Why accountability is a now problem

Nobody serious disputes that automated decisions will eventually have to be explained. The dispute is always about when. Accountability, the comfortable view holds, is a mature-market concern — something you address once the systems are settled, the regulations are final, and the failure modes are understood. Until then it is prudent to wait, because acting early means building against requirements that have not stabilized. This is a respectable-sounding position, and it is wrong, but it is not wrong in the way most urgency arguments are wrong. It is wrong specifically about timing, which happens to be the only thing it claims to be right about.

An urgency argument earns its keep by identifying a clock the skeptic has not noticed. Most fail because there is no such clock: the thing being sold is genuinely optional, and "act now" is a manufactured pressure. The case for decision accountability is different, because there are two clocks, they are both already running, and they are running toward each other. The mistake in the wait-and-see position is not a failure of nerve. It is a failure to look at the calendar.

The first clock: deployment is removing the human who used to answer

For most of computing's history, the accountable party in any consequential decision was a person. Software sat beside that person as an instrument — it retrieved, it calculated, it recommended — but a human made the call and, crucially, could later be asked to account for it. If a loan was denied, a claim rejected, a candidate screened out, there was someone whose judgment produced the outcome and who could reconstruct, however imperfectly, why. Accountability was not a system feature because it lived in a human memory and a human obligation to explain.

The current wave of deployment is quietly dissolving that arrangement, and it is doing it now, not on some future date when the technology "matures." As systems move from advising to acting — approving, pricing, routing, escalating, allocating on their own authority — the human who used to be the accountable party is being lifted out of the loop for reasons that have nothing to do with accountability and everything to do with speed and cost. That removal is the entire point of the deployment; it is what makes it valuable. But it has a side effect that arrives on the same schedule as the benefit: the moment the human leaves the loop, the capacity to answer for the decision leaves with them, unless something was built to catch it. There is no lag. The accountability gap opens at the exact instant the efficiency is captured, because they are the same event seen from two sides.

This is why "later" misreads the situation. The skeptic imagines a future in which decisions are automated and then, subsequently, accountability becomes an issue to address. But the automation and the gap are simultaneous. Every flow moved from human judgment to autonomous action is, in the same stroke, a flow whose decisions can no longer be reconstructed from anyone's memory. The backlog of unaccountable decisions is not something that starts accumulating once the market matures. It starts accumulating the day the system goes live, and it is accumulating right now, in production, in silence.

The efficiency and the gap are not sequential events with time between them for a considered response. They are one event, and it has already begun.

The second clock: outside demand arrives on its own schedule

The wait-and-see position also assumes that the demand to account for a decision is something the institution controls the timing of — that it can prepare its records once it decides the moment is right. This misunderstands who does the demanding. The pressure to explain a consequential decision does not originate inside the organization that made it. It originates outside: with the person who was denied and now contests the denial, the regulator who opens an inquiry, the plaintiff's counsel who files for discovery, the insurer who prices the risk, the auditor who tests the control. None of these parties consults the institution's readiness before they act. They arrive on their own schedule, and that schedule is set by events — a complaint, a headline, a lawsuit, a rule taking effect — that no vendor and no buyer gets to postpone.

Regulatory timelines are the most visible face of this clock, and they are not hypothetical or distant. Rules governing automated and algorithmic decision-making in consequential domains are being written, phased in, and enforced across multiple jurisdictions on published calendars. But the deeper point is not any single statute. It is that once a rule exists requiring an institution to account for an automated decision, the relevant date is retrospective: the regulator will ask about decisions the institution already made, in a period when it may not have been keeping the record needed to answer. A control you stand up the day the demand arrives cannot reach backward to cover the decisions already made without it. Preparedness in this domain is not something you can acquire on the spot, because the evidence it depends on had to be captured at the moment of decision — a moment that, for every unaccounted-for decision, has already passed and will not return.

Where the two clocks meet

Put the clocks together and the shape of the problem changes. One clock is opening the gap — steadily, invisibly, one automated flow at a time, with no pause between the efficiency and the exposure. The other clock is bringing the demand — on a schedule set by outsiders, pointed backward at decisions the institution has already made. The dangerous interval is the space between them: the period in which decisions are being made without the capacity to account for them, before anyone has yet demanded that account. That interval feels calm. Nothing is going wrong. The systems are performing; the efficiencies are real; no one has asked the hard question yet. It is precisely the calm that makes "later" seem reasonable.

But the calm is the exposure. Every decision made in that interval is a decision the institution will, at some unpredictable later date, be asked to defend, using evidence it either captured at the time or did not. The cost of the wait-and-see posture is not that the problem gets marginally worse each quarter. It is that an irreversible backlog of indefensible decisions accumulates during exactly the window in which the problem looks least pressing. When the demand finally arrives — and the second clock guarantees it will — the institution cannot go back and re-decide those cases with a record attached. It can only discover, case by case, which of its past decisions it can stand behind and which it cannot.

This is what makes accountability a now problem and not a later one. Not that a deadline is close, though some are. Not that the harm is dramatic, though it can be. It is that the two things which set the timing — the opening of the gap and the arrival of the demand — are both already in motion, moving toward each other, and the only variable the institution actually controls is whether it began capturing a defensible record before the decisions it will be asked about were made. That variable has a hard property: it cannot be exercised retroactively. You are either keeping the record now, for the decisions happening now, or you are not — and the decisions happening now are the ones you will someday have to answer for. There is no version of this problem that gets easier by waiting, because waiting is the mechanism by which it gets worse.

— Dispatches · Summit Cognitive