A Harvard working paper on coding-agent adoption found that once developers started using AI coding tools, review time per proposed change rose 49 percent. The share of proposals that needed corrections nearly doubled, and comments per review went up 35 percent. Intermediate coding activity climbed as well. The number of issues actually resolved did not.
So the tools produced more code, a larger share of that code needed fixing, and people spent more time checking it, without more problems getting solved.
That arithmetic describes a role now forming in organizations that deploy agents: the person who signs off on what the machine produced. Their approval carries organizational weight. Three pressures are converging, though, to hollow out what "checked" can mean.
Volume is the most obvious. A neuroradiology study found error rates rose 63 percent when radiologists interpreted more than 26 studies per shift. Vigilance decay is subtler. In a controlled experiment, participants caught 82 percent of automation failures when the system's reliability varied, but only 33 percent when it was consistently reliable. When the machine is usually right, the person monitoring it stops looking closely. Then there is anchoring. A mammography study found that incorrect AI suggestions swayed radiologists at every level of experience; the recommendation became where they started rather than the thing they were evaluating.
Any one of these can be managed. Together, they describe conditions in which rubber-stamping is a rational choice for the individual reviewer even when it is dangerous for the organization. A randomized experiment with Taobao's customer-service agents adds a structural wrinkle: human intervention repaired technical problems effectively but not emotional ones, where customers felt dismissed rather than confused. Some kinds of output resist meaningful review however attentive the reviewer is.
Regulators already treat human approval as a real safeguard. The SEC requires annual CEO certification of controls on automated trading. Aviation law makes the pilot in command responsible for the aircraft's operation, automated systems included. Both frameworks assume the person approving can actually inspect what they approve. As that assumption erodes, the signature stays on the page with its legal force intact, while what it attests to grows thinner.
At Wedbush Securities, two employees put together a one-page review of the firm's automated trading controls over a single weekend. It didn't mention the relevant regulation or test specific settings. A senior executive signed the certification anyway. The SEC sanctioned the employees who had authority over the controls, but the case shows how formal certification can drift away from actual inspection while the signature's legal weight stays exactly where it was.
What happens to this role as agent output grows? One path is professionalization. The UK's Financial Reporting Council has issued guidance stating that audit partners keep full responsibility for AI-assisted work and that approving output, by itself, doesn't show the proper method was followed. That points toward credentialed reviewers with published accuracy rates and a defined scope. Another path is automation, with a second model reviewing the first model's output. That fixes throughput, but someone still has to decide the automated reviewer is trustworthy, and that person inherits the approval burden the automation was supposed to relieve.
The third possibility is that the role persists as organizational theatre, and here the question of power comes in. In 2018, an Uber self-driving test vehicle struck and killed a pedestrian in Arizona. The safety operator, whose job was to monitor the system and intervene, pleaded guilty to endangerment. The NTSB also found that Uber's organizational failures contributed to the crash, so liability landed on both. Of everyone involved, though, she had the least power to redesign the system that failed.
Our institutions are very good at producing this particular outcome, in which liability flows toward the people least able to refuse it. Organizations now have to decide whether the widening gap between what an approval signature claims and what the signer actually inspected is a design problem worth paying to fix, or someone else's risk to carry. So far, the incentives favor the second answer.