Show Your Work

A manager gets the pay letter. The raise is fair — genuinely fair, defensible against the market, the budget, the performance record. She reads it twice, says nothing, and over the next quarter gives a little less than she used to. Not out of spite. She just can't see how the number was reached, and in the absence of seeing, she has filled in the gap herself. Usually with the worst available story.

This is the part of organizational life that the accuracy people keep missing. The outcome was right. The effort dropped anyway.

We are about to spend a decade pointing artificial intelligence at exactly these moments — who gets the raise, the rating, the promotion, the territory, the flag on the performance dashboard — and the entire public conversation about whether that's a good idea is being conducted in the wrong currency. The currency on the table is accuracy. Is the model right? Is it less biased than the manager it replaces? Does it predict better? Those are real questions. They are not the load-bearing one. The load-bearing one is whether the people on the receiving end can see how the decision was made — because a century of evidence says that's what governs whether they accept it, trust the place, and keep giving the discretionary part of their effort that no one can compel.

Most AI cannot show that. Not because its builders are careless, but because the thing was built as a box you can't see into. And that, not accuracy, is where the contest is going to be won or lost.

What we've known since before the transistor

Start with the finding that should have ended the debate forty years ago, and somehow didn't reach the people now buying AI.

Thibaut and Walker, studying how people respond to dispute procedures in the 1970s, found something that violates the economic intuition: people will accept an outcome they didn't want — a worse outcome — if they had a voice in the process that produced it. The effect is robust enough to have its own name in the literature, the fair-process effect. Give someone a say, run the procedure consistently, and they will sign off on a result that went against them. Deny them the process and hand them the identical result, and they fight it.

Leventhal, in 1980, turned that into a checklist — the conditions a procedure has to meet to read as fair: apply it consistently, suppress the decider's private bias, use accurate information, make it correctable when it's wrong, give the affected parties a voice, keep it ethical. Read that list again with an algorithm in mind. We'll come back to it.

Then the part that explains the manager and the pay letter. Lind and Van den Bos, building the fairness-heuristic account, established when fairness matters most: under uncertainty. When people don't know whether they can trust an authority — when information about the process is missing — they reach for fairness as the heuristic that tells them how much to invest. The less they can see, the harder that judgment swings. Absent real information, they don't suspend judgment. They manufacture it, and the manufactured version is rarely generous.

That is the mechanism under the oldest intuition every manager has: rumors fill the vacuum. It isn't folk wisdom. It's a measured property of how humans process authority. Where the process is invisible, suspicion is not a risk — it is the default, and it drives behavior as hard as the facts would have.

And the behavior it drives is the expensive kind. Across the empirical record the chain is consistent: perceived justice feeds organizational commitment, and commitment is what predicts whether people stay and how much they give. Aquino, Griffeth, Allen, and Hom traced justice through satisfaction and withdrawal cognitions into turnover in 1997. Gim and Desa, studying 226 employees in Malaysia, found that affective commitment fully mediated the path from justice to turnover intention — the justice perception didn't act on people directly, it acted by changing how attached they felt. Vandenberghe and Tremblay, across two samples, found that the effect of pay satisfaction on intended turnover was, again, fully mediated by commitment. Read that one slowly, because it's the whole argument in miniature: it isn't the dollar that keeps people — it's the felt thing the dollar is supposed to signal. Get the felt thing wrong and the dollar doesn't save you.

Where I have to be precise

Here is where the honest version of this argument parts company with the marketing version, and the precision is the point.

Procedural fairness does not beat distributive fairness at everything. The pattern in the literature is two-handed: how fair the outcome was drives satisfaction with the outcome itself — your pay, specifically — while how fair the process was drives the organization-directed things: commitment, trust, intention to stay, the willingness to go past the job description. So I am not going to tell you that transparency makes people love their paychecks. It doesn't. What it does is govern whether they trust the institution that issued the paycheck and keep spending discretionary effort on its behalf. That happens to be the exact thing every leader says they want from AI adoption and the exact thing a botched rollout destroys first.

