🔍 Read the full analysis: How Reviewing AI’s Work Adds A Human Step To The Workday on ThorstenMeyerAI.com
Get business pricing on office and shipping supplies
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
TL;DR
Cognitive engineer Karen Feigh argues that reviewing AI output can offset some of the time saved by generating it. Studies show mixed results: experienced developers in one trial took longer with AI, while other experiments found faster work on tasks suited to AI. The findings support measuring the full workflow, including review, rather than draft speed alone.
Georgia Tech cognitive engineer Karen Feigh argued in a November 16 article in The Conversation that workers can spend time and mental effort checking AI-generated material, offsetting some of its apparent speed gains. The issue, which Feigh calls the “verification burden,” is reflected in mixed research findings: one randomized trial found experienced software developers took longer with AI, while other studies found faster work on tasks within AI’s capabilities.
In a randomized trial by research group METR, 16 experienced open-source developers worked on 246 tasks in large software projects. Tasks were randomly assigned to allow or disallow AI tools. Developers took 19% longer on average when AI was allowed. The trial used tools including Cursor Pro and Anthropic’s Claude 3.5 and 3.7 Sonnet. Its result measured task completion, including work after the AI generated code, rather than only the time required to produce an initial draft.
The trial also found a gap between measured time and developers’ expectations. Participants had forecast a 24% time saving before the study and, after taking part, estimated that AI had saved them 20%. METR reported that developers accepted fewer than 44% of the 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, researchers estimated that reviewing and cleaning AI output took about 9% of working time; waiting for the AI accounted for another 4%.
The results differ across settings. In a preregistered experiment with 758 Boston Consulting Group knowledge workers, participants using AI completed 25.1% more quickly on average across 18 tasks within the technology’s capabilities, while completing 12.2% more tasks and producing higher-quality solutions, according to the study authors. On a separate complex task outside those capabilities, AI users were 19% less likely to produce a correct solution. A 2023 study of professional writing tasks also found that ChatGPT reduced completion time by 40% and increased measured quality by 18%.
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 Changes the Productivity Picture
These findings matter because an AI tool’s speed at producing a draft does not by itself show whether it saves time across a whole job. Workers may need to check facts, test code, correct errors, or decide whether an answer fits the task. Those steps can absorb time saved earlier, and their scale may differ by occupation, task difficulty, and the AI system being used.
Accuracy also depends on what reviewers are asked to catch. In an experiment involving 2,784 people checking AI-style suggestions, participants usually caught typos, but achieved only 31% accuracy on errors requiring an understanding of the task’s rules. That result points to a potential weakness in relying on surface-level proofreading alone. It does not show that all workers or review processes perform at that rate, but it suggests that fluent wording can make substantive mistakes harder to identify.
For employers, the practical question is not simply whether AI generates work faster. It is whether the total process improves once review and correction are included. Feigh’s argument also raises a responsibility issue: people may remain accountable for work produced with tools whose reasoning and weaknesses are not always visible to them.
Products Worth Considering

Apple Mac mini (2018), 3.0 GHz 6-Core Intel Core i5, 32 GB RAM, 256 GB SSD

Logitech MX Creative Console with 9 Customizable Keys, Graphite

Logitech MX Master 4 Ergonomic Wireless Mouse with Haptics – Graphite
As an affiliate, we earn on qualifying purchases.
Evidence Varies Across Work Tasks
Feigh’s article is an expert analysis, not a new experiment. It connects research on AI-assisted work with a longstanding concern in automation: systems can exercise practical influence over a task while a human remains responsible for the outcome. Feigh refers to this as a responsibility-authority double bind, drawing on cognitive engineer David Woods’s work from 1985.
The studies she discusses measure different tasks, workers, tools, and outcomes, so their results should not be treated as a single estimate of AI’s effect on productivity. The METR trial focused on experienced developers working in mature repositories, while the consulting experiment tested a range of knowledge-work tasks and included one designed to fall outside AI’s abilities. The writing study measured time and quality on professional writing tasks. Together, they show that effects can depend on whether a task is a good fit for the tool and how much human checking it requires.
“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 Much Review Each Job Requires
The research does not establish a single verification cost for all AI-assisted work. METR’s coding result came from a small group of experienced developers and particular tools tested from February to June 2025; it does not establish that AI will slow developers in every setting or with later systems. The study’s authors said their findings did not mean current AI tools are useless in economically relevant work or rule out future improvements in the same setting.
The amount of checking needed in everyday workplaces also remains uncertain. The cited studies cover selected tasks and conditions, and their results do not establish how often AI-generated errors go unnoticed in routine use. Feigh’s argument identifies a concern, but the available figures do not measure the total review burden across every occupation or show whether organizations consistently give reviewers enough time and training.
Track the Full Work Process
The evidence points toward evaluating AI by the full workflow: time spent prompting and generating, followed by review, correction, testing, and any rework. Employers and researchers can compare those steps with the same work done without AI, while tracking accuracy and quality as well as speed. A faster first draft is not, on its own, proof of a faster or better result.
Feigh’s recommendation is to give people who remain accountable for AI-assisted work the time and training to check it. Further research across more tasks and work settings will be needed to show when review costs outweigh time savings, when AI delivers a net gain, and how those effects change as tools develop.
Key Questions
What does “verification burden” mean?
It is the time and mental effort workers spend checking whether AI-generated output is accurate and appropriate for the task, as described by Karen Feigh.
Did the METR trial show 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 that trial. METR said the result does not rule out AI being useful in other settings or future tools producing different results.
Can AI make professional work faster?
Some studies found faster work on certain tasks. For example, a BCG experiment found AI users completed tasks within the tested capability range 25.1% faster on average. Results varied when the task fell outside that range.
Why is checking AI output difficult?
Some mistakes are not obvious from wording alone and require knowledge of the task’s rules. In one experiment with 2,784 people, accuracy was 31% on errors requiring that understanding, even though participants usually caught typos.
What should employers measure when evaluating AI?
They should assess the whole workflow, including generation time, review, corrections, testing, and quality—not just how quickly the first draft appears.
Source: ThorstenMeyerAI.com
