🔍 Read the full analysis: AI In The Office: The Overlooked Effort Of Quality Checking 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
Research reviewed in a November 16, 2026, article by Georgia Tech cognitive engineer Karen Feigh highlights the time and mental effort workers spend checking AI output. Results vary by task: one trial found experienced developers took 19% longer with AI, while other studies found faster work on tasks suited to AI. The evidence points to a need to measure the full workflow, including review.
Georgia Tech cognitive engineer Karen Feigh says workers’ review of AI-generated work can consume time and mental effort that productivity claims often overlook. In a November 16 article in The Conversation, she calls this work the “verification burden”, citing studies with mixed results: AI sped up some professional tasks, but experienced developers in one randomized trial took 19% longer when AI tools were allowed.
Feigh’s article is an expert commentary, not a new experiment. It brings together research on how people use AI at work and argues that the speed of an initial draft does not show whether the overall task became faster. Workers may still need to inspect, test, correct or reject the output, particularly when its errors are not obvious.
The most detailed example in the source material is a randomized trial by METR, a research group studying AI capabilities. Sixteen experienced open-source developers completed 246 tasks in large software projects. Tasks were randomly assigned to allow or disallow AI. For the AI-allowed work, developers primarily used Cursor Pro and Anthropic’s Claude 3.5 or 3.7 Sonnet, tools available during the study’s February-to-June 2025 period. On average, the developers took 19% longer with AI allowed.
The METR findings also capture a gap between perceived and measured speed. Before the trial, developers expected AI to cut their completion time by 24%; afterwards, they estimated it had saved 20%. The researchers reported that developers accepted fewer than 44% of AI code generations. In a survey, 75% said they read every line of AI-generated code, while 56% said they often had to make major changes to clean it up. On a subset of 44 tasks with usable screen recordings, about 9% of AI-allowed working time went to reviewing and cleaning output.
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
Why Review Time Changes the Productivity Picture
The findings matter to employers because a faster first draft is not the same as a faster completed task. If workers must spend time checking output, correcting errors and taking responsibility for the result, productivity measures that stop at generation can overstate the benefit. The METR result is specific to its participants, tasks and tools; it does not establish that AI slows all software work. It does show why the full workflow should be measured.
Review is also a question of quality, not just time. The source material describes an experiment involving 2,784 people checking AI-style suggestions. Participants caught typos most of the time, but achieved only 31% accuracy on errors that required understanding the task’s rules. That result suggests fluent-looking output may be harder to check when the mistake depends on deeper knowledge. It does not, by itself, establish how often workers miss such errors in ordinary workplaces.
For workers, this can mean added responsibility without a matching reduction in effort. Feigh describes people as supervising a fast, polished intern: useful in some situations, but capable of producing incorrect work in a confident tone. The practical implication is to give employees who remain accountable for AI-supported work time and training to review it, rather than treating checking as invisible overhead.
Products Worth Considering

Apple Watch Series 8 GPS 41mm Smartwatch with Midnight Aluminum Case and Midnight Sport Band (S/M)

Sony WH-1000XM5 Wireless Noise-Canceling Over-Ear Headphones, Black, International Version

Samsung Galaxy Tab A11+ Kids Edition, 128GB, Royal Blue
As an affiliate, we earn on qualifying purchases.
Different Tasks, Different AI Results
The evidence cited by Feigh does not show a uniform productivity effect. In a 2023 study of mid-level professional writing tasks, researchers found that ChatGPT users completed work 40% faster and produced output rated 18% higher in quality. Those results concern the writing tasks tested and should not be assumed to apply to every office role or assignment.
A separate preregistered experiment with 758 knowledge workers, conducted with Boston Consulting Group, tested tasks that varied in how well they matched AI capabilities. On 18 tasks within what researchers called AI’s “frontier,” participants using AI completed 12.2% more tasks and worked 25.1% faster on average, with higher solution quality. On a complex task outside that frontier, AI users were 19% less likely to produce a correct solution than participants without AI.
The studies therefore point to a task-dependent picture: AI can help when work fits its abilities, while less suitable work may bring slower or less accurate results. Feigh also links the issue to an older problem in automation: people may retain responsibility for outcomes even when a tool has substantial influence over the work. Her essay connects this idea to cognitive engineer David Woods’ 1985 discussion of the responsibility-authority double bind.
“AI’s productivity gains are often countered by time and mental energy spent determining whether the output is accurate and appropriate.”
— Karen Feigh, Georgia Tech cognitive engineer, in The Conversation
Where the Evidence Does Not Reach
The studies cover different tasks, participants, tools and periods, so their results cannot be combined into a single estimate of AI’s effect on office productivity. The METR trial involved 16 experienced developers working on established open-source projects, and its tools reflected the period from February to June 2025. It does not settle how current or future systems perform across other workplaces.
The source material also does not provide a broad measure of how much time employees spend verifying AI output across occupations, or how frequently review catches consequential mistakes in routine work. The 31% accuracy figure comes from a particular experiment about rule-dependent errors, not a general measure of workplace checking. The evidence supports scrutiny of review costs, but does not show that AI consistently raises or lowers productivity overall.
Measure the Full Work Cycle
Feigh’s proposed response is to evaluate the whole workflow: time spent prompting, checking, testing and correcting output, alongside speed and quality of the finished work. Employers also need to account for who is responsible for the result and whether those workers have enough time and relevant knowledge to review it.
The studies summarized here do not set a single standard for measuring verification or prescribe a universal workplace policy. Further research across more roles and task types will be needed to show when AI saves time after review, and when its errors create additional work. For now, the clearest next step for organizations is to treat checking as part of AI-assisted work rather than as a cost outside the productivity calculation.
Key Questions
What does “verification burden” mean?
It is the time and mental effort spent checking AI output for accuracy, suitability and errors before using it. Karen Feigh uses the term in her November 16, 2026, article in The Conversation.
Did AI make developers slower in the METR trial?
In that specific randomized trial, 16 experienced developers took 19% longer on average when AI tools were allowed across 246 software tasks. The finding is limited to the study’s participants, tasks and tools; it is not a universal estimate for all developers.
Does the research show AI always reduces productivity?
No. Other studies cited in the source found faster work on professional writing tasks and on knowledge-work tasks within AI’s capabilities. Results differed by task, and the evidence does not support a single productivity claim for all jobs.
Why can checking AI output be difficult?
Some errors are visible, such as typos, while others require knowledge of the task’s rules. In one experiment described in the source material, participants achieved 31% accuracy on rule-dependent errors. That result applies to the experiment, not necessarily to every workplace review process.
What should employers measure when evaluating AI tools?
They should measure the completed task from start to finish, including time spent reviewing and correcting output, as well as its accuracy and quality. The source also recommends allowing workers who remain accountable for AI-supported work time and training to check it.
Source: ThorstenMeyerAI.com
