The Human Effort Behind AI At Work: Checking What It Produces
AIThis post was created with the assistance of artificial intelligence (AI).

🔍 Read the full analysis: The Human Effort Behind AI At Work: Checking What It Produces on ThorstenMeyerAI.com

Before you orderOffer from Amazon

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
Start your free Prime trial Free trial for eligible customers · Cancel anytime
As an affiliate, we earn on qualifying purchases.

TL;DR

Cognitive engineer Karen Feigh argues that workers can spend time and mental effort checking AI output, a burden that may offset productivity gains. Studies show mixed results: AI slowed experienced developers in one trial, helped consultants on tasks within its abilities, and made them less accurate on a task outside them.

Georgia Tech cognitive engineer Karen Feigh argued in a November 16 essay in The Conversation that checking AI output can consume enough time and mental effort to offset some of its productivity gains. Research she cited shows sharply different results by task: a trial found experienced software developers took 19% longer with AI tools, while a separate consulting experiment found faster work on tasks within AI’s capabilities and lower accuracy on a task outside them.

Feigh calls the work of determining whether AI output is accurate and appropriate “the verification burden.” She describes workers as supervising a fast, polished intern whose output can be useful but may contain errors, while the human remains responsible for what is delivered. Her essay is an expert perspective, not a new experiment; the evidence comes from separate studies in different work settings.

The METR randomized trial involved 16 experienced open-source developers completing 246 tasks in large software projects. Tasks were assigned to allow or disallow AI tools. Developers using AI took 19% longer on average, despite expecting before the trial that AI would save them 24% of their time. Afterward, they still estimated that it had saved 20%. The result is specific to the study’s participants, projects and tools; it does not establish that AI slows all software development.

METR reported that developers accepted fewer than 44% of AI code suggestions; 75% said they read every line of generated code, and 56% said they often had to make major changes to clean it up. In screen recordings covering a subset of 44 issues, about 9% of AI-allowed working time went to reviewing and cleaning AI output. Those figures describe the study’s sample and measured tasks, not a universal share of time spent checking AI.

At a glance
reportWhen: Feigh’s essay was published November 16…
The developmentA November 16 essay by Georgia Tech cognitive engineer Karen Feigh has drawn attention to research showing that checking AI-generated work can add time and that productivity effects vary by task.

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

Checking Can Change the Productivity Picture

The findings matter because a quick first draft is not the same as a completed, reliable piece of work. If employees must test, correct or reject generated material, measuring only how quickly AI produces an answer can overstate the time saved. METR counted work through completion, including what happened after code review, and found that checking and revising were part of the overall task.

Other evidence cautions against treating that result as a general verdict on workplace AI. In a study with Boston Consulting Group involving 758 knowledge workers, participants using AI completed 12.2% more tasks and worked 25.1% faster on average across 18 tasks within the tools’ capabilities, with higher-quality solutions. On a separate task set outside those capabilities, AI users were 19% less likely to produce a correct solution. The contrast suggests that task fit and the ability to recognize mistakes can affect whether AI helps.

For employers, the practical question is not only whether AI can generate work quickly, but who checks it, how long review takes and whether reviewers can spot consequential errors. Feigh’s argument is that the person accountable for an output may need time and training to verify it, even when a tool appears confident.

Products Worth Considering

Samsung 32-Inch Movingstyle Essential 4K UHD Smart Monitor with Rolling Stand

Samsung 32-Inch Movingstyle Essential 4K UHD Smart Monitor with Rolling Stand

Canon PIXMA MegaTank G3270 Wireless All-in-One Color Inkjet Printer

Canon PIXMA MegaTank G3270 Wireless All-in-One Color Inkjet Printer

Apple MacBook Air 15-inch Laptop with M5 Chip, 16GB Memory, 512GB SSD, Liquid Retina Display, and Touch ID (Midnight)

Apple MacBook Air 15-inch Laptop with M5 Chip, 16GB Memory, 512GB SSD, Liquid Retina Display, and Touch ID (Midnight)

As an affiliate, we earn on qualifying purchases.

