🔍 Read the full analysis: Checking AI’s Work Takes Effort: What Research Says About The Impact 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
Research discussed by Georgia Tech cognitive engineer Karen Feigh highlights the time and mental effort workers spend checking AI-generated work. A randomized trial found experienced developers took 19% longer with AI, while separate studies found faster, higher-quality results on some writing and consulting tasks. The findings vary by task and do not establish a universal productivity penalty.
Research on workplace AI finds that checking generated output can consume time and attention, and that the effect on productivity depends on the task. In a randomized trial involving experienced open-source developers, those allowed to use AI took 19% longer to complete their assigned work, even though they expected the tools to save time. Other experiments found gains in writing and consulting tasks within AI’s demonstrated capabilities.
In a November 16 article in The Conversation, Georgia Tech cognitive engineer Karen Feigh describes this checking work as a “verification burden”: the time and mental energy spent deciding whether AI output is accurate and appropriate. Her article is expert commentary, not a new experiment. It brings together earlier research on how people use and evaluate AI-generated work.
The most detailed example is a METR randomized trial of 16 experienced open-source developers completing 246 real tasks in large software projects. Tasks were randomly assigned to allow or disallow AI. In the study, conducted from February to June 2025, developers mainly used Cursor Pro and Claude 3.5 or 3.7 Sonnet. With AI allowed, developers took 19% longer on average. The researchers measured the overall task, including work after the initial code was generated.
Developers accepted fewer than 44% of AI code generations, and 75% said they read every line of AI-generated code. On a subset of 44 tasks with usable screen recordings, about 9% of AI-allowed working time went to reviewing and cleaning generated output. The study also found different results in other settings: one experiment on mid-level professional writing reported that ChatGPT cut completion time by 40% and raised quality by 18%. In a consulting experiment, AI users completed tasks within its capability range 25.1% faster, but were 19% less likely to get a task outside that range right.
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 AI Productivity Depends on the Task
The findings caution against judging workplace AI by how quickly it produces a first draft. A tool may save time on a bounded task, but workers still need to check whether the output fits the assignment, follows relevant rules and is factually sound. That review can change the net time saved, particularly when the work is complex or the AI’s capabilities are uncertain.
The results also matter for organizations that remain accountable for work produced with AI assistance. Feigh argues that responsibility stays with people even when automated systems shape the work. The experiments do not show that AI is generally unproductive; instead, they show that benefits and risks vary by task, and that relying on fluency or speed alone can give an incomplete picture.
Products Worth Considering

Bose Smart Dolby Atmos Soundbar with Alexa, A.I. Dialogue Mode, and Bluetooth, Black

LG UltraGear 27-inch QHD OLED Gaming Monitor, 240Hz, 0.03ms, G-SYNC Compatible, FreeSync Premium Pro, HDR True Black 400

DJI Mic 3 (2 TX + 1 RX + Charging Case), Wireless Microphone
As an affiliate, we earn on qualifying purchases.
What the Cited Experiments Measured
Feigh’s account draws on research from different fields and settings, so the results should not be treated as one direct comparison. The METR study tracked experienced developers working in mature software repositories, while the writing experiment examined mid-level professional writing tasks. The consulting research involved 758 knowledge workers in a preregistered experiment conducted with Boston Consulting Group.
In that consulting study, researchers reported that AI users completed 12.2% more tasks and finished 25.1% faster on 18 tasks within the technology’s capability range, with higher-quality solutions. On a complex task selected to fall outside that range, AI users were 19% less likely to produce a correct solution. These are results from that experiment, not a general estimate for every workplace or AI system.
A separate experiment involving 2,784 people examined how well reviewers spotted errors in AI-style suggestions. Participants usually caught typos, but achieved just 31% accuracy on errors that required understanding the task’s rules. That result suggests that some mistakes are harder to verify than surface-level flaws, though it does not measure the same workflow or outcome as the developer trial.
““are often countered by time and mental energy spent determining whether the output is accurate and appropriate””
— Karen Feigh, cognitive engineer at Georgia Tech, in The Conversation
Where the Evidence Has Limits
The METR trial involved 16 developers working on particular open-source repositories with specific AI tools available in early 2025. Its authors said the findings do not establish that AI is unhelpful in other economically relevant settings or rule out productivity gains from future systems. The trial also does not establish how often comparable time costs occur across other jobs.
The studies cited use different participants, tasks, tools and measures, so their results cannot be combined into a single estimate of AI’s overall workplace impact. It is also unclear how results would change with different training, review procedures or newer tools. Feigh’s commentary offers an interpretation of existing work; it does not itself test whether the verification burden cancels out productivity gains across an organization.
Measure the Full Work Process
The research points toward evaluating complete workflows rather than counting how quickly AI produces an initial answer. Employers would need to account for review, testing, correction and the consequences of errors, as well as the time saved on drafting or routine work.
Feigh’s recommendation is to give workers who remain responsible for AI-assisted output enough time and training to check it. The studies summarized here do not identify one review method that works for every task. Further research across jobs, tools and levels of expertise will be needed to show when verification costs outweigh gains, and when AI assistance still leaves workers with a net time or quality improvement.
Key Questions
Did the METR study find that AI slows all workers down?
No. It found that 16 experienced open-source developers took 19% longer on average across 246 tasks in the study. The researchers cautioned that this result does not show AI is unhelpful in other settings.
Why can checking AI output take so much time?
Workers may need to read, test or revise generated work and decide whether it is accurate and appropriate. In the METR trial, developers accepted fewer than 44% of AI code generations, and many reported reading every line.
Did other studies find productivity gains?
Yes. A professional-writing experiment reported that ChatGPT cut time by 40% and increased quality by 18%. In a consulting experiment, AI users finished tasks within the system’s capability range 25.1% faster. These results apply to the specific tasks studied.
What is the main uncertainty in the research?
The studies cover different jobs, participants, tools and task types, so they do not establish a single productivity effect for workplace AI. It remains unclear how broadly the findings apply or how review practices and newer systems may change the balance.
Source: ThorstenMeyerAI.com
