2 Sources
[1]
Catching AI's mistakes is a mental burden - and a drag on productivity
By now, almost everyone has asked an AI agent a question, only to realize that the answer seems not quite right. Something might be missing or wrong or out of date. An executive assistant told me she uses AI tools for drafting itineraries and briefings, but she's caught enough issues that she now
[2]
Catching AI's mistakes is a mental burden -- and a drag on productivity
By now, almost everyone has asked an AI agent a question, only to realize that the answer seems not quite right. Something might be missing or wrong or out of date. An executive assistant told me she uses AI tools for drafting itineraries and briefings, but she's caught enough issues that she now
Share
Copy Link
A cognitive engineer who studies how people work with automated systems argues that gains from AI are often offset by the time and mental effort spent checking whether its output is right
1
. The essay's point is that measuring AI productivity by how fast it generates a first draft misses the second half of the work: someone still has to catch the errors, and that someone is usually the worker rather than the machine.
Source: Phys.org
The sharpest finding cuts against the feeling of speed. In one randomized trial, experienced software developers completed their work 19% slower once a verification step was counted, and they still believed AI had made them faster
1
. What people feel when AI hands them polished output is generation speed; what the stopwatch catches is the checking that follows.Other studies point the same way. A large study of consulting teams found more people reached wrong conclusions when they used AI outside their own area of expertise
1
. That is exactly where they are least able to check it. A separate study found AI's explanations of its reasoning helped people spot errors on easy questions but not hard ones, where help was needed most1
. And a meta-analysis of 86 studies found people remember and understand material they produced themselves far better than material they merely read1
.The author identifies three burdens. Cognitively, there is no clear stopping point and no way to know where an error hides. On time, each error found starts a cycle of correction and rechecking. Motivationally, a checking session that finds nothing feels unproductive, which erodes vigilance; research on automated decision support documented that pattern two decades ago
1
. Two more findings make under-checking likely: easy-to-read text reads as credible text, and people tend to sort AI agents into either capable or incapable rather than judging each output. Explaining the AI's reasoning did not remove the reluctance to spend effort checking1
. Research on error correction found that the more work fixing a mistake requires, the less likely people are to fix it1
.
Source: The Conversation
The author frames this as an old problem wearing new clothes, citing cognitive engineer David Woods' 1985 description of the double bind: "Automation exercises practical authority, yet responsibility remains with the human"
1
. In a recent survey, knowledge workers described their own critical thinking moving from production toward verification1
. The author's analogy is "supervising an extraordinarily fast and polished intern", one who delivers strong and weak work in the same confident voice and bears no responsibility for either1
.The counter-evidence is notable because economics research has documented real AI productivity gains in office work
1
. The author is careful not to overreach: the studies cited do not establish a universal verification tax, and the evidence is study-by-study1
. If workers stay accountable for what AI produces, the author argues, employers owe them the time, the training and the authority to check it1
. On the author's account, productivity should be measured by whether the whole human-AI workflow turns out trustworthy work at lower total cost1
. One executive assistant the author spoke with now schedules time to check AI itineraries and briefings for inaccuracies and biases, after catching repeated issues1
.Related Stories
Summarized by
Navi
[1]
10 Feb 2026•Business and Economy

17 Oct 2025•Technology

14 Jan 2026•Business and Economy
