When did your AI reviewer last send a piece back? If they have to think about it, I’d say your human in the loop is approving on autopilot, and that’s the exact failure a human in the loop audit is built to catch. The loop didn’t break. It faded one clean approval at a time, until your safeguard became a person who has learned to trust the machine’s answer.
In Canada this check carries weight beyond quality. The privacy commissioners’ principles for generative AI say accountability for decisions rests with the organization, not with the automated system that supported them.
Most firms describe their safeguard in one sentence. A person checks every piece of AI work before it goes out. That sentence was true the week the process started. The reviewer was new to the tool, read every line, and sent work back often. Months of clean output later, the sentence still sits in the process document. The reviewer it describes has changed jobs, from checker to witness.
Why does a careful reviewer go soft?
Because the machine keeps being right, and enough clean runs in a row teach anyone to stop hunting for the one that is wrong. There is a name for it, automation bias, and it needs no laboratory to take hold.
It takes hold at a desk, in a queue. Attention is a budget, and a reviewer spends it where trouble has shown up before. Forty drafts that read fine tell the budget holder the forty-first is a safe place to save. Careful reviewers drift too, in a quieter way, because an AI’s lean can become your own call without you noticing.
The shapes go familiar too. The reviewer has seen this report, this letter, this summary a hundred times, and familiar work gets read the way a familiar road gets driven, eyes open, mind somewhere else. Approving takes one click. Sending a piece back starts a conversation, maybe an argument, and always more work. The path of least effort runs in one direction, and it runs downhill toward yes.
Clean runs teach trust faster than they teach care.
What does soft reviewing look like from the outside?
It looks like speed. A full batch approved before the coffee cools. It looks like a change count of zero across a week of output, or changes so small they are spelling and spacing. It looks like a reviewer who reads the first paragraph and the last, trusts the middle, and cannot remember the last piece they sent back.
Nothing about it looks careless, which is what makes it dangerous. The reviewer is experienced. The output mostly deserves the trust. The piece that finally carries an error into a client’s hands will not announce itself. It will read like the forty clean ones before it, because errors that survive a faded check look like normal work. They arrive dressed as the ordinary.
A safeguard that never stops anything is just a step in the workflow.
How do you audit the approver?
You count. Take one week of AI output your reviewer approved and count the pieces they actually changed in ways that matter. Leave the spelling fixes out of it. Count the claims softened, the figures corrected, the paragraphs cut, the pieces sent back whole.
Then read a sample of that week’s work yourself, starting with the pieces that were approved fastest. The fastest approvals are where the safeguard runs thinnest. That’s where I’d start reading.
The record helps you here. The same action record you would ask to show last Tuesday also carries the approvals, who made them and how the week flowed. If the change count comes back at or near zero, believe it. A check that never changes anything has stopped checking, whatever the process document says. And treat the reviewer fairly while you do this. They drifted the way every person drifts. The audit is of the check, not of the character.
What do you rebuild once you have the number?
You rebuild the conditions, because the workload went soft before the person did. Give the reviewer smaller batches and a hard stop on how much one sitting holds. Rotate a second reader through a sample each week, because fresh eyes have no trust built up to spend.
Make sending work back a normal, expected act that gets thanked out loud, rather than a small act of rebellion. And give client-facing work a slower lane with a longer look, since that is where a waved-through error lands on your name, the same risk an agent carries when it inherits your firm’s habits.
This belongs inside the same monthly discipline as the rest of your AI governance, a habit, not a binder. Add one line to that habit. This month, how much did our reviewers change? A check on the checker, run on a schedule, is what keeps the original check worth having.
You cannot order a person to pay attention. You can build a week where attention survives.
How do you audit your AI reviewer before Friday?
Run the audit before Friday. Pull last week’s approved AI output, count the changes that matter, and re-read the fastest approvals yourself. Whatever number you find, rebuild one condition around the reviewer this month, the batch size, the second reader, or the slower lane for client work.
Then put the approver audit on the monthly list, so the next fade gets caught as routine rather than as a surprise. When I build AI Automation Systems, the approver audit is designed in beside the approval step.
Book the approver audit for the first Monday of every month. I go through checks like that each Friday in the AI Newsletter with ZAK.
Human in the loop audit, quick answers
What is automation bias in AI review?
The pull to trust a machine’s output after enough clean runs. A reviewer who approves forty good drafts starts saving attention on the forty-first. Nothing looks careless, so the error that finally reaches a client reads like the normal work around it.
How do you audit a human in the loop?
Take one week of approved AI output and count the pieces the reviewer changed in ways that matter, leaving spelling fixes out. Then read a sample yourself, starting with the fastest approvals. A change count at or near zero means the check has stopped checking.
How do you keep AI reviewers sharp?
Rebuild the conditions around them. Smaller batches with a hard stop per sitting, a second reader rotating through a weekly sample, thanks for every piece sent back, and a slower lane for client-facing work. Then add the approver audit to your monthly AI check.
8 Comments