🔍 Read the full analysis: AI And The Review Burden: What Research Says About Checking Its Work on ThorstenMeyerAI.com
Get smart everyday buys delivered free with Prime
- Fast, free delivery on millions of items
- Prime Video, Amazon Music and more included
- Member-only deals all year
TL;DR
Research reviewed by Georgia Tech cognitive engineer Karen Feigh finds that AI can save time on some tasks while adding review work and increasing the risk of errors on others. A randomized trial found experienced developers took 19% longer with AI, while separate experiments recorded faster work on tasks suited to AI and weaker accuracy on a task outside its capabilities.
Research on workplace AI finds that faster first drafts do not always mean faster completed work: experienced software developers in one randomized trial took 19% longer when they could use AI, while other experiments found time savings on tasks suited to the tools. The findings were highlighted November 16 by Karen Feigh, a cognitive engineer at Georgia Tech, who describes the time and mental effort required to check AI output as a “verification burden.”
The developer trial, run by the research group METR, involved 16 experienced open-source developers completing 246 tasks in large software projects. Tasks were randomly assigned to allow or disallow AI. In the study setting, participants mainly used Cursor Pro and Anthropic’s Claude 3.5 or 3.7 Sonnet. The researchers reported that developers with AI access took 19% longer on average, counting work after the initial code was generated.
Participants’ expectations did not match the measured result. Before the trial, developers predicted AI would reduce their completion time by 24%; afterwards, they estimated it had saved them 20%. METR reported that developers accepted fewer than 44% of AI-generated code suggestions. In a survey, 75% said they read every line of generated code, while 56% said they often had to make major changes to clean it up.
Those figures describe a particular group and task setting, not every workplace use of AI. On a subset of 44 tasks with usable screen recordings, METR estimated that reviewing and cleaning generated output took about 9% of AI-allowed working time; another 4% went to waiting for the tool. The researchers cautioned that the result does not show current AI is unhelpful in other economically relevant settings or establish what future systems will do.
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
Why Review Time Changes the Productivity Calculation
The findings matter because organizations often judge AI by how quickly it produces a draft or suggestion. The studies point to a broader measure: the time to produce accurate, usable work, including checking, correction and any consequences of a missed error. A tool can shorten one stage and still leave the overall workflow unchanged or slower.
Results also vary with the task. In a preregistered experiment with 758 knowledge workers working with Boston Consulting Group, participants using AI completed 25.1% more quickly, on average, on 18 tasks within the tools’ tested capabilities; they also completed 12.2% more tasks, with higher-quality solutions. On a separate complex task selected as outside those capabilities, AI users were 19% less likely to produce a correct solution. These results describe those experiments, not a universal performance rate.
That difference places a practical burden on people who remain responsible for the work. If output can be fluent but wrong, review requires more than proofreading: it may require checking facts, applying domain rules and deciding whether the tool was suitable for the task. Employers evaluating productivity need to account for that labor and give reviewers time, relevant expertise and clear accountability.
Products Worth Considering

Apple Mac mini M6, 12-core CPU and GPU, 24GB Memory, 1TB SSD

Logitech MX Mechanical Wireless Illuminated Keyboard, Tactile Quiet, Graphite

Apple Mac mini (2018), 3.0 GHz 6-Core Intel Core i5, 32 GB RAM, 256 GB SSD
As an affiliate, we earn on qualifying purchases.
Studies Behind the Verification Burden
Feigh’s essay in The Conversation is an expert commentary, not a new experiment. It connects workplace concerns to a group of existing studies, including research on coding, consulting and professional writing. The source material also cites a 2023 study by Shakked Noy and Whitney Zhang: on mid-level professional writing tasks, ChatGPT reduced completion time by 40% and increased assessed quality by 18%. That result applies to the tasks and measures used in that study.
Evidence about review quality raises another concern. In an experiment involving 2,784 people checking AI-style suggestions, participants caught typos most of the time, but achieved 31% accuracy on errors that required understanding the task’s rules. The result suggests that errors needing substantive judgment may be harder to identify than surface mistakes; it does not establish a single error-detection rate for all AI users or jobs.
Feigh frames the accountability problem through earlier work by cognitive engineer David Woods, who described a “responsibility-authority double bind” in 1985: automated systems can exercise practical influence while people retain responsibility. Her analogy is that AI can resemble a fast, polished intern whose confident delivery does not signal which parts need scrutiny. That is Feigh’s interpretation of the research, rather than a measured outcome from one study.
“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
Where the Findings May Not Apply
The studies cover different occupations, tasks, tools and measures, so they do not establish a single productivity effect for workplace AI. The METR trial involved 16 experienced developers working in mature open-source projects and tools available between February and June 2025. Its result may not carry over to less complex coding, other software, newer systems or different review practices.
The reported 9% review estimate came from a subset of 44 issues with valid labeled screen recordings, not from every task in the trial. The available findings also do not show how review burden changes with experience, training, organizational safeguards or repeated use. It remains unclear how often workers miss consequential errors in ordinary workplaces, and how often checks catch them before they cause harm.
Nor do the study summaries establish that the gains and losses persist as tools change. The consulting experiment shows contrasting outcomes across tasks, but does not by itself provide a general test for predicting whether a new task falls inside a model’s capabilities. Broader, task-specific evidence is still needed before drawing conclusions about overall productivity across an organization.
Measure the Full Work Cycle
The practical next step suggested by the evidence is to measure completed work rather than the speed of the first AI-generated response. Employers and researchers can track time spent prompting, checking, correcting and validating output, alongside its accuracy and quality. Those measures should identify the task and tool used so comparisons do not blur unlike work.
Organizations also need to decide who checks AI-assisted work and whether that person has enough time and subject knowledge to do so. Feigh’s central recommendation is to support the workers who remain accountable with training and time for verification. The studies reviewed do not establish one review process that fits every job, so checks will need to match the consequences and complexity of the task.
Further research can test whether newer tools and different workflows change the balance seen in the developer trial, and which types of errors people are most likely to miss. Until then, the evidence supports a measured conclusion: AI can speed some tasks, but its net effect depends on the work and the cost of checking its output.
Key Questions
Did the METR study find that AI always slows developers down?
No. It found that 16 experienced developers took 19% longer on average across 246 tasks in the study’s particular setting. METR said the result does not show AI is unhelpful in other settings or predict the effect of future systems.
What does “verification burden” mean?
It is Karen Feigh’s term for the time and mental effort spent checking whether AI output is accurate and appropriate. That work can include reading, testing, correcting or rejecting generated material.
Can AI still make work faster?
Yes. In a consulting experiment, AI users completed tasks within the tools’ tested capabilities 25.1% faster on average. Results differed on a task outside those capabilities, where AI users were less likely to produce a correct answer.
Why can AI errors be difficult to catch?
Some mistakes require understanding the task’s rules rather than spotting a typo. In one experiment with 2,784 participants, accuracy on those rule-dependent errors was 31%; that finding is specific to the experiment, not a general rate for all reviewers.
How should organizations evaluate AI productivity?
Measure the full workflow, including review, correction, accuracy and quality, rather than only how quickly a tool produces an initial draft. The studies do not identify one approach that works for every task.
Source: ThorstenMeyerAI.com
