🔍 Read the full analysis: The Workday After AI: Reviewing Results Takes Time And Attention on ThorstenMeyerAI.com
Get business pricing on tech for your team
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
TL;DR
An article published by cognitive engineer Karen Feigh on November 16 argues that checking AI output can consume time and mental effort that offset productivity gains. Research varies by task: a METR trial found experienced developers took 19% longer with AI, while other experiments found faster completion on tasks suited to the tools.
Checking AI-generated work can absorb time and attention that would otherwise count as a productivity gain, cognitive engineer Karen Feigh of Georgia Tech argued in an article published November 16 in The Conversation. The research she discusses points to different outcomes across tasks: a randomized trial found experienced developers took 19% longer when allowed to use AI, while experiments in other work settings found faster completion when tasks matched AI’s abilities.
Feigh calls the work of judging whether AI output is accurate and appropriate “the verification burden.” Her article is an expert analysis, not a new experiment. It draws on research in software development, consulting and professional writing to argue that measuring only how quickly a tool produces a draft can miss time spent reviewing, testing and correcting it.
The most direct measurement comes from a METR randomized trial involving 16 experienced open-source developers completing 246 tasks in large, established software projects. Tasks were assigned either with or without AI access. In the trial, developers took an average of 19% longer with AI allowed. The tools used were mainly Cursor Pro and Anthropic’s Claude 3.5 or 3.7 Sonnet, during work conducted from February to June 2025.
METR reported that developers accepted fewer than 44% of AI code generations. In a survey, 75% said they read every line of AI-generated code, and 56% said they often had to make major changes to clean it up. For a subset of 44 tasks with usable screen recordings, about 9% of working time went to reviewing and cleaning AI output, while another 4% went to waiting for the system. Those figures describe that subset, not all tasks in the trial.
AI and work · explainer
The verification loop: where AI’s time savings can go
The draft is fast. Knowing whether it is right is the work that follows.
Time for one task (schematic)
Draft: secondsgood and bad parts in the same confident voice About 9% of the timewent to reviewing and cleaning AI output in METR’s trial with AI allowed Speed is not productivitycount the whole loop, errors included
How the hidden cost builds up
- DraftThe AI answers in seconds, fluent and confident“few clues about where scrutiny is needed”
- CheckDecide what to check and how, without knowing if or where an error iscognitive load
- FixFinding an error does not end the work, it starts a correction cycletime
- Re-checkEvery fix needs checking again; when nothing is wrong, the effort feels wastedmotivation
- Count it allProductivity is the whole human-AI workflow, not the first draftnet, not gross
Source: Karen Feigh, Georgia Tech, The Conversation, Nov 16, 2026 (the three burdens, the quotes); METR, arXiv 2507.09089, p. 20 (time spent reviewing and cleaning AI output). The bars are a schematic, not data.
Graphic: Powered by Thorsten Meyer AI
The Cost of Checking AI Work
The findings matter because many workplace claims about AI productivity focus on the speed of producing a first draft. A more useful measure is whether the entire task takes less time once review, corrections and responsibility for the final result are included. In the METR trial, participants expected AI to reduce their completion time by 24% before starting. Afterward, they still estimated a 20% reduction, despite taking longer on average. That gap suggests workers’ impressions of speed may not capture the full workflow in every setting.
Other evidence cautions against treating the software result as a general rule. In a study of consulting tasks with 758 Boston Consulting Group knowledge workers, participants using AI completed tasks within the tools’ demonstrated capabilities 25.1% faster on average and completed 12.2% more tasks. On a complex task designed to fall outside those capabilities, however, AI users were 19% less likely to produce a correct solution. The result makes the fit between a task and a tool’s capabilities relevant to both productivity and accuracy.
For employers, the practical issue is not only whether AI can produce useful work. It is also whether people have enough time, expertise and authority to check it, and whether that checking is counted when organizations evaluate costs and benefits. If accountability stays with employees while output volume rises, unmeasured review work can shift the workload rather than remove it.
Products Worth Considering

