🔍 Read the full analysis: AI’s Workplace Impact Includes The Effort Of Checking Its Work on ThorstenMeyerAI.com
Get business pricing on monitors, keyboards and dev gear
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
TL;DR
A November 16 essay by Georgia Tech cognitive engineer Karen Feigh highlights the time and mental effort workers spend checking AI output. Research she discusses finds mixed results: AI sped some writing and consulting tasks, but a randomized trial found experienced developers took 19% longer when AI tools were allowed.
AI can make workplace tasks faster, but checking its output can consume time and attention, Georgia Tech cognitive engineer Karen Feigh argued in a November 16 essay in The Conversation. Research she cites reports both productivity gains and cases where AI users took longer or made more errors, underscoring that results depend on the task and the work needed to verify the answer.
Feigh calls the checking work a “verification burden”: the effort workers spend deciding whether AI output is accurate and appropriate, then reviewing or correcting it. Her essay is an expert commentary drawing on existing research, not a new experiment. She describes AI as resembling a fast, polished intern whose work may be useful but whose confident presentation does not reliably signal where errors are.
One of the studies she discusses, a randomized trial by research group METR, involved 16 experienced open-source developers completing 246 tasks in established software projects. When AI tools were allowed, developers took 19% longer on average. The tools used in the trial included Cursor Pro and Anthropic’s Claude 3.5 and 3.7 Sonnet. Researchers measured completion time, including work after reviewing the code, rather than just how quickly an initial draft appeared.
The trial also found a gap between perceptions and measured results. Developers had expected AI to reduce their completion time by 24% and, after participating, still estimated a 20% reduction. METR reported that developers accepted fewer than 44% of generated suggestions; 75% said they read every line of AI-generated code, while 56% said they often had to make major changes. In a subset of 44 tasks with usable screen recordings, roughly 9% of working time went to reviewing and cleaning AI 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 Checking Can Erase Time Savings
The findings matter because organizations often judge AI tools by how quickly they produce a first draft or suggestion. That measure can miss the review, testing and correction required before the output is ready to use. Where workers remain responsible for errors, faster generation does not necessarily mean faster completion.
Other research cited by Feigh shows why a single verdict on AI productivity would be misleading. In a study of mid-level professional writing tasks, ChatGPT users completed work 40% faster and received an 18% increase in quality. In a Boston Consulting Group experiment, participants using AI completed tasks within its capabilities 25.1% faster on average and completed 12.2% more tasks. On a task designed to fall outside AI’s capabilities, however, AI users were 19% less likely to produce a correct solution than participants without AI.
For employers, the practical issue is not just whether a tool can produce useful work. It is whether teams can identify when its output needs scrutiny, and whether their time estimates include that scrutiny. Feigh’s argument is that workers asked to supervise AI should have the time and training to check it, particularly when errors carry consequences.
Products Worth Considering

Bose QuietComfort Headphones (2nd Gen), Wireless Noise-Cancelling Over-Ear Headphones, White Smoke

Kindle Paperwhite Signature Edition 32GB (2024 Model) – Metallic Black

Bose QuietComfort Ultra Wireless Noise Cancelling Earbuds, 2nd Gen, with 20W 2-Port Wall Charger
As an affiliate, we earn on qualifying purchases.
Studies Show Uneven Results
The research does not measure one universal effect of workplace AI. It covers different tasks, workers and tools. The METR trial focused on experienced developers working in large, familiar software repositories. Its findings apply to that tested setting; METR said they do not show that AI is unhelpful in other economically relevant work or that future systems could not speed up similar tasks.
The Boston Consulting Group experiment involved 758 knowledge workers and examined 18 realistic tasks judged to be within AI’s capabilities, as well as a complex task outside that range. The results illustrate how performance can change with task difficulty and fit. A separate experiment Feigh cites, involving 2,784 people checking AI-style suggestions, found that participants often caught typos but achieved only 31% accuracy on errors requiring an understanding of the task’s rules.
Feigh also points to earlier writing research by Noy and Zhang, published in Science in 2023. Taken together, these studies provide evidence for gains in some settings and costs or risks in others. They do not establish that every worker faces the same verification burden or that the burden always outweighs the time saved.
“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, in The Conversation
How Broad Is the Burden?
The available findings do not establish how often verification cancels out AI’s gains across occupations. The studies use particular tasks, participants and tools, so their results should not be treated as a single productivity estimate for all workers or workplaces.
It is also unclear how the balance will change as AI systems and workplace practices evolve. The METR trial tested specific tools during a defined period, while the cited research covers different kinds of work. The studies do not settle how much review time is appropriate for each use, how reliably workers can spot subtle errors, or how much training changes checking accuracy.
Measure the Full Work Cycle
Feigh’s essay points toward measuring the whole workflow: time spent prompting, reviewing, testing, correcting and finalizing work, as well as the quality of the result. For organizations evaluating AI, that means comparing completed tasks rather than relying only on the speed of a generated draft.
Further research will be needed to establish how those costs and gains vary by role, experience and task. For now, the cited evidence supports a limited conclusion: AI can improve performance in some settings, but workers may still need to check its work, and that labor belongs in any assessment of productivity.
Key Questions
What does “verification burden” mean?
It is the time and mental effort workers spend checking whether AI output is accurate and appropriate, then correcting it when needed.
Did AI slow all workers in the studies?
No. Results varied by task. In the METR trial, experienced developers took 19% longer with AI tools allowed. Other studies found faster completion and improved quality on some writing and consulting tasks.
What did the METR developer trial measure?
It measured task completion time for 16 experienced open-source developers working on 246 real tasks, with AI allowed for some randomly assigned tasks and disallowed for others.
Why can AI help on one task but hurt on another?
The cited consulting experiment found gains on tasks within the tested AI capabilities but lower accuracy on a task outside them. The results suggest that task fit matters; they do not show that every task has the same outcome.
What should employers measure when evaluating AI?
They should track the full work cycle, including review and correction time, and assess the quality of finished work rather than only the speed of AI-generated drafts.
Source: ThorstenMeyerAI.com
