The machine generates the draft. The human reviews it. The output goes forward.
That arrangement appears on more slides than any other workflow I encounter, usually labelled responsible deployment. It is nothing of the kind. It is the most consequential design error I meet, and there is forty years of evidence for that.
Start with what the reviewer actually does, because it is not what the diagram implies. You open the document and skim it. The structure looks sound, the tone is about right. You tidy a heading, soften a phrase that overclaims, fix an inconsistency, and move it along. It took eleven minutes and it felt like diligence.
The substantive decisions were made by the machine: how the argument is built, which evidence was selected, what the conclusion asserts. You approved them on the basis of plausibility rather than verification, because verification would have meant reconstructing the reasoning, and you had three more of these before lunch.
Nobody involved is being negligent. The role trains people to work exactly this way. When I ask teams how often they review a colleague's work as thoroughly as the approval chain implies, I get a nervous chuckle from the team and an outright laugh from leadership. I include myself in that. The honest test of any review step is whether the professional is still in the work or merely near it.
None of this is new, which is the part that should trouble us. In 1983 a British researcher named Lisanne Bainbridge published a paper on automation in chemical plants and aircraft cockpits, and identified two ironies that describe the office of 2026 with uncomfortable precision.
The first is skill atrophy. You automate a process because the machine is more reliable, then retain a human to intervene when it fails, and the automation deprives that human of the practice they would need to intervene effectively. The safety net exists on paper. It is not there in practice.
The second is vigilance. Attention drifts on systems that rarely fail, which is biology rather than a character flaw. And the errors that get past a drifting reviewer are not the obvious ones. They are the contextual misalignments and quiet logical faults that need real engagement to catch, which is precisely the engagement the role discourages. Kahneman supplies the mechanism: fluent text recruits fast automatic thinking, real review requires slow effortful thinking, and under deadline the fast system wins.
Here is why this bites harder in an office than in a plant. A worker reduced to watching a machine is disengaged, which is a shame, but the machine still produces. A professional reduced to reviewing work they did not shape and cannot interrogate is being stripped of the one thing that made them valuable. In knowledge work the human mind is not the operator of the process. It is the process. Reduce it to passive supervision and the work only looks like knowledge work. The substance has quietly gone.
There is a consequence senior leaders should take personally. Madeleine Clare Elish called it the moral crumple zone: the human nominally in control of an automated system, who absorbs the blame when it fails. Just as a car's crumple zone deforms to protect the passengers, the passive reviewer deforms to protect the organisation. They did not produce the output. They could not meaningfully have verified it. But they approved it, and that approval is the fig leaf the firm shelters behind.
The regulator has noticed. Article 14 of the EU AI Act requires high-risk systems to be designed so humans can effectively oversee them, and it names the failure mode directly: automation bias, the tendency to over-rely on system output. The drafters had read their Bainbridge. Most of those obligations took effect on 2 August 2026, and a rubber stamp does not satisfy the plain intent.
The fix is to move the human upstream, to the framing and the interrogation, where their judgment actually bears on the outcome.
Something to try this week. Pull up the last five things you approved. For each, write one line: the substantive change you made. Not a formatting fix or a softened phrase. A decision, a rejection, a reframing, a piece of context the machine could not have had.
If you cannot name one for three of the five, you are not reviewing. You are absorbing liability. And the number you find is a property of the role rather than a verdict on you, which means it can be redesigned.