🔍 Read the full analysis: The Hidden Labor Of AI: What Reviewing Its Work Means For Employees 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
Research cited in a November 16, 2026, article by Georgia Tech cognitive engineer Karen Feigh highlights the work employees do to verify AI-generated material. A METR trial found experienced developers took 19% longer with AI tools, while other studies found productivity gains on suitable tasks and weaker results beyond AI’s capabilities.
Employees can spend time and mental effort checking AI-generated work, even when the tools produce drafts quickly, Georgia Tech cognitive engineer Karen Feigh argued in The Conversation on November 16, 2026. Research she cites shows that AI’s effect on productivity varies by task: a trial of experienced software developers found they took longer with AI, while studies of other professional work found faster completion on tasks suited to the tools.
Feigh calls the checking work the “verification burden”: AI may produce a fluent answer, but a person still has to judge whether it is accurate and appropriate. She describes workers as supervising “an extraordinarily fast and polished intern” whose output can contain both good and bad work, without taking responsibility for the consequences. Her article is an expert commentary, not a new experiment.
The most detailed example is a randomized METR trial involving 16 experienced open-source developers and 246 tasks in large, mature software projects. Tasks were assigned with or without AI access. In this setting, developers took 19% longer when AI tools were allowed. The tools included Cursor Pro and Anthropic’s Claude 3.5 and 3.7 Sonnet, available during the trial from February to June 2025.
The developers’ expectations did not match the measured result. They predicted a 24% time saving before the trial and still estimated a 20% saving afterwards. METR reported that developers accepted fewer than 44% of AI-generated code suggestions; 75% said they read every line of AI code, and 56% often needed major changes to clean it up. In screen recordings covering 44 issues, about 9% of working time went to reviewing and cleaning AI output. That figure applies to the recorded subset, not all tasks in the trial.
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
Review Time Can Erase Speed Gains
The findings matter because AI’s visible speed—how quickly it produces a draft or code suggestion—does not capture the whole work process. If staff must check, test, revise or reject output, those tasks belong in any assessment of whether a tool saves time. The METR result is specific to its participants, projects and tools; it does not establish that AI slows all developers or workers.
Other studies Feigh cites show why a single productivity claim can mislead. In an experiment on mid-level professional writing tasks, ChatGPT users completed work 40% faster and with quality 18% higher. In research with Boston Consulting Group involving 758 knowledge workers, people using AI completed tasks within its capabilities 25.1% faster on average and completed 12.2% more tasks. On one complex task outside those capabilities, however, AI users were 19% less likely to produce a correct solution.
That uneven performance can shift work rather than simply remove it. Employees may gain time on one task while taking on more responsibility for checking outputs, particularly where errors carry consequences. Feigh’s argument is that organizations should account for that effort and provide workers with time and training to review AI work, rather than judging adoption only by how quickly a first draft appears.
Popular right now

De'Longhi Magnifica Duo Fully Automatic Espresso Machine with Milk Frother, Silver

Dyson Airwrap Co-anda2x Multi-Styler and Dryer (2026 Model)

Roborock Q7 M5+ Robot Vacuum and Mop with Self-Emptying Dock
As an affiliate, we earn on qualifying purchases.
Studies Show Task-Specific Results
Feigh’s article draws together research on AI use at work and earlier ideas about human responsibility for automated systems. She points to cognitive engineer David Woods, who described a responsibility-authority double bind: automation can exercise practical authority while a human remains responsible. The concept helps frame why reviewing AI output matters, but it is not itself a new finding from the studies she discusses.
The evidence comes from different settings and measures. The METR trial tested experienced developers working in their own repositories. The Boston Consulting Group experiment examined knowledge workers performing realistic consulting tasks, while the writing study measured time and quality on mid-level writing work. These results should not be treated as directly comparable or as a single estimate of AI’s effect across occupations.
One cited experiment involving 2,784 people checking AI-style suggestions also found that errors were not equally easy to spot: participants caught typos most of the time, but achieved only 31% accuracy on errors requiring an understanding of task rules. That result suggests polished presentation alone may not make substantive mistakes obvious. It does not establish how often such errors occur in everyday workplaces.
“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
How Often Checking Cancels Gains
The studies do not establish a single, general measure of the time employees spend verifying AI output. The METR estimate that review and cleanup took about 9% of working time comes from screen recordings for 44 issues, a subset of the trial. It should not be read as a universal rate or as a standalone estimate of how much time AI adds.
It also remains unclear how these results apply to other jobs, less experienced workers, different AI products, or newer versions of the tools. METR said its findings do not rule out future systems speeding up developers in the same setting. The consulting research, meanwhile, found better results on tasks within AI’s capabilities and worse accuracy on a task outside them; it does not establish how reliably workers can identify that boundary in routine work.
Feigh’s commentary synthesizes existing research and expert interpretation. It does not provide a new workplace survey measuring how widely the verification burden occurs, how it affects employees’ workload, or whether employers currently give reviewers enough time and training.
Measure the Full Work Process
The practical next step suggested by the evidence is to measure end-to-end task performance: include time spent prompting, reviewing, correcting and testing output, alongside quality and error rates. Results should be broken out by task and experience level rather than treated as one productivity figure. That approach can show whether a tool saves time in a particular workflow or shifts effort to a later stage.
Feigh argues that workers who remain accountable for AI-assisted work need time and training to check it. The studies described do not set a standard for what that training should involve, and no new policy or employer requirement was announced in her article. Further evidence across occupations and tools will be needed to determine how common these costs are and when AI’s gains outweigh them.
Key Questions
What does “verification burden” mean?
It is Feigh’s term for the time and mental effort spent checking AI output for accuracy and appropriateness after a tool produces it.
Did the METR trial show AI makes all workers slower?
No. It found that 16 experienced open-source developers took 19% longer on 246 tasks when AI tools were allowed. The result is limited to that trial’s participants, tasks and tools.
Do studies show AI can improve productivity?
Yes, in some settings. One writing study reported work completed 40% faster with quality 18% higher, and a consulting experiment found faster completion on tasks within AI’s capabilities. Results differed on a task outside those capabilities.
Why can reviewing AI output be difficult?
Fluent output can contain mistakes that require knowledge of the task to detect. In one experiment with 2,784 people, accuracy was 31% when participants checked errors requiring understanding of task rules; the result does not indicate how often those errors arise at work.
What should employers measure when introducing AI?
They should track the full workflow, including review, correction and testing, as well as completion time and quality. Feigh also says workers responsible for AI-assisted output need time and training to check it.
Source: ThorstenMeyerAI.com
