A Closer Look At The Cost Of Checking AI At Work
AIThis post was created with the assistance of artificial intelligence (AI).

🔍 Read the full analysis: A Closer Look At The Cost Of Checking AI At Work on ThorstenMeyerAI.com

Before you orderOffer from Amazon

Get hardware and tech essentials 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 highlighted by Georgia Tech cognitive engineer Karen Feigh warns that checking AI output can offset some productivity gains. A randomized trial of experienced developers found tasks took 19% longer with AI, while other studies found faster work on suitable tasks and weaker accuracy on work outside AI’s capabilities.

Georgia Tech cognitive engineer Karen Feigh warned in a November 16 essay that the time workers spend checking AI output can counter its productivity gains, a concern underscored by a randomized trial in which 16 experienced developers took 19% longer to complete tasks when AI tools were allowed. The findings show that AI’s effect on work depends on the task and on the effort required to verify the result.

Feigh described the extra effort of judging whether AI output is accurate and appropriate as the “verification burden.” Her essay in The Conversation draws on existing research and professional experience; it is an expert commentary, not a new experiment. Feigh said an executive assistant told her that after catching inaccuracies and bias in AI-generated itineraries and briefings, she now sets aside time to check them.

The most direct evidence on review time comes from a METR randomized trial involving 16 experienced open-source developers and 246 real tasks in large software projects. Tasks were randomly assigned to allow or disallow AI. In the February-to-June 2025 study, participants mainly used Cursor Pro and versions of Anthropic’s Claude. With AI allowed, developers took 19% longer on average, including the work done after reviewing the code.

METR reported that developers accepted fewer than 44% of AI code generations. In a survey, 75% said they read every line of AI-generated code, and 56% said they often needed major changes to clean it up. For 44 tasks with usable screen recordings, about 9% of AI-allowed working time went to reviewing and cleaning output; another 4% went to waiting for the tool. These figures describe that study’s participants and tasks, not a universal workplace cost.

At a glance
reportWhen: Feigh’s essay was published November 16…
The developmentA November 16 essay by Karen Feigh draws attention to research showing that AI’s workplace speed gains can be offset by the time and effort workers spend checking its output.
Crypto market snapshot
Fear & Greed Index
12/100 — Extreme Fear
Bitcoin BTC$65,058▲ 1.4%
Ethereum ETH$1,823▲ 1.2%
Tether USDT$0.9992▼ 0.0%
BNB BNB$594.03▲ 1.2%
USDC USDC$0.9996▼ 0.0%
XRP XRP$0.9742▲ 2.5%
Solana SOL$73.95▲ 4.4%
TRON TRX$0.3059▲ 0.8%
Live data · CoinGecko · alternative.me (24h change)

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

Review Time Can Erase Speed Gains

The research complicates the common way of measuring workplace AI: how quickly a tool produces a first draft. If workers must spend additional time checking facts, correcting errors, or adapting code, that initial speed does not necessarily translate into faster completion. METR’s trial measured the full task, and found slower completion in its specific software-development setting.

Other studies show the effect can run in the opposite direction. In research on professional writing tasks, ChatGPT reduced completion time by 40% and increased quality by 18%. In a preregistered experiment with 758 knowledge workers and Boston Consulting Group, AI users completed tasks judged to be within its capabilities 25.1% faster and completed 12.2% more tasks. On a complex task outside those capabilities, however, AI users were 19% less likely to give a correct answer.

For employers, the implication is that productivity assessments need to count review and correction, not just drafting. Workers remain accountable for many AI-assisted outputs, while the systems may present correct and incorrect material in a similarly confident style. That gap can affect quality, deadlines, and the amount of training and review time teams need.

Products Worth Considering

Apple AirPods Pro 3 Wireless Earbuds with Active Noise Cancellation

Apple AirPods Pro 3 Wireless Earbuds with Active Noise Cancellation

Bose Smart Ultra Dolby Atmos Soundbar with Voice Control, HDMI eARC, Black

Bose Smart Ultra Dolby Atmos Soundbar with Voice Control, HDMI eARC, Black

Bose Portable Smart Speaker with Alexa and Google Assistant, Black

Bose Portable Smart Speaker with Alexa and Google Assistant, Black

As an affiliate, we earn on qualifying purchases.