And one more caveat, because the effect has an edge. The fair-process boost can shrink, or even invert, when people are highly uncertain about the outcome and inclined to read a fair process as a setup — the boundary the Brockner tradition mapped. Fair process is a powerful lever. It is not a free one in every condition. Anyone who tells you it always works is selling.

With those two qualifications in hand, the claim is strong and it is defensible: for the outcomes organizations actually care about under AI — trust, retention, discretionary effort — the process and what people can see of it is not a soft factor. It's the factor.

The box that can't pass the test

Now put the black box back on Leventhal's list.

Consistency — maybe; a model is at least consistent in its inconsistencies. But accuracy you can inspect? Correctability when it's wrong? A voice for the person it scored? An account of why, in terms they can check? A black-box model fails these not occasionally but by construction. Its opacity isn't a bug to be patched in version two; it's the architecture. And recall what the fairness-heuristic work established: fairness judgments swing hardest exactly when information is missing. So a black box doesn't merely risk being unfair. It builds the information vacuum the science says suspicion rushes into, and then points that vacuum at the highest-stakes decisions in a person's working life. It is, almost precisely, the worst possible instrument for a procedurally-sensitive species — optimized for the one currency (accuracy) that doesn't govern acceptance, blind to the one (legible process) that does.

You cannot explain your way out of this after the fact. An explanation bolted onto a system that didn't actually decide that way is not transparency; it's a press release, and people are very good at smelling the difference. The legibility has to be load-bearing — the system has to actually decide the way it says it decides — which means it has to be built that way from the first commit. Transparent by construction, or not transparent at all.

What you build instead

So build it the other way. Make every step of the data work visible and testable; when something can't be verified, make it fail loudly and sit in quarantine with a reason attached, rather than slip through and quietly bend a decision. Say, in plain language, what each analytic does, what evidence stands behind it, and where it stops being trustworthy. Expose the services over typed, inspectable contracts instead of behind a wall. None of that is ethics garnish. It is the engineering expression of Leventhal's list.

And then measure the thing directly, which is the part almost nobody does. For every consequential decision, ship two numbers: the measured fairness — the residual disparity you can compute from the data — and the felt fairness — the perceived-justice reading from the people the decision landed on. The literature has been telling us for decades these are different axes; treat them as different axes. Because the most valuable signal isn't either number. It's the gap between them. A small measured disparity sitting next to a large felt unfairness is not noise to be averaged away — it's a diagnosis: the pay is fine, the process is the problem, and the fix is cheaper than the one you were about to buy. That gap is transparency proving its own worth in a number — evidence that the legible lever is the better lever, on the organization's own data.

I'll concede the obvious thing: building this way is harder, and it surrenders the marginal accuracy you might wring from a model you can't explain. The wager is that the trade runs the other direction from where the field assumes — that the legible decision people accept beats the accurate one they resist, on every outcome that pays the bills. The evidence says the wager is sound.

The hill

Here's why this is the whole argument and not a feature.

A black-box incumbent cannot copy transparency. It's not a setting they can toggle; their model was built opaque, their pipeline was built opaque, and "explainable AI" pasted on top is the press release again. Legibility is the one move that can't be retrofitted — which makes it the rare differentiator that an established competitor with more data and more money structurally cannot match. The thing that's hard to build is exactly the thing that's hard to copy.

It also happens to be where three arguments that look separate turn out to be one. The study of how humans actually interact with AI keeps returning the same finding — that acceptance runs through process and visibility, not through accuracy. People analytics is the discipline that knows how to measure that — fairness, trust, commitment — defensibly rather than by vibe. And the way you build software so it can be seen through is the engineering that makes the first two real instead of aspirational. Behavioral science tells you what to build; measurement tells you whether it's working; the architecture makes it legible enough to check. Same argument, three faces.

The manager with the pay letter was never going to be won by a better number. She was going to be won, or lost, by whether she could see the work. So show your work. It's the one thing the box can't do — and on the evidence, the one thing that was ever going to matter.