The inheritance of a record

We know the Roman Empire in part because it was a bureaucracy that could not stop writing things down. Grain receipts, land surveys, the census returns of a province, the discharge papers of a soldier, the complaint of a farmer about a neighbor's cattle — these survive, and through them a world that ended sixteen centuries ago can still be reconstructed, argued over, held to a kind of account. The records were never made for us. They were made for the ordinary business of a working state. And yet they became an inheritance, because the practice of keeping them outlasted the reasons for keeping any one of them. What a civilization leaves behind is not mainly its monuments. It is its paperwork.

I want to put a question to our own era that we mostly manage not to ask. We are, at this moment, delegating an enormous and growing share of consequential decisions to machines — who is approved, who is flagged, who is admitted, who is denied, at a volume no prior generation could have processed by hand. That is a fact about the present. The question is about the future: when someone fifty or a hundred years from now wants to understand how this transition actually happened — not the press releases about it, but the substance of it, the millions of small verdicts that added up to how people were treated — will there be anything for them to read? Or will they find, where the record should be, a silence?

Because a decision made by a machine and then discarded is not merely unaccountable now. It is unrecoverable later. The person who wanted to contest it in the moment loses their standing today; the historian who wants to understand it loses their evidence forever. And these are the same loss, seen from two distances. Contestability and history are the near and far ends of the same requirement, which is simply that the decision left a trace that outlives the moment of deciding.

We are writing the history of the machine age in real time, and the default setting is to throw it away.

What the future will not be able to ask

Consider what it means to be legible to those who come after. It does not mean they approve of what we did — the Roman records do not flatter Rome. It means they can find out. They can locate a specific decision, see the evidence it rested on, read the rule it claimed to follow, and judge whether it held. Legibility is not the same as legitimacy, but it is the precondition of ever assessing legitimacy at all. An era you cannot read is an era you cannot fairly judge, which means it is also an era that has escaped judgment — not by being righteous, but by being blank.

The machines of this period are, by construction, poor keepers of their own history. They optimize, they update their weights, they overwrite the state that produced yesterday's output with the state that will produce tomorrow's. Left to themselves they retain almost nothing of why a particular decision came out the way it did, because retaining it was never their objective. This is not a flaw to be scolded; it is a property to be corrected from the outside, deliberately, by building the record as a separate and durable thing. A Decision Receipt that carries the evidence actually consulted, the rule as it stood at the time, and enough state to replay the decision is, whatever else it is, a message to the future. It says: here is what happened, and here is how you can check that I have told you the truth.

The alternative is not neutral. A generation that decides at machine scale and keeps nothing is not simply failing to write its history. It is making a choice — a choice to be, to its descendants, a set of outcomes with no discoverable reasons, a period in which enormous things were done to people and no one can now say on what basis. That is a strange thing to bequeath. We would think poorly of an institution that shredded its files the moment each decision cleared. We should think no better of an age that does the same at scale and calls it efficiency.

The inheritance is a practice, not an archive

What survives across the centuries is never a single document. It is a habit that produced documents reliably enough that some of them lasted. The Roman clerk did not know which of his receipts would matter; he kept them all because keeping them was the discipline of the office. The inheritance we can actually leave is not a curated archive of the important cases — those we would choose to remember are precisely the ones least in need of independent memory. It is the boring, universal practice of producing a real record for every consequential decision, whether or not anyone expects to need it, precisely because the ones that will be needed cannot be identified in advance.

This reframes what the infrastructure is for. When we build the means to produce and preserve a defensible record of an automated decision, we tend to justify it by the near term — the appeal, the audit, the regulator, the individual who was harmed. All of that is real and sufficient on its own. But it undersells the thing. The same infrastructure that lets a person contest their decision this week is the infrastructure that lets a grandchild reconstruct this era's decisions in fifty years. We are not choosing between serving the present and serving the future. The record that serves one serves the other, because they are the same record, read at different distances.

So the honest way to see the work of this decade is as an act of inheritance, whether or not we mean it that way. We are deciding, right now, what our descendants will be able to know about how the machines of our time treated the people of our time. There is no version where they get to decide that later; the evidence either exists by then or it never will. A civilization is partly the sum of the records it kept — and we are, at this moment, writing that sum or erasing it, one discarded decision at a time. The question is only whether we will do it on purpose.

— Dispatches · Summit Cognitive