🔍 Read the full analysis: AI Can Produce Work Quickly, But What Does Checking It Require? on ThorstenMeyerAI.com
Get the latest gadgets delivered free with Prime
- Fast, free delivery on millions of items
- Prime Video, Amazon Music and more included
- Member-only deals all year
TL;DR
A commentary by Georgia Tech cognitive engineer Karen Feigh highlights the time and mental effort workers spend checking AI-generated work. Studies show results vary: AI slowed experienced developers in one trial, while helping consultants and writers on some tasks; errors beyond reviewers’ expertise were often missed.
Georgia Tech cognitive engineer Karen Feigh argued in a November 16 commentary that checking AI-generated work can consume the time and attention saved by producing it quickly. The research she cites shows mixed results: AI slowed experienced developers in one controlled trial, while improving speed on some writing and consulting tasks, but users and reviewers can still miss consequential errors.
Feigh calls the effort spent deciding whether AI output is accurate and appropriate the “verification burden.” In The Conversation, she described workers as supervising a fast, polished intern whose output can be correct or wrong in the same confident tone, while the human remains responsible for the consequences. Her article is expert commentary linking to existing research, not a new experiment.
The clearest example of a possible productivity reversal comes from a randomized METR trial involving 16 experienced open-source developers completing 246 tasks in established software projects. Developers took 19% longer when AI tools were allowed. Before the trial, they expected AI to cut completion time by 24%; afterward, they still estimated it had saved 20%. The measured time included review and work after code generation.
METR reported that developers accepted fewer than 44% of AI code suggestions. In a survey, 75% said they read every line of generated code, and 56% said they often needed major changes to clean it up. On a subset of 44 tasks with usable screen recordings, about 9% of working time went to reviewing and cleaning AI output. These figures describe that study’s developers and tasks, not a universal cost for AI use.
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
When Review Erases Speed Gains
The findings matter because organizations often assess AI by how quickly it produces a draft or completes an isolated step. But the relevant measure is the whole workflow: generation, checking, correction, and the consequences of errors. A faster first version does not necessarily mean a faster completed task.
Accuracy can be difficult to judge even when a result looks polished. In an experiment involving 2,784 people reviewing AI-style suggestions, participants usually caught typos, but correctly identified only 31% of errors that required understanding the task’s rules. The result suggests that proofreading for surface problems may not be enough when correctness depends on expertise. It does not establish that all AI review fails or that reviewers perform the same way in every workplace.
That distinction affects staffing and accountability. If employees are expected to approve AI work, they need time and relevant knowledge to check it. Feigh’s argument is that responsibility remains with the human even when automation has practical influence over the work. Employers evaluating productivity should count review effort and error rates, rather than relying only on output volume or drafting speed.
Products Worth Considering

Amazon Kindle Scribe Colorsoft 64GB with Premium Pen

LG 27GR83Q-B 27-inch Ultragear QHD IPS Gaming Monitor

DJI Osmo 360 Motorcycle Handlebar POV Combo
As an affiliate, we earn on qualifying purchases.
Results Depend on the Task
Other studies cited by Feigh show why the METR result should not be read as a general verdict on AI productivity. A 2023 experiment on professional writing tasks found that ChatGPT users completed work 40% faster and produced results rated 18% higher in quality. Those results apply to the tasks and participants in that experiment, not every kind of writing or professional work.
In a preregistered experiment with 758 Boston Consulting Group knowledge workers, researchers examined tasks both within and beyond what they described as AI’s “jagged” capability frontier. On 18 tasks within that frontier, AI users completed 12.2% more tasks and worked 25.1% faster on average, with higher-quality solutions. On one complex task outside it, AI users were 19% less likely to produce a correct solution than participants without AI.
The studies used different people, tools, work, and measures. METR’s software trial used tools including Cursor Pro and Claude 3.5 or 3.7 Sonnet during February to June 2025. Its authors cautioned that the result does not show AI is unhelpful in other economically relevant settings, or rule out future gains in the tested setting. Results from one task or tool cannot settle the question for all workplaces.
“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
How Much Checking Is Needed
The cited research does not establish a single review cost for AI-generated work. The METR estimate of about 9% came from screen recordings for a subset of tasks, while the reported acceptance and line-reading figures came from the study’s particular developers and tools. How review time varies by job, employee experience, and AI system remains unsettled by these examples.
It is also unclear how well the findings carry over to other professions, newer models, or routine use over longer periods. The consulting study’s result outside AI’s measured capability frontier concerned one complex task, not every task where AI may struggle. Nor do the studies establish whether workplace training or different review processes would reduce errors or recover time.
Feigh’s commentary makes a broader case about responsibility and verification, but it is not itself a controlled test of how organizations should allocate review work. The available evidence supports measuring both gains and checking costs; it does not show that AI always increases or always reduces productivity.
Measure the Full Work Cycle
The practical next step suggested by the evidence is to evaluate completed work, not just AI’s first draft. Organizations testing these tools can track time spent generating, checking, correcting, and finishing work, alongside quality and error rates. Comparisons should use tasks that reflect the actual job and identify where the system is likely to be reliable.
Feigh argues that people who remain accountable for AI-assisted work need enough time and training to verify it. The cited studies do not set a universal review standard or say which checks should be required in each profession. Further workplace research will be needed to determine how tools, task design, and reviewer expertise affect both speed and accuracy.
For now, the evidence points to a conditional conclusion: AI can accelerate some work, but its net value depends on what follows generation. Whether a particular deployment saves time remains a question for measurement in that setting.
Key Questions
Did AI make developers faster in the METR trial?
No. In METR’s randomized trial of 16 experienced open-source developers working on 246 tasks, completion took 19% longer when AI was allowed. The result applies to the study’s participants, tools, and tasks.
Did the other studies also find that AI slowed work?
No. A writing experiment found participants completed tasks 40% faster with ChatGPT, while a consulting study found AI users were 25.1% faster on tasks within the system’s capabilities. Results differed by task and study.
What does “verification burden” mean?
Feigh uses the term for the time and mental effort spent checking whether AI output is accurate and appropriate. It describes her interpretation of a workplace challenge, not a standardized metric.
Why can AI output be risky if it reads well?
Polished language does not guarantee correctness. In one experiment with 2,784 reviewers, participants correctly identified only 31% of errors requiring knowledge of task rules, even though they generally caught typos.
What should employers measure when adopting AI?
The cited evidence supports measuring the full work cycle: time to generate, review, correct, and finish a task, as well as quality and errors. The studies do not establish one measurement rule for every workplace.
Source: ThorstenMeyerAI.com
