🔍 Read the full analysis: The Cost Of Checking AI: Lessons From Workplace Studies on ThorstenMeyerAI.com
TL;DR
A November 16 essay by Georgia Tech cognitive engineer Karen Feigh highlights workplace studies showing that AI’s apparent speed gains can be offset by time spent checking its output. Results vary by task: a trial found experienced developers took longer with AI, while writing and consulting experiments found gains on tasks suited to the tools.
Georgia Tech cognitive engineer Karen Feigh argued in a November 16 essay that workplace AI’s productivity gains can be offset by the time and mental effort people spend checking its output. The research she cites shows markedly different results across tasks: one randomized software-development trial found experienced developers took longer with AI, while experiments in writing and consulting found faster work when tasks fit the tools’ abilities.
Feigh’s essay, published by The Conversation, calls the work of checking AI output a “verification burden.” She describes a workplace in which a person remains responsible for results even when an AI system produces much of the work. The essay is an expert perspective, not a new experiment; its argument draws on existing studies.
The sharpest counterexample to claims of automatic time savings comes from a randomized METR trial involving 16 experienced open-source developers and 246 tasks in large, established software projects. Developers took an average of 19% longer when AI tools were allowed. Before the trial, they had expected AI to reduce completion time by 24%; afterwards, they still estimated that it had saved them 20% of their time.
METR reported that developers accepted fewer than 44% of AI-generated code suggestions. In a survey, 75% said they read every line of AI-generated code, and 56% said they often had to make major changes to clean it up. For 44 tasks with usable screen recordings, about 9% of working time went to reviewing and cleaning AI output. The trial measured work through review, rather than counting only the initial draft.
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
Checking Can Erase AI’s Time Savings
The findings matter to employers and workers because a fast first draft does not necessarily mean a faster completed task. If people must inspect, test and repair AI output, measuring only how quickly a tool produces a response can miss part of the workload. METR’s results show that this gap can be large enough to leave experienced developers slower overall in the setting studied.
Other experiments caution against treating that result as a general verdict on workplace AI. A 2023 study of mid-level professional writing tasks found that ChatGPT reduced completion time by 40% and increased quality by 18%. In a preregistered experiment with Boston Consulting Group involving 758 knowledge workers, participants using AI completed tasks within the tools’ capabilities 25.1% faster on average and completed 12.2% more tasks. On a complex task outside those capabilities, however, AI users were 19% less likely to produce a correct solution.
The practical implication is to judge AI by the whole workflow and the quality of the final work, not speed alone. The studies do not establish a single productivity effect for every job. They suggest that task type, tool capability and the work required to catch errors all affect the result.
Popular right now

Bose QuietComfort Ultra Earbuds (2nd Gen), Wireless Noise-Cancelling Earbuds, Black

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

Sony WH-1000XM6 Wireless Noise-Canceling Headphones, Black
As an affiliate, we earn on qualifying purchases.
Different Tasks, Different Results
The studies Feigh discusses cover different populations and measures, so their results are not directly interchangeable. The METR trial focused on experienced developers working in familiar, mature repositories, using Cursor Pro and Anthropic’s Claude 3.5 or 3.7 Sonnet during February to June 2025. Its authors said the findings did not show that current AI systems are useless in other economically relevant settings or that future systems could not help in the same work.
The Boston Consulting Group experiment examined knowledge workers doing realistic consulting tasks. Its “jagged frontier” framing describes a uneven boundary: a tool may perform well on some tasks and poorly on others. Separately, the writing study by Shakked Noy and Whitney Zhang, published in Science in 2023, measured mid-level professional writing tasks. Those settings help explain why evidence of gains in one field does not settle what will happen in another.
Feigh also points to research on the difficulty of checking suggestions. In an experiment involving 2,784 people, participants usually caught typos but achieved only 31% accuracy on errors that required understanding the task’s rules. The result indicates that reviewing fluent output may be harder when mistakes are substantive rather than obvious, though it does not by itself quantify checking performance across all workplaces.
„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
How Broadly the Results Apply
The studies do not establish how often AI saves or costs time across all occupations. The METR trial was small and focused on experienced developers in particular repositories; its findings may not describe novice programmers, other software work or newer tools. The writing and consulting experiments used different participants, tasks and measures.
It also remains unclear how checking time and error rates change as workers gain experience, organizations adjust their processes, or AI systems improve. The reported 9% review figure came from screen recordings for a subset of 44 tasks, not every task in the trial. None of these results should be read as a universal estimate of AI’s workplace productivity effect.
Measure the Full Work Cycle
Feigh’s essay points toward a practical test for employers: measure the time required to complete and check a task, alongside the quality and correctness of the final result. That includes time spent reviewing, testing and correcting AI output, rather than treating the initial response as finished work.
Organizations also need to decide who is accountable for the result and give that person time and training to assess it. The cited studies do not set a single standard for workplace adoption. Further evidence across different jobs, task types and tool generations will be needed to show where AI reliably saves time after verification is counted.
Key Questions
Did the METR trial find that AI always slows developers down?
No. It found that 16 experienced open-source developers took 19% longer on average across 246 tasks when AI was allowed in the trial. The researchers said the result does not show that AI is unhelpful in other settings.
Why did the developers take longer with AI?
The trial reported that developers accepted fewer than 44% of AI code suggestions and often reviewed, tested or changed generated code. Those activities added work, although the study does not assign the entire slowdown to one cause.
Do workplace studies show that AI can save time?
Yes. A professional-writing study found a 40% reduction in time and an 18% increase in quality on the tasks tested. The consulting experiment also found faster completion for tasks inside AI’s capabilities, while results were worse on one task outside them.
What is the verification burden?
It is Feigh’s term for the time and mental effort workers spend deciding whether AI output is accurate and appropriate. It includes reviewing and correcting work before a person can rely on it.
Source: ThorstenMeyerAI.com