🔍 Read the full analysis: Checking AI Output: What The Research Could Mean For Your Workplace 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 cited by Georgia Tech cognitive engineer Karen Feigh highlights a verification burden: workers may spend time checking and correcting AI output, even when it arrives quickly. Studies show mixed results, from slower completion in a trial of experienced developers to faster work on some writing and consulting tasks.
Georgia Tech cognitive engineer Karen Feigh argued in a November 16, 2026, essay in The Conversation that checking AI-generated work can offset some of the time saved by producing it. The research she cites does not show a single productivity effect across workplaces: a randomized trial found experienced developers took longer when AI tools were allowed, while other experiments found faster completion on tasks suited to AI.
Feigh calls the time and mental effort spent deciding whether AI output is accurate and appropriate the “verification burden.” In her account, workers remain responsible for work produced with AI, even when the system’s confident presentation offers little indication of where errors may lie. She likens the role to supervising a fast, polished intern whose work can be useful but still needs review.
The most detailed example is a randomized trial by the research group METR, involving 16 experienced open-source developers working on 246 real tasks in large software projects. In the study, which used tools including Cursor Pro and versions of Claude 3.5 and 3.7 Sonnet, developers took 19% longer on average when AI tools were allowed. The result measures task completion, including work after reviewing generated code, rather than only the time needed to produce an initial draft.
METR reported that developers accepted fewer than 44% of AI code generations. In a survey, 75% said they read every line of generated code, and 56% said they often had to make major changes to clean it up. Screen recordings from a subset of tasks indicated that about 9% of working time went to reviewing and cleaning AI output. The sample and setting are limited; METR said its result does not establish that AI is unhelpful in other economically relevant work.
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
For employers, the findings suggest that a fast first draft is not the same as a faster completed task. If workers must inspect, test, correct, or replace generated material, those steps belong in any credible measure of productivity. Counting only time to first output can make a tool appear more effective than it is across the full workflow.
The risk is not limited to lost time. In a study of 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. That result points to a distinction between proofreading and substantive verification: polished language or plausible code may still contain errors that require domain knowledge to detect.
That does not mean workers should reject AI tools. It means organizations need to match tools and review procedures to the work, and give accountable staff time and training to check outputs. The studies indicate that the balance between speed and checking can differ sharply by task and by how well the task fits the system’s capabilities.
Products Worth Considering

Bose Smart Ultra Dolby Atmos Soundbar with Voice Control, HDMI eARC, Black

Logitech MX Keys S Combo Wireless Keyboard and Mouse, Graphite

LG 27GS60QC-B Ultragear 27-inch Curved Gaming Monitor (QHD, 180Hz, 1ms, 1000R)
As an affiliate, we earn on qualifying purchases.
Why Results Vary by Task
Other research cited by Feigh found gains in settings where AI performed well. A study of mid-level professional writing tasks reported that ChatGPT reduced completion time by 40% and increased output quality by 18%. Those figures describe that study’s particular tasks and participants, not a general workplace effect.
A preregistered experiment with Boston Consulting Group involving 758 knowledge workers tested tasks inside and outside what researchers called AI’s “jagged technological frontier.” On 18 tasks within the frontier, participants using AI completed 12.2% more tasks and finished 25.1% faster on average, with higher-quality solutions. On one complex task placed outside the frontier, AI users were 19% less likely to produce a correct solution. The researchers’ results, later published in Organization Science, show why performance on one task may not predict performance on another.
Feigh’s essay is an expert perspective, not a new experiment. It brings together earlier research and applies it to workplace oversight. The studies use different populations, tools, tasks, and measures, so their results should not be treated as directly comparable or as a universal estimate of AI’s effect on productivity.
“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 Has Limits
The available studies do not establish how much verification time is typical across jobs, industries, or current AI tools. The developer trial involved a small group of experienced contributors working in mature open-source repositories, and its tools reflect the period from February to June 2025. Results may differ with other users, software, tasks, or later systems.
Nor do the studies settle whether AI leads to net gains for an organization after accounting for review, training, software costs, and the quality of completed work. The findings vary by task and by the measure used. Feigh’s essay draws attention to that variation, but it is not itself a new workplace-wide measurement.
Measure the Full Work Process
For organizations adopting AI, the practical next step suggested by the research is to measure the whole workflow: initial generation, checking, corrections, and final quality. Employers can compare results across specific tasks rather than assume that faster drafting means shorter completion time. They should also make clear who is accountable for AI-assisted work and give those workers the time and training to review it.
Further studies across more roles and tasks will be needed to establish how often verification offsets speed gains and which review practices work best. Until then, the evidence supports a task-by-task approach, not a blanket conclusion that AI either raises or lowers workplace productivity.
Key Questions
Does the research show that AI makes workers less productive?
No. Results vary by task. METR’s trial found experienced developers took 19% longer with AI allowed, while writing and consulting experiments found faster completion on certain tasks. These studies do not establish one effect across all workplaces.
What does “verification burden” mean?
Karen Feigh uses the term for the time and mental effort workers spend checking whether AI output is accurate and appropriate, and correcting it when needed.
How large was the developer study?
METR’s randomized trial involved 16 experienced open-source developers completing 246 tasks. The result applies to that study setting and should not be treated as a universal estimate.
Why did the consulting experiment find both benefits and risks?
Researchers reported faster, higher-quality work on 18 tasks within AI’s tested capabilities, but lower accuracy on one complex task outside those capabilities. The findings show that performance can depend on the task.
What should employers measure when adopting AI?
They should track the full process, including generation, review, correction, completion time, and final quality, rather than measuring only how quickly AI produces a first draft.
Source: ThorstenMeyerAI.com
