competence and its outward signs traveled together.

This was never a perfect arrangement, and the history of fraud is the history of its failures. But it was a workable one, and it was workable for a specific structural reason: producing a claim that looked authoritative was roughly as expensive as producing one that was authoritative. Confidence was costly to fake. Detail was costly to fabricate. A name was costly to stake. The friction that made knowledge hard to acquire also made counterfeit knowledge hard to manufacture, and that symmetry is what let us get away with judging truth by its outward signs. We were not really asking "is this true." We were asking "does this carry the marks that true things tend to carry," and trusting the correlation to hold.

The settlement that machine-made decisions break

That correlation is exactly what machine-made output severs. A system can now produce fluent, confident, internally consistent claims at a volume and a polish that no human source could match, and it can do so with no relationship between how sure the output sounds and how sound it is. Confidence has come unbundled from competence. The texture we used to read as evidence of underlying work is now available for free, in unlimited supply, attached to nothing. The cost of looking authoritative has collapsed toward zero while the cost of being authoritative has not moved at all — and the moment those two costs diverge, every shortcut built on their symmetry stops working.

Worse, the third proxy fails alongside the first two. There is no maker who can be held to a machine-made claim in the way a person can be held to theirs. When a colleague tells you something and it turns out to be wrong, you can go back to them; they own the assertion, they can be asked to justify it, they bear some cost for having made it. A generated output owns nothing. It cannot be summoned, cannot be cross-examined, cannot be made to account for the gap between what it said and what was so. The claim arrives detached from any party who can be put under obligation for it. So all three of the old signals fail at once — the speaker dissolves, the confidence decouples, and the volume makes checking everything by hand impossible. The question "is it true" still matters as much as it ever did. We have simply lost the apparatus we used to answer it.

The cost of looking authoritative has collapsed to nothing. The cost of being authoritative has not moved at all. Every shortcut we built on their symmetry is now broken.

The temptation, faced with this, is to try to rebuild the old apparatus — to demand that machines sound appropriately uncertain, to bolt confidence scores onto their outputs, to insist that some named human "sign off" on what the system produced. Each of these is an attempt to restore a proxy that has already failed. A calibrated confidence number is still just a number the system emits about itself; it inherits the system's blind spots. A human signature on a decision the human did not actually make is a name staked on nothing. We keep trying to make the output more trustworthy when the output was never where the trust lived. The trust lived in the conditions of production — and those are what we lost sight of.

Admissibility, not truth, is the successor question

Here is the move I want to propose. Stop asking, of a machine-made decision, whether its output is true. Start asking whether it would be admissible — and on what record. Truth is a property of the claim in isolation, and in isolation a generated claim gives you nothing to grip. Admissibility is a property of the claim plus the record it comes with: not "is this assertion correct" but "is this a thing I am entitled to rely on, given how it was produced and what it can show about its own production." The word is borrowed deliberately. Courts long ago abandoned the fantasy that they could directly verify the truth of everything put before them. They cannot rerun the crime. So they built, over centuries, a discipline for deciding which claims are even allowed to count — rules about where evidence came from, who is offering it, whether it can be tested by the other side. Admissibility is not a lower standard than truth. It is the standard a system adopts once it admits it cannot establish truth by inspection and must instead govern the conditions under which a claim earns the right to be relied upon.

That is precisely our situation with machine-made authority, and it is why admissibility is the right successor and not merely an analogy. We cannot inspect a generated claim and see its truth; the inspection that used to work — reading the confidence, knowing the speaker — has been defeated. So the question has to move from the claim to the record around it. A claim is admissible when it arrives with enough of the right record to be relied upon and, crucially, to be contested. That record, I think, rests on four properties, and the rest of this is an attempt to name them precisely.

The first is provenance: an account of what the decision actually drew on. Not a gesture at "relevant data" but the specific inputs that were in front of the process at the moment it ran — the sources, the documents, the prior state — recorded as they were, not summarized after the fact. Provenance is what replaces "I know who said this." You can no longer trust a speaker, so you trace the lineage instead. A claim whose provenance is unknown is not a humble claim awaiting verification; it is a claim that has structurally opted out of being checked, and that should lower your confidence rather than leave it untouched.

The second is replay: the ability to re-run the decision against what was known at the time and watch where it lands. This is what replaces "I can tell they sound like they did the work." You no longer trust the texture of confidence, so you reconstruct the process instead. Replay is the difference between a record that tells you a story about how it reached its conclusion and a record that lets you reproduce the conclusion yourself, without taking the system's word for anything. An explanation is the system's opinion about its own reasoning. A replay is a fact you can establish independently. When the replay reproduces the outcome, the record has earned standing; when it diverges, the divergence is itself the finding, and it tells you exactly what was wrong.

The third is standing: a clear answer to who may demand the record and contest what it shows. This is what replaces the maker who could be held to the claim. The generated output owns nothing and can be summoned by no one — so admissibility has to specify, in advance, which parties have the right to pull the record, to challenge it, to make it answer. A record that exists but that no one is entitled to demand is not accountability; it is a sealed file. Standing is the property that turns a record from a private artifact into a public obligation, and without it the other three properties are inert.

The fourth is contestability: the record has to be built, from the start, to be fought. This is the property that binds the other three together and gives them their point. Provenance, replay, and standing are not ends in themselves; they exist so that a decision can be argued against by someone with reason to. A record assembled to reassure is a different object from a record assembled to be refuted, even when the two look identical on the surface. The first closes the question. The second exposes its own inputs and its own reproducibility precisely so that a motivated party can try to knock it over — and either fails, in which case the record has earned its authority, or succeeds, in which case the record has done its job by surfacing the error. This is the heart of what we call a Decision Receipt: not a summary that asks to be believed, but a record built to survive contact with someone who disagrees.

What we are deciding by default

Put those four together and admissibility stops being abstract. A machine-made decision is admissible when you can see what it drew on, re-run it against what was known, identify who is entitled to demand that re-run, and contest the result on a record built for the fight. None of these asks the impossible. None of them requires that the decision be correct — a wrong decision can still be admissible, and that is a feature, because an admissible wrong decision can be caught, while an unaccountable right one cannot be trusted and an unaccountable wrong one cannot be found. Admissibility is not a promise that the machine got it right. It is the precondition for ever being able to tell.

I want to be plain about the stakes, because they are larger than any single deployment. We are, right now, setting the standard for what machine-made authority will be allowed to get away with — and we are mostly setting it by default, by accepting outputs that carry no provenance, that cannot be replayed, that no one has standing to contest, that were never built to be fought. Every time a consequential decision is handed down by a system and accepted because it sounded sure and arrived at scale, we are ratifying a standard. The standard being ratified is that fluency is sufficient, that confidence substitutes for accountability, that a claim with no maker behind it and no record beneath it can nonetheless govern people who had no way to push back. That is the settlement forming around us, and it is forming whether or not anyone chooses it.

The old question, "is it true," assumed a world in which truth left visible marks we could read. That world is gone for machine-made claims, and no amount of wishing restores it. But the failure of the old proxy is not the failure of the underlying demand. We still need to know what we may rely on. The successor standard is admissibility — a claim plus a record built to be provenance-traced, replayed, demanded by someone with standing, and fought. The work in front of us is to make that standard the default before its absence becomes the default, because one of the two is going to set, and only one of them leaves us able to ask the question that started all of this and still get an answer worth having.

— Dispatches · Summit Cognitive