The Hidden Cost Of Checking AI’s Work: What Studies Suggest For The Workplace
AIThis post was created with the assistance of artificial intelligence (AI).

🔍 Read the full analysis: The Hidden Cost Of Checking AI’s Work: What Studies Suggest For The Workplace on ThorstenMeyerAI.com

Before you orderOffer from Amazon

Get movie nights delivered free with Prime

  • Fast, free delivery on millions of items
  • Prime Video, Amazon Music and more included
  • Member-only deals all year
Start your free Prime trial Free trial for eligible customers · Cancel anytime
As an affiliate, we earn on qualifying purchases.

TL;DR

Research reviewed by Georgia Tech cognitive engineer Karen Feigh points to a workplace verification burden: workers must spend time checking AI output while remaining responsible for it. Results vary by task: one trial found experienced developers took 19% longer with AI, while other studies found faster work on tasks within AI’s capabilities.

Research on workplace AI finds that faster drafting does not always mean faster completion: checking and correcting AI output can consume time, and some workers may miss errors that require deeper knowledge of a task. In a November 16 article in The Conversation, Georgia Tech cognitive engineer Karen Feigh calls this work the “verification burden,” drawing on studies that show results differ by task and setting.

The clearest example comes from a randomized trial by METR, a research group that studied 16 experienced open-source developers completing 246 real tasks in large software projects. Tasks were randomly assigned to allow or disallow AI. In that study, developers using AI tools took 19% longer on average. The researchers said the measured time included work after code review, rather than counting only the creation of an initial draft.

The developers’ expectations did not match the measured result. Before the trial, they expected AI to reduce completion time by 24%; after using it, they still estimated that it had saved them 20%. METR reported that developers accepted fewer than 44% of AI code suggestions. In survey responses, 75% said they read every line of generated code, while 56% said they often had to make major changes to clean it up. In screen recordings for a subset of 44 tasks, about 9% of AI-allowed working time went to reviewing and cleaning AI output.

Other research found productivity gains in different settings. A study of professional writing tasks reported that ChatGPT users completed work 40% faster and produced work rated 18% higher in quality. In a preregistered experiment involving 758 knowledge workers and Boston Consulting Group, participants using AI completed tasks within the tools’ capabilities 25.1% faster on average and completed 12.2% more tasks. On a complex task designed to fall outside those capabilities, however, AI users were 19% less likely to produce a correct solution.

At a glance
reportWhen: Feigh’s article was published November…
The developmentA November 16 article by Georgia Tech’s Karen Feigh highlights research showing that the time and attention needed to check AI output can offset productivity gains.

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)

AI draft Checking Fixing Re-checking What a speed metric sees What the work costs

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

  1. DraftThe AI answers in seconds, fluent and confident“few clues about where scrutiny is needed”
  2. CheckDecide what to check and how, without knowing if or where an error iscognitive load
  3. FixFinding an error does not end the work, it starts a correction cycletime
  4. Re-checkEvery fix needs checking again; when nothing is wrong, the effort feels wastedmotivation
  5. Count it allProductivity is the whole human-AI workflow, not the first draftnet, not gross
The real measure: “whether the entire human-AI workflow produces trustworthy work at a lower total cost”

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

These findings complicate claims that AI adoption automatically improves workplace productivity. A tool may produce an answer quickly, but the relevant measure for an employer is whether the whole task takes less time and produces work that is accurate enough to use. Review, correction and testing belong in that calculation, alongside drafting speed.

There is also an accountability issue. Feigh argues that workers remain responsible for AI-assisted work even when the system produces much of it. If employees are expected to catch errors without adequate time, training or authority to reject the output, an apparent productivity gain may come with greater risk. The studies do not establish one universal cost for verification; they show that it depends on the task, the worker and the AI’s capabilities.

Ninja Creami Scoop & Swirl Ice Cream Maker – Bone & Champagne (Amazon Exclusive)

Ninja Creami Scoop & Swirl Ice Cream Maker – Bone & Champagne (Amazon Exclusive)

Bose QuietComfort Ultra Wireless Noise Cancelling Earbuds, 2nd Gen, with 20W 2-Port Wall Charger

Bose QuietComfort Ultra Wireless Noise Cancelling Earbuds, 2nd Gen, with 20W 2-Port Wall Charger

As an affiliate, we earn on qualifying purchases.