What the Studies Actually Tested

The studies cited by Feigh cover different tasks, participants, and measures, so their results should not be treated as a single estimate of AI’s effect. The METR experiment examined experienced developers working in repositories they knew well. It used completion time for real tasks and recorded how participants worked, rather than measuring only how quickly AI generated code.

The consulting experiment, published in Organization Science, tested the idea of a “jagged technological frontier”: AI may help on tasks within its capabilities but perform poorly beyond them. Its results showed faster, higher-quality work on 18 tasks inside that frontier, alongside lower correctness on one deliberately challenging task outside it. A separate study of professional writing, published in Science in 2023, measured time and quality in mid-level writing tasks and found gains with ChatGPT.

Feigh also cites 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. That result suggests checking fluent output can require more than proofreading; the reviewer may need to know the subject and recognize when an answer violates its requirements.

“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

How Broadly the Results Apply

The findings do not establish a single net productivity effect for AI across workplaces. The METR result applies to a small group of experienced open-source developers using particular tools on particular tasks in 2025. It does not show that AI slows every developer, or predict how newer tools will perform in other settings.

The writing and consulting studies measured different work and used different definitions of speed, quality, and correctness. The available material does not quantify the overall cost of verification across occupations, nor establish how often workers catch consequential errors in day-to-day use. It also remains unclear how results change with additional training, different review processes, or later AI systems.

Measure the Whole Work Process

Feigh’s recommendation is to measure the whole workflow, from the first AI-generated draft through review, correction, and final delivery. That approach can show whether a tool saves time on a task overall, rather than only making one step faster.

Organizations adopting AI will also need to decide who checks its output and how much time and subject-matter training that review requires. The research cited here does not set a common standard for those decisions. Further evidence across different jobs, tools, and review practices will be needed to determine where speed gains persist after verification.

Key Questions

What does “verification burden” mean?

It is the time and mental effort spent checking whether AI output is accurate and appropriate, as described by Georgia Tech cognitive engineer Karen Feigh.

Did the developer study find that AI always slows work?

No. The METR trial found that 16 experienced developers took 19% longer on average across 246 tasks when AI was allowed. The result is specific to those participants, tools, and software tasks.

Do other studies show productivity gains?

Yes. A professional-writing study reported that ChatGPT cut task time by 40% and raised quality by 18%. In a consulting experiment, AI users completed tasks within AI’s capabilities 25.1% faster, while performing worse on a task outside those capabilities.

Why can checking AI output be difficult?

AI can present errors in fluent language, and spotting them may require knowledge of the task rather than proofreading alone. In an experiment with 2,784 people, accuracy was 31% on errors that required understanding task rules.

What should employers measure when evaluating AI?

They should measure the full task, including review, correction, and final quality, rather than counting only how quickly AI produces a draft. Feigh also argues that workers responsible for AI-assisted output need time and training to check it.

Source: ThorstenMeyerAI.com

Nothing in this article is financial or investment advice. Cryptocurrency and precious-metal investments carry significant risk — do your own research and consider a licensed advisor.
You May Also Like

AmenGate: The Moment Before The Scroll

AmenGate introduces a faith-based phone lock that interrupts distractions with prayer, aiming to foster meaningful digital habits rooted in trust and tradition.

Which Laptops Are Leading AI Innovation For Content Creators In 2026?

Discover the laptops at the forefront of AI-driven content creation in 2026, including key features, performance, and what makes them ideal for creators.

Technology Operations Signal Monitor: How Google Helped Destroy Adoption Of RSS Feeds (2023)

New analysis shows Google’s platform and tooling changes contributed to the decline of RSS feed usage, impacting small software companies’ product decisions.

Grimfaste: Operations for a Fleet

Grimfaste introduces a new control plane designed to manage and monitor large publishing fleets, addressing operational bottlenecks and revenue leaks.