Essay · The accountable human

Proof of Consult

Regulations and customers increasingly demand meaningful human involvement in AI decisions. Everyone claims it. Almost nobody can evidence it. The consult record: what it contains, and how it separates judgment from stamping.

Somewhere in your organization, a document claims that consequential AI decisions involve meaningful human review. Now imagine the least friendly reader that document will ever have, an auditor, an opposing counsel, a journalist with a leaked incident report, asking one flat question: prove it. Not assert it, not describe the process diagram. Produce the artifact that shows a qualified human actually weighed this decision. For most companies the honest answer is a signature in a log, and a signature, as every rubber-stamp scandal in history demonstrates, proves presence and nothing else. The gap between involvement claimed and involvement evidenced is where the next generation of AI liability lives.

A signature is not a judgment

I have written about the Rubber-Stamp Rate: the fraction of human approvals that change nothing because the human is approving on autopilot. The uncomfortable corollary is that a stamped approval and a genuine judgment look identical in most logs. Same checkbox, same timestamp, same name. If your evidence cannot distinguish the two, then in a dispute it evidences neither, and the claim of meaningful involvement collapses into the claim that someone was clocked in. Meaningful is doing all the work in that phrase, and meaningful is precisely what a bare signature cannot carry.

Anatomy of a consult record

A record that carries it has six parts, and none is exotic. Who was asked: a named person, with the qualification that made them the right person, the expertise-matching problem HumanChain is built around. What they were shown: the actual information in front of them, because judgment over a summary is judgment over the summary. How long they had: thirty seconds and thirty minutes are different events, whatever the checkbox says. What they said: the substance, not just approve or reject; a consult that cannot be quoted cannot be evaluated. What changed: did the decision move, the single strongest signal that judgment occurred, the logic of the Meaningful Override Rate. And when, anchored to what the system itself was at that moment, because a consult about a model that has since silently changed is evidence about a system that no longer exists.

Stamp and judgment, told apart in the log

Collect those six fields and the statistics start speaking. A reviewer whose approvals take four seconds each, at any hour, with no recorded reasoning and a zero deviation rate, is a stamp, and now you can see it before the auditor does. A reviewer whose time varies with case difficulty, who overrides sometimes, whose notes reference the actual case, is exercising judgment, and now you can prove it. The consult record does not just satisfy the hostile reader. It gives you the internal instrument to find where your oversight is theater while that is still a fixable engineering fact rather than a finding with a docket number.

Consult records as compliance currency

Watch what regimes like the EU AI Act actually require of deployers: oversight by competent people, monitoring, logs. Every one of those duties is discharged in evidence or not at all. The consult record is the natural unit of that evidence for the human leg, the way the independent behavioral record is for the machine leg. Together they close the loop I keep drawing: a named human, in genuine command, of a system whose identity is established, with both facts on a record neither side of a dispute can rewrite. That pair is what meaningful involvement looks like when it grows up.

Start with one decision class

You do not need to instrument everything by Friday. Pick the single decision class that would look worst in the newspaper and build the consult record for it: six fields, one pipeline, one named owner. The first week of real records will teach you more about your actual oversight than a year of process diagrams, and some of what it teaches will sting. Better a sting now than an exhibit later.

And notice what the consult record does for the humans themselves, because this is not only armor for the company. The expert who gave careful judgment under time pressure is currently invisible in most logs, indistinguishable from the stamper two seats over. A real record makes expertise legible: it shows who caught what, whose overrides held up, whose judgment moved outcomes. In a working system, proof of consult is not surveillance of the reviewers. It is the first time their actual work becomes visible enough to value.