Different Tasks, Different Results

Feigh’s Conversation article is an expert analysis, not a new experiment. It brings together existing research, including METR’s software trial and studies of writing and consulting tasks. The METR trial used Cursor Pro and versions of Anthropic’s Claude 3.5 and 3.7 Sonnet during a period in 2025. Its result describes that group, those tools and those repository tasks; it should not be read as a finding about every developer or current AI system.

The consulting experiment illustrates why performance can vary within a single study. Participants benefited on tasks the researchers placed within the AI’s capabilities, but performed worse on a task outside that range. Separately, an experiment involving 2,784 people checking AI-style suggestions found that participants usually caught typos but achieved only 31% accuracy on errors that required understanding the task’s rules. That result suggests polished presentation does not guarantee that reviewers will spot substantive mistakes.

Feigh connects this tension to an older concern in automation research: systems can exercise practical influence over decisions while people retain responsibility for the outcome. She describes AI at work as resembling a fast, polished intern whose output can be useful or wrong, sometimes without clear signals about which parts need scrutiny.

“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, writing in The Conversation

How Broadly the Results Apply

The studies cover particular groups, tools and tasks, so they do not establish a single productivity effect for all workplaces. METR’s trial involved 16 experienced developers working in mature open-source projects with specific AI tools available in 2025. The writing and consulting studies used different tasks and measures. The findings cannot be directly combined into one estimate of AI’s overall workplace impact.

It is also unclear how results will change as models, software interfaces and workers’ experience change. METR said its result did not show that AI systems are unhelpful in every economically relevant setting, nor did it rule out future time savings in the same type of work. The research presented here does not determine how much review time organizations should budget, or which checking practices reliably catch difficult errors.

Measure the Full Work Cycle

For employers adopting AI, the studies point to a practical next step: evaluate the complete workflow, including review, correction, testing and the consequences of mistakes, rather than measuring only how quickly a first draft appears. Results should be tracked by task type, since the same tool may help with one assignment and hinder another.

Feigh’s article also argues that workers who remain accountable for AI-assisted output need time and training to check it. Further evidence will be needed to show how those practices affect accuracy and completion time as tools change. For now, the research supports neither a blanket claim that AI saves time nor that it always adds a net cost; it shows that verification is part of the work and can alter the outcome.

Key Questions

Did the METR study find that AI slowed every developer down?

No. It found that 16 experienced developers took 19% longer on average across 246 tasks in that trial. The result is specific to that sample, setting and tools, not a finding about every developer.

Why did AI users spend time checking generated work?

Participants reviewed, tested and modified AI-generated code before deciding whether to use it. METR reported that developers accepted fewer than 44% of suggestions, and screen recordings for a subset of tasks showed about 9% of working time spent reviewing and cleaning output.

Did other studies find productivity gains from AI?

Yes. A professional-writing study reported faster completion and higher quality ratings with ChatGPT. In a Boston Consulting Group experiment, AI users were faster on tasks within the tools’ capabilities, but were less likely to answer a task outside those capabilities correctly.

What should employers measure when evaluating AI?

They should count the whole task cycle: drafting, checking, correction, testing and the cost of errors. The cited research does not establish a universal review-time requirement, so results should be evaluated for the specific tasks and tools in use.

Source: ThorstenMeyerAI.com

You May Also Like

Meet ‘Little Gen Z’

A new poll reveals growing political divergence within Gen Z, especially among men aged 18-22, indicating a split between older and younger members of the generation.

A War Room for Your Next Idea: Inside IdeaClyst

Discover how IdeaClyst provides founders with a local AI-driven war room to validate ideas, assess risks, and make confident decisions efficiently.

How Business Owners Can Build Camera Confidence With A Practice App

IdeaNavigator AI proposes a 30-day test of short camera exercises for business owners who struggle to record video. No app launch is confirmed.

Millions Are Quitting After Lunch—Inside the ‘Silent Afternoon’ Rebellion

Breaking workplace norms, millions are quitting after lunch in the ‘silent afternoon’ rebellion—discover what this means for your organization’s future.