🔍 Read the full analysis: The Hidden Cost Of AI Assistance: Reviewing What It Gets Done 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 argues that AI productivity gains can be offset by the time and mental effort workers spend checking results. Research she cites shows mixed outcomes: AI slowed experienced developers in one randomized trial, while other studies found faster work on tasks suited to AI. The evidence does not establish a universal productivity penalty.
Georgia Tech cognitive engineer Karen Feigh argued November 16 that workers’ time spent checking AI-generated work can offset some of the speed gains the tools promise. The studies she cites show the effect varies by task: a randomized trial found experienced software developers took 19% longer when AI tools were allowed, while experiments in other settings found faster work when tasks fell within AI’s capabilities.
Feigh calls the effort spent determining whether AI output is accurate and appropriate the “verification burden.” In an essay published by The Conversation, she describes workers as supervising a fast, polished assistant whose answers can mix useful material with errors, while the human remains accountable for the result. The essay is an expert commentary, not a new experiment.
The clearest example she discusses is a randomized trial by research group METR. Sixteen experienced open-source developers completed 246 tasks in large software projects, with each task assigned to allow or disallow AI. In the study, conducted using tools including Cursor and Claude models available from February to June 2025, developers took 19% longer on average when AI was allowed. That measure included work after the initial code was produced.
The developers’ expectations did not match the measured result. Before the trial, they expected AI to reduce their completion time by 24%; after it, they still estimated that it had saved them 20%. METR reported that developers accepted fewer than 44% of AI code generations, and 75% said they read every line of AI-generated code. On screen-recorded tasks with valid labels, about 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
Checking Can Erase Some Speed Gains
The findings matter because the speed of an AI-generated draft is not the same as the time needed to complete a reliable piece of work. If staff must inspect, test, correct, and take responsibility for AI output, organizations that measure only first-draft speed may overstate the productivity benefit.
That burden is not limited to code. Feigh cites an experiment involving 2,784 people checking AI-style suggestions: participants usually caught typos, but accuracy fell to 31% for errors requiring an understanding of the task’s rules. The result suggests that polished presentation alone does not make difficult errors easy to spot. It does not, by itself, show how often such errors occur in everyday workplaces.
For workers, the practical issue is whether employers allow enough time and provide suitable training for review. For organizations, the cited evidence supports measuring the whole workflow—including corrections and checking—rather than treating faster output as proof of net time saved.
Products Worth Considering

DJI Mini 5 Pro Fly More Combo with DJI RC 2, 1-Inch CMOS

Sonos Roam (Gen 1) Portable Smart Speaker, White

Apple Watch SE 3 GPS 40mm Smartwatch
As an affiliate, we earn on qualifying purchases.
Results Depend on the Task
The studies do not support a single conclusion that AI always speeds work or always slows it. In a 2023 study of professional writing tasks, researchers found ChatGPT users completed work 40% faster and produced work rated 18% higher in quality. Those results concern the tasks and participants in that experiment; they should not be assumed to apply to every occupation.
A separate experiment with 758 knowledge workers, conducted with Boston Consulting Group and published in Organization Science in 2026, found that AI users completed 18 tasks within the tools’ assessed capability range 25.1% faster on average and completed 12.2% more tasks. On a complex task outside that range, however, AI users were 19% less likely to produce a correct answer than participants without AI.
METR also cautioned that its developer trial does not show that current AI systems are unhelpful in all economically relevant settings, or that future tools could not improve performance in the same setting. Its result is tied to a particular group, set of mature software projects, and tools available during the trial.
“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, writing in The Conversation
How Broad Is the Burden?
The studies cited here measure different tasks, participants, tools, and outcomes, so they do not establish one average effect for AI across workplaces. The METR trial involved 16 experienced developers working in established open-source projects; its findings may not transfer directly to less experienced programmers, other software work, or newer AI systems.
It is also unclear how frequently the reported review time translates into a net cost in ordinary organizations. The cited experiments do not settle how results vary with training, task design, access to source material, or changes in workplace review policies. The 9% figure comes from a subset of tasks with valid screen-recording labels, not every task in the trial.
Feigh’s essay presents an interpretation of existing research, rather than new measured data. Whether a specific business gains time or quality from AI will depend on the work being done and on the time required to verify the output.
Measure the Full Work Cycle
The next step for employers is to evaluate AI against the full process: drafting, checking, correcting, and completing the work. Comparisons should track both time and accuracy, and distinguish tasks where the tool performs reliably from tasks where errors are harder to detect.
Feigh’s recommendations focus on the people who remain accountable for AI-assisted work: give them time and training to check outputs, rather than assuming that a fast response is a finished result. The source material does not identify a new trial or policy deadline; further evidence will be needed to show how these effects change across occupations and as tools develop.
Key Questions
Did AI make developers slower in every case?
No. In the METR trial, 16 experienced open-source developers took 19% longer on average when AI was allowed across the study’s tasks. That result applies to the trial setting, not every developer or software task.
What does “verification burden” mean?
It is Feigh’s term for the time and mental effort spent checking whether AI output is accurate and appropriate before relying on it.
Did the cited studies find that AI always reduces productivity?
No. Results differed by task. A professional-writing study found faster work with ChatGPT, and a consulting experiment found gains on tasks within AI’s assessed capabilities. The METR developer trial found slower completion in its setting.
Why can polished AI output still need careful review?
AI-generated work can include errors that are difficult to recognize without understanding the task’s rules. In one experiment cited by Feigh, accuracy in spotting those rule-dependent errors was 31%, despite participants often catching typos.
What should employers measure when evaluating AI tools?
They should measure the full workflow, including review, testing, corrections, completion time, and quality, rather than relying only on how quickly an initial draft appears.
Source: ThorstenMeyerAI.com