Bose QuietComfort True Wireless Earbuds with Active Noise Cancellation, Bluetooth 5.3, USB-C Charging, IPX4, Black

Anker SOLIX S2000 Portable Power Station, 2,010Wh, 1,500W

Jackery Explorer 2000 v2 Portable Power Station, 2042Wh, 2200W
As an affiliate, we earn on qualifying purchases.
Results Depend on the Task
The research Feigh cites covers different tasks and methods, so its results should not be combined into a single estimate of AI’s workplace effect. A 2023 study by economist Shakked Noy and researcher Whitney Zhang, published in Science, examined professional writing tasks. It found that ChatGPT users completed tasks 40% faster and received quality scores that were 18% higher than those of participants who did not use the tool. Those results concern the specific writing tasks studied, not every kind of work.
The BCG experiment, led by Fabrizio Dell’Acqua, Ethan Mollick, Karim Lakhani and colleagues, examined what its authors describe as a “jagged technological frontier”: AI performance can be strong on some tasks and weaker on others. Its positive results applied to 18 realistic tasks judged to fall within AI’s capabilities; the negative accuracy result came from a separate, complex task outside that range. The METR software trial, by contrast, involved experienced developers working in mature codebases, where understanding project-specific details may matter.
Feigh also connects the verification burden to an older concern in automation research: people can remain responsible for outcomes even as automated systems take on practical authority. She describes workers as supervising a fast, polished intern whose output may include both useful and flawed work. That analogy is Feigh’s interpretation of the problem, not a result measured by the studies.
“AI’s productivity gains “are often countered by time and mental energy spent determining whether the output is accurate and appropriate.””
— Karen Feigh, cognitive engineer at Georgia Tech, writing in The Conversation
Where the Evidence Stops
The available studies do not establish a single net productivity effect for AI across jobs. They examine different tasks, participants, tools and measures. The METR trial involved a small group of experienced developers and particular software tools available in 2025; its result cannot show how less experienced workers, other tools or other coding tasks would fare. The 9% review estimate comes from a subset of recorded tasks, rather than the full trial.
It is also unclear how much review time organizations currently record or how that time changes as workers gain experience with AI. The BCG result indicates a sharp difference between tasks inside and outside the tested capability range, but the studies do not provide a simple method that employers can use to identify that boundary for every assignment. Feigh’s article presents a synthesis of existing research and an argument about accountability; it does not report a new workplace-wide measurement.
Measure the Whole Workflow
The next step for organizations evaluating AI is to track completion time, review effort and quality together, rather than relying on how quickly a first draft appears. That means recording time spent checking, correcting and testing output, as well as whether the final work meets the task’s requirements. Comparisons should specify the task, tool, participant experience and time period.
Feigh argues that workers who remain responsible for AI-assisted work need time and training to review it. The studies summarized here support the need to distinguish between tasks where AI can help and tasks where it may produce errors that are harder to spot. Whether newer tools reduce the review burden, and whether that leads to net gains across particular occupations, remains to be demonstrated through further measurement.
Key Questions
Did the METR study find that AI always slows workers down?
No. The trial found that 16 experienced developers took 19% longer on average across 246 software tasks when AI was allowed. METR said the result does not mean AI is unhelpful in other economically relevant settings.
How much time did developers spend reviewing AI code?
For a subset of 44 tasks with usable screen recordings, METR estimated that developers spent about 9% of working time reviewing and cleaning AI output. The figure applies to that subset, not every task in the trial.
Did other studies find productivity gains?
Yes. A 2023 professional-writing study found ChatGPT users finished tasks 40% faster, with quality scores 18% higher. In a BCG consulting experiment, AI users completed tasks within the tested capabilities 25.1% faster on average, but performed less accurately on one task outside those capabilities.
What does Feigh mean by the verification burden?
She means the time and mental effort needed to judge whether AI output is accurate and appropriate, then review or correct it before use. Her article argues this work can offset some of the time saved by generating a draft.
Source: ThorstenMeyerAI.com