Evidence Varies by Work and Task

Feigh’s essay draws together research rather than reporting a single new workplace trial. The studies measure different activities and use different designs, so their results should not be combined into one estimate of AI’s effect. The METR trial tested experienced developers working on real tasks in mature repositories with tools available in early 2025. Its findings apply to that setting and do not rule out gains from other tools, users or kinds of software work.

In another line of research cited by Feigh, a study of mid-level professional writing tasks found that ChatGPT cut completion time by 40% and increased assessed quality by 18%. That result concerns those writing tasks, not all office work. Across these examples, AI’s effects differ according to what the work requires and whether the system can perform it reliably.

Feigh also points to an experiment involving 2,784 people asked to check AI-style suggestions. Participants usually caught typos, but achieved only 31% accuracy on errors that required understanding the task’s rules. The result raises a distinct concern: reviewing output is not automatically effective just because a human is present. The study’s measure concerns that experiment and should not be read as a general error-detection rate for workers.

“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

Where the Evidence Does Not Reach

The studies do not establish one overall productivity effect for AI across occupations. They cover different tasks, participants and tools, and the METR trial used a small group of experienced developers. Its authors said the result does not show that current AI systems are unhelpful in other economically relevant settings or that future tools cannot speed up work in the same setting.

It also remains unclear how often workers catch serious errors in everyday use, how much verification time varies across jobs, and whether training or changes to AI systems would improve review. The 31% accuracy result comes from a specific error-checking experiment; it is not a measure of all human review. Feigh’s essay makes a case for attention to verification, but does not itself supply a new estimate of its cost across the workforce.

Measure the Whole Work Process

The research points toward evaluating AI across the full workflow: generating, checking, revising and delivering the final work. Employers and researchers can compare task completion time and quality, while recording how much effort goes into reviewing output and whether errors are caught. Results should be reported by task and setting rather than treated as a single productivity number.

Feigh’s recommendation is to give workers who remain accountable for AI-assisted work the time and training needed to check it. The next question for organizations is whether those practices can preserve gains on suitable tasks while reducing mistakes when AI output is unreliable. The cited studies do not yet provide a universal answer.

Key Questions

What does Feigh mean by the verification burden?

She means the time and mental effort workers spend deciding whether AI output is accurate and appropriate, then checking or correcting it.

Did AI make the developers in the METR trial faster?

No. The 16 experienced developers took 19% longer on average with AI allowed in the trial’s 246 software tasks. The finding is limited to that study’s participants, tasks and tools.

Did AI help the consultants in the other study?

On 18 tasks judged to be within AI’s capabilities, participants using AI completed more tasks and were faster on average. On a separate task outside those capabilities, they were 19% less likely to be correct.

Does the research show that AI always adds review work?

No. The studies report different outcomes across software, consulting and writing tasks. The evidence does not establish a single effect for every job or show that every AI-assisted task requires the same amount of review.

What should employers measure?

They can track the whole task, including review, correction, completion time and quality, rather than counting only how quickly AI generates an initial answer.

Source: ThorstenMeyerAI.com

This content is for general information only and is not financial, tax or legal advice. Consult a qualified professional for decisions about your money.
You May Also Like

A Guide To The Fed’s 87% AI Risk Survey Result And Australia’s Test Plan

The Fed says 87% of surveyed market contacts cited AI risk. Australia’s APRA plans to test AI-driven disruption in its 2027 system stress test.

A Reality Check On AI: 50 Signs Of Superintelligence And The AGI Question

A review of 50 proposed signs of AI superintelligence finds major definition gaps, mixed evidence and no verified arrival of ASI.

Grok AI Privacy Lawsuit: SpaceXAI Class Action Raises Illinois Law Claims

Top Class Actions reports a class action alleging Grok violates Illinois privacy law. The plaintiffs, court, statute and case status are not confirmed.

China Sphere Capability Gap, Q2 2026 Update: Five Labs, Five Strategies, One Narrowing Frontier

Chinese labs shipped five frontier-tier models within four weeks in April 2026, narrowing the capability gap with US leaders. This shapes AI deployment strategies worldwide.