AI At Work: The Time And Effort Behind Reviewing Its Output
AIThis post was created with the assistance of artificial intelligence (AI).

🔍 Read the full analysis: AI At Work: The Time And Effort Behind Reviewing Its Output 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 cited by Georgia Tech cognitive engineer Karen Feigh shows that AI’s speed gains can be offset by the time and attention workers spend checking its output. A METR trial found experienced developers took 19% longer with AI on the tasks studied, while other experiments found faster work on tasks suited to AI and poorer results on a task beyond its capabilities.

Georgia Tech cognitive engineer Karen Feigh argued in a November 16 essay that workers can lose time and mental energy checking AI-generated work, a cost she calls the “verification burden.” Research she cites shows that AI’s effects vary by task: experienced developers in one trial took longer with AI, while other studies found time and quality gains on work better suited to the tools.

The clearest measured slowdown came in a randomized METR trial involving 16 experienced open-source developers completing 246 tasks in large, established software projects. Tasks were randomly assigned to allow or disallow AI tools. In that setting, developers took 19% longer when AI was allowed, according to the researchers. The result covers the full task, including work after the initial AI output, rather than measuring only how quickly code was drafted.

The trial also found a gap between expectations and measured performance. Before the study, developers expected AI to reduce their completion time by 24%; after participating, they still estimated that it had saved them 20%. 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 had to make major changes to clean it up. Screen recordings for a subset of 44 issues indicated that about 9% of working time went to reviewing and cleaning AI output.

Other results point in a different direction. A study of mid-level professional writing tasks found ChatGPT reduced completion time by 40% and raised quality by 18%. In a preregistered Boston Consulting Group experiment involving 758 knowledge workers, AI users completed tasks within the tools’ capabilities 25.1% faster on average and completed 12.2% more tasks. On one task outside those capabilities, however, they were 19% less likely to produce a correct solution than participants without AI.

At a glance
reportWhen: Feigh’s essay was published November 16…
The developmentA November 16 essay by Georgia Tech’s Karen Feigh highlights research showing that AI’s workplace productivity gains can be offset by the effort required to verify its output.
Crypto market snapshot
Fear & Greed Index
12/100 — Extreme Fear
Bitcoin BTC$64,983▲ 1.3%
Ethereum ETH$1,822▲ 1.2%
Tether USDT$0.9993▲ 0.0%
BNB BNB$593.64▲ 1.1%
USDC USDC$0.9996▲ 0.0%
XRP XRP$0.9711▲ 2.2%
Solana SOL$73.6▲ 4.1%
TRON TRX$0.3056▲ 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

Why Verification Changes Productivity

The findings matter because a fast first draft is not the same as a faster completed job. If employees must check facts, test code, correct omissions or assess whether an answer fits the task, those steps belong in any credible measure of productivity. The METR result is specific to its participants, projects and tools; it does not establish that AI slows all software work. It does show why speed estimates based on generation alone can miss substantial follow-up effort.

Verification also affects accountability. Feigh describes workers as supervising a fast, polished intern whose output can contain good and bad work in the same confident voice. The analogy highlights a practical concern: people may remain responsible for decisions even when AI has produced much of the material, while the points that need extra scrutiny are not always obvious. Organizations adopting these tools need to account for review time and give staff the expertise and authority to challenge output.

Errors can be difficult to catch when they require domain knowledge rather than surface proofreading. In an experiment with 2,784 people reviewing AI-style suggestions, participants usually caught typos, but achieved only 31% accuracy on errors requiring an understanding of the task’s rules. That result suggests fluent presentation alone is not a reliable signal of correctness. The experiment does not establish how often such errors occur in every workplace, but it underscores the limits of review that relies on appearance rather than substantive checking.

Products Worth Considering

LG UltraGear 27-inch QHD OLED Gaming Monitor, 240Hz, 0.03ms, G-SYNC Compatible, FreeSync Premium Pro, HDR True Black 400

LG UltraGear 27-inch QHD OLED Gaming Monitor, 240Hz, 0.03ms, G-SYNC Compatible, FreeSync Premium Pro, HDR True Black 400

LG 27G610A-B 27-inch UltraGear QHD (2560 x 1440) IPS Gaming Monitor, 200Hz, 1ms, AMD FreeSync Premium, VESA DisplayHDR 400, HDMI, DisplayPor

LG 27G610A-B 27-inch UltraGear QHD (2560 x 1440) IPS Gaming Monitor, 200Hz, 1ms, AMD FreeSync Premium, VESA DisplayHDR 400, HDMI, DisplayPor

As an affiliate, we earn on qualifying purchases.

