🔍 Read the full analysis: Checking AI’s Work: A Workplace Question Studies Are Starting To Explore 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 November 16 essay by Georgia Tech cognitive engineer Karen Feigh highlights a workplace challenge researchers are beginning to measure: workers must verify AI output while remaining accountable for it. Studies report both productivity gains and slowdowns, depending on the task; a METR trial found experienced developers took 19% longer when AI tools were allowed. The evidence does not establish one universal verification cost across jobs.
Feigh published the essay in The Conversation on November 16, 2026. She calls the effort of deciding whether an AI response is accurate and appropriate the “verification burden.” The essay is an expert commentary, not a new experiment; it connects findings from existing studies to the practical question of who checks AI-assisted work and how much time that takes.
The strongest quantified slowdown in the material comes from a randomized METR trial involving 16 experienced open-source developers working on 246 tasks in large, established software projects. The study reported that developers took 19% longer on average when AI was allowed. Its measure included work after the initial code was generated, rather than counting only how quickly a first draft appeared.
The trial also found a gap between expectations and measured results. Developers expected AI to reduce their completion time by 24% before the trial and still estimated a 20% reduction afterward. METR reported that developers accepted fewer than 44% of AI code generations; 75% said they read every line of AI-generated code. In recordings for a subset of 44 tasks, about 9% of working time went to reviewing and cleaning AI output. Those figures describe that study’s participants and tasks, not all workers or 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 Checking Changes the Productivity Math
For employers, the findings complicate a simple measure of AI productivity: how quickly a tool produces a draft. A faster first response may not mean a faster completed task if workers need to test, revise, or reject parts of the result. METR’s developer trial illustrates why the whole workflow matters, from generation through review and completion.
Verification also carries responsibility. Feigh argues that employees can remain accountable for work even when AI systems supply parts of it. If a polished answer contains a hidden error, the worker may need the expertise and time to find it. That makes review capacity and training relevant to how organizations introduce AI, alongside the tool’s apparent speed.
The evidence does not show that AI is generally unproductive. It suggests that outcomes depend on the task and whether the tool can handle it reliably. For workers and managers, a key issue is whether promised time savings persist after corrections and quality checks are counted.
Popular right now

Vitamix VX1 Professional-Grade Blender, 64 oz, Black

Vitamix 5200 Blender with 64 oz Container, Black

Hisense E7 Cinema Series 55-Inch Hi-QLED Mini-LED 4K UHD Smart Fire TV
As an affiliate, we earn on qualifying purchases.
Different Tasks, Different Results
Other studies cited in Feigh’s essay report substantial gains in specific settings. Research on mid-level professional writing tasks found that ChatGPT users completed work 40% faster and produced work rated 18% higher in quality. These results concern the tasks and measures in that study; they do not establish the same effects across occupations.
A preregistered experiment with 758 knowledge workers, conducted with Boston Consulting Group, found a split based on AI’s capabilities. On 18 tasks within what the researchers called the “jagged” frontier of AI capability, users completed 12.2% more tasks and worked 25.1% faster on average, with higher-quality solutions. On a complex task outside that frontier, AI users were 19% less likely to produce a correct solution than participants without AI.
Another experiment involving 2,784 people checking AI-style suggestions found that participants usually caught typos but achieved only 31% accuracy on errors requiring an understanding of the task’s rules. The result points to a distinction between spotting surface-level mistakes and judging whether an answer is substantively right.
“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, in The Conversation
The Limits of Current Evidence
The studies do not establish one average verification cost for AI use across workplaces. They cover different occupations, tools, tasks, and measures, so their results cannot be treated as directly interchangeable. The METR findings, for example, apply to a small group of experienced developers working in mature software repositories with specific tools in 2025.
It is also unclear how results would change as AI systems, work practices, and employee experience evolve. METR said its findings do not show that current AI systems are useless in economically relevant settings, nor do they rule out productivity gains in the same setting from future systems. The cited evidence also does not settle how much review time is needed in every job, or which workplace processes best catch errors that require subject-matter knowledge.
Measure Work Through Final Review
The practical implication in Feigh’s essay is to evaluate the complete task workflow, not just the time needed to generate an initial answer. That means counting review, testing, correction, and quality outcomes when assessing whether AI saves time in a particular role.
Employers considering AI-assisted processes will need to decide who is responsible for checking outputs and give those workers time and training to do it. The next useful evidence will come from studies that track completed work and error rates across more jobs and real-world conditions. The material cited here does not identify a single next study or a universal review standard.
Key Questions
What is the “verification burden”?
It is Feigh’s term for the time and mental effort workers spend deciding whether AI-generated output is accurate and appropriate.
Did AI make the developers in the METR trial faster?
No. In that randomized trial, 16 experienced developers took an average of 19% longer on 246 tasks when AI tools were allowed. The finding is specific to the study’s participants, tasks, and tools.
Do studies show that AI always slows work down?
No. Other cited experiments found faster completion and higher quality for some writing and consulting tasks. Results differed when tasks fell outside the AI’s capabilities.
Why can checking AI output be difficult?
Some mistakes are not obvious from wording alone. In one experiment, people caught typos more often than errors requiring them to understand the task’s rules.
What should employers measure when evaluating AI?
They should track the full process, including review, corrections, and final quality, rather than measuring only how quickly the AI produces a draft.
Source: ThorstenMeyerAI.com
