How this editorial can be challenged
Who carries the verification cost when producing an answer becomes cheaper than checking it?
An organization can record savings at the point of production while moving checking, exception handling and recovery to another worker, customer or public institution. Counting only the visible production step can therefore mistake cost transfer for productivity.
Human work already contains errors and expensive delays. Requiring exhaustive review of every automated output can destroy useful gains and make care or services less accessible.
The relevant comparison is total cost and quality against the existing workflow, with risk-proportionate checks. Low-stakes reversible outputs do not need the same review as a clinical decision or court record. Automation that reduces checking effort as well as drafting effort is a genuine gain.
The clinical review cannot establish a universal oversight failure rate or the cost of review. OpenAI reports its own containment work. Amex has not disclosed that affected roles were reviewers. None of these sources measures a common economy-wide verification cost.
Prospective comparisons showing lower total human effort, fewer harmful errors and reliable recovery after automation—without exporting work to customers or other teams—would demonstrate that verification costs are being reduced rather than hidden.
The work after the answer
There is a familiar moment at work: somebody sends you something that looks finished, and your job becomes finding out whether it is. A polished document can make that job harder. The prose no longer tells you where the writer was uncertain. The appearance of completion arrives before the work of justification.
That is the economic question I want to put beside today's AI headlines. We are getting very good at counting what a system produces. We are less consistent about counting the effort required to trust it. If the first number rises and the second quietly lands on somebody else's desk, we may call a transfer of work a productivity gain.
A signature is not a unit of safety
The new clinical review makes this distinction unusually concrete: human participation did not consistently prevent reliance on wrong automated advice. Its heterogeneous evidence cannot tell us one universal failure rate. It can tell us why the existence of a reviewer is an inadequate description of a safeguard.
Imagine two hospitals using the same tool. In one, a clinician can inspect the underlying evidence, compare an independent judgment and pause the workflow. In the other, checking is added to an already full queue. These are hypothetical examples, not hospitals measured in the review. They illustrate why the organizational conditions belong in the evaluation, rather than being assumed equal because both institutions promise human oversight.
The denominator belongs to the whole process
A useful productivity measure follows a decision to completion. It includes the initial task, checking, corrections, escalations and recovery when something goes wrong. It also records the outcome. A cheaper decision that harms someone is not made efficient by leaving the harm outside the spreadsheet.
This does not require an imaginary universal price for trust. Start locally: how many minutes did the old process require; how many does the new one require across all participants; what happened to error rates and unresolved cases? A clinic, newsroom and bank will need different measures. The principle is the same: do not stop the clock when the machine finishes speaking.
The savings and the bill can go to different people
American Express's restructuring is not evidence that it is removing safety reviewers. We do not know that, and should not imply it. The announcement does, however, make the distinction between an immediate organizational decision and a demonstrated change in work impossible to ignore.
A broader possibility deserves testing wherever automation and restructuring meet. The department purchasing software may claim the saving while another team handles the exceptions. Or a customer may spend an evening repairing a transaction that used to be resolved by staff. Those are hypotheses about cost distribution, not findings about Amex. They become useful only when we trace the work rather than attach them automatically to a layoff headline.
A stronger machine can also help pay the bill
The strongest objection to this argument is straightforward: manual work is not a pristine baseline. People miss errors, wait too long and struggle with limited capacity. AI may make checking easier, surface relevant evidence or help security teams find problems they would otherwise miss.
OpenAI's reported use of agents to test its own containment illustrates that constructive possibility, while remaining a developer account rather than independent certification. The lesson is not to prohibit automation of review. It is to avoid treating an automated checker as the end of the evidentiary chain. Test the complete arrangement, including what happens when both the first answer and its checker are wrong.
Buy the outcome, not the appearance of review
I would want a procurement or deployment decision to state which errors matter, who can interrupt the process, what evidence they can access and how their workload will be measured. These are our proposed management questions. They are not a claim that one standard design has already been proven across sectors.
The tradeoff is real. Checking every low-stakes suggestion as though it were a surgical decision wastes time and can deny people useful services. Risk-proportionate checks, reversible actions and targeted testing offer a more defensible route. WHO's emphasis on intended use and lifecycle monitoring provides a baseline for health applications; each actual service still needs its own evidence.
The decision we should make before the next rollout
My position is not that AI is secretly uneconomic, nor that every efficiency claim is a trick. We do not have the evidence for either sweeping verdict. It is that the claimed gain should survive an honest accounting of the work that follows the answer. Show reduced total effort and better or equivalent outcomes, including the hard cases, and the case for adoption becomes stronger.
Before approving the next rollout, decide where the second look will live. Give it a budget, a person or team with the power to act, and a way to demonstrate that it catches what matters. If the plan cannot say who will do that work, its efficiency claim is unfinished. The cheapest answer is not necessarily the cheapest decision.
Read the reporting
Opinion is ours. The factual record is linked below.
npj Digital Medicine — human oversight review OpenAI — containment and training security American Express — Q3 filing WHO — AI health oversight considerations