What the Cited Studies Tested

Feigh’s piece in The Conversation is an expert essay, not a new study. It brings together research on AI use and human oversight. The evidence spans different tasks and methods, so the findings should not be treated as one universal productivity estimate. The METR trial concerned experienced developers working in mature repositories with tools including Cursor Pro and Anthropic’s Claude 3.5 and 3.7 Sonnet during February to June 2025.

The consulting experiment, led by Fabrizio Dell’Acqua and colleagues and conducted with Boston Consulting Group, compared results on tasks inside and outside what researchers described as AI’s “jagged” capability frontier. Its reported gains apply to the tasks within that boundary; the poorer result came from a separate complex task selected to sit outside it. A writing experiment by Noy and Zhang, published in 2023, examined mid-level professional writing. Different populations, tools and task designs help explain why results differ.

Feigh also connects the current debate to a longstanding issue in automation research: practical control can shift toward a system while responsibility remains with a person. Her point is that the verification burden is not only a question of seconds spent editing. It includes the attention needed to decide whether an output is appropriate and reliable for its intended use.

“AI’s productivity gains are often countered by time and mental energy spent determining whether the output is accurate and appropriate.”

— Karen Feigh, Georgia Tech cognitive engineer, in The Conversation

Where the Evidence Has Limits

The available studies do not establish a single net productivity effect for AI across occupations. The METR trial involved only 16 developers and particular open-source tasks and tools; its findings may not apply to less experienced users, other software projects or newer systems. The consulting results likewise distinguish between tasks within AI’s capabilities and one task outside them, rather than measuring every kind of professional work.

It is also unclear how often workers catch consequential errors in routine workplace use, how much review time organizations currently record, and whether repeated use changes the balance between drafting and checking. Feigh’s essay synthesizes existing evidence but does not report a new workplace survey or experiment. Results from controlled studies and reported experiences cannot by themselves establish the net effect for a particular employer.

Measure the Whole Work Process

The evidence points toward measuring complete workflows rather than counting the time needed to generate a first draft. Employers evaluating AI tools can track review, correction, testing and rework alongside output speed and quality, while comparing similar tasks with and without AI. They also need to identify who is accountable for checking material before it is used.

Further research across occupations, task types and newer AI systems will help clarify where the tools save time and where verification absorbs those gains. Until then, the studies support a task-by-task assessment rather than a blanket claim that AI either raises or lowers productivity. For workers and managers, the immediate issue is whether the time and skill needed to check an output are included in the job’s real cost.

Key Questions

What does “verification burden” mean?

Verification burden is Feigh’s term for the time and mental effort workers spend deciding whether AI output is accurate and appropriate, then correcting it when needed.

Did AI make developers slower in the METR trial?

In a randomized trial of 16 experienced open-source developers completing 246 tasks, participants took 19% longer when AI was allowed. That finding applies to the study’s participants, tools and tasks; it is not a result for all developers or software work.

Do studies show that AI always reduces productivity?

No. A professional-writing study reported faster work and higher quality with ChatGPT, and the Boston Consulting Group experiment found faster completion on tasks within AI’s capabilities. The results vary by task, and the consulting experiment also found lower accuracy on one task outside those capabilities.

Why can checking AI output be difficult?

Some errors require knowledge of a task’s rules, not just proofreading. In one experiment with 2,784 reviewers, accuracy was 31% on errors that required understanding those rules, even though participants usually caught typos.

What should employers measure when evaluating AI?

They should measure the full workflow, including review, corrections, testing and rework, as well as drafting speed, output quality and task completion. The cited studies do not establish one productivity result for every workplace.

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

Forward-Deployed: The Integration Wall, and the Role That Now Pays $700K to Climb It

Forward-Deployed Engineers now command up to $700K in total compensation, transforming enterprise AI deployment and redefining high-paying tech roles.

Meet S1: the Intriguing, Under-$50, Budget Powerhouse From Openai O1.

Learn about S1, OpenAI’s groundbreaking AI model that defies expectations with its low cost—what revolutionary advancements could it bring next?