The second chance the model forgot

Consider a person who did something, once, that they are not proud of — a conviction in their early twenties, a bankruptcy after a failed business, a debt that went to collections during a bad year. Consider that the law, in its slow and deliberate way, decided that these things should eventually stop following them: the record sealed, the conviction spent, the debt aged off the report, the bankruptcy discharged and then, after a term of years, gone from the file. And consider that none of this care survived contact with the systems that now decide, in a fraction of a second, whether that person gets the apartment, the account, the job, the loan. The law let them start over. The model did not get the memo. This essay is about the gap between those two facts, and about what it is like to live in it.

We are used to hearing that machines forget too little only as a privacy complaint — data hoarded, records leaked, a past dredged up by a search engine. But the deeper problem is not that the past is stored. It is that the past continues to count. A determination that a human institution had already decided should lapse keeps arriving in the present with full force, its expiry ignored, its context stripped away, weighed today as if the intervening years and the deliberate act of forgiveness never happened. From the receiving end, this is not a data-retention issue. It is the quiet cancellation of a second chance the law meant you to have.

Forgetting on purpose

It is worth being clear that forgetting, in human institutions, is not an accident or a limitation we grudgingly tolerate. It is a designed feature, built into law at real cost and defended against the obvious objection that it hides the truth. The sealed juvenile record exists because a society decided that a child's mistake should not be a lifelong sentence. Expungement exists because we concluded that some debts to justice, once paid, should be allowed to close. The statute of limitations exists because we judged that a claim has a shelf life — that after enough time, the interest in repose outweighs the interest in pursuit. The debt that ages off a credit file, the bankruptcy that clears after its term, the spent conviction that need no longer be disclosed: each is an institution saying, in effect, this was true, and it will stop counting anyway.

Notice what these mechanisms have in common. None of them denies that the underlying thing happened. The conviction was real; the default occurred; the bankruptcy was filed. They are not corrections of the record — they are decisions about the record's standing. The fact is left intact and its power to determine your future is retired on schedule. This is a distinct move, and a hard-won one. It took centuries of argument to establish that a just society is not one that remembers everything and forgives nothing, but one that builds, into its own machinery, an expiry date on consequence. The reason is not sentimentality. It is that a life under permanent surveillance by its own worst moment is not a life anyone can actually live. A person who can never move past anything has no reason to try to be better, because the ledger never resets. We forgive on purpose because the alternative — total, permanent, unforgiving recall — is unbearable, and we knew it.

The machine that forgets nothing

Automated systems, left to their defaults, do the precise opposite of all this. Their native posture is to retain everything and to resurface it whenever it is asked for. A model does not have a statute of limitations. A scoring pipeline does not observe an expungement. A record pulled into a decision does not carry, by default, the metadata that would say this determination was meant to lapse in year three, and it is now year nine. The machine treats a nine-year-old flag and a nine-day-old flag as the same kind of input: a signal, present and countable. Its memory is flat. Everything is now.

And here the trouble compounds, because a system's perfect recall gets rebranded as a virtue. We are told the model is more accurate precisely because it forgets nothing — that it sees the whole history, misses no detail, weighs every fact. But accuracy about what? A determination can be perfectly faithful to something that happened and still be something the law itself has decided you should no longer be held to. Remembering it accurately is not the same as being entitled to hold it against you. The machine collapses that distinction. It confuses the truth of a fact with the fact's continuing right to count, and dresses the confusion up as rigor. What looks like scrupulous memory is, from the receiving end, a refusal to honor a forgiveness that was already granted.

The law decided your mistake should stop following you; the model never got the memo, and it has a better memory than the law.

So the person on the receiving end is judged today by a version of themselves that no longer legally exists. They are denied by a determination whose expiry date has passed — one a human process would have let fade, its provenance long gone, its context unrecoverable, resurfaced by a system that has no concept of the interval that was supposed to have redeemed them. They cannot even contest it cleanly, because to the machine there is nothing anomalous to contest: the fact is true, the input is present, the score is computed. The very thing that was supposed to protect them — the deliberate lapse — is invisible at the layer where the decision is actually made.

The right to have your past stop counting

What the affected party is owed here is not deletion, and not a lie. It is that the machine honor the same distinction the law already draws — between what is true and what is still allowed to count. This is a form of standing over your own history: the position that some determinations about you have expired, and that an accountable system must treat expiry as a first-class fact about its inputs rather than an inconvenience it is free to ignore. A record that a human institution sealed should arrive at the machine already marked as sealed. A determination with a shelf life should carry that shelf life, and a system reaching for it past its term should have to justify why, not silently spend it.

This is the opposite of treating total recall as an unalloyed good. An account built for the individual — a Decision Receipt worth the name — should be able to show not only which determinations a decision rested on but how old they were, whether they had passed their expiry, and by what warrant a stale one was allowed to count. Provenance and expiry are not separate concerns; a determination without a live sense of its own age is a determination pretending the calendar stopped the day it was recorded. The right to a fresh start, translated into the vocabulary of machine decisions, is the right to have stale determinations lapse — to insist that a system prove a fact still has standing before it is permitted to weigh on your future.

None of this asks a system to be less truthful. It asks it to be as forgiving as the law that governs it already is — to build in, as human institutions painfully learned to, an expiry on consequence. A society that made forgetting a feature of its own record-keeping did so because it understood something the machine has quietly unlearned: that a past which never stops counting is not accuracy. It is a sentence no one agreed to serve. The account we are owed is one that remembers we were meant to be allowed to begin again.

— Dispatches · Summit Cognitive