What Studies Reveal About The Cost Of Reviewing AI-Generated Work
AIThis post was created with the assistance of artificial intelligence (AI).

🔍 Read the full analysis: What Studies Reveal About The Cost Of Reviewing AI-Generated 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 on AI-assisted work finds that productivity effects vary by task: some experiments report faster, higher-quality work, while a randomized trial of experienced developers found they took 19% longer when AI tools were allowed. The studies point to review and correction as part of the workload, but do not establish a single cost of checking AI output across jobs.

Research on AI-assisted work shows that reviewing and correcting outputs can absorb some productivity gains, but the effect differs sharply across tasks. In a November 16 article in The Conversation, Georgia Tech cognitive engineer Karen Feigh describes this checking work as a “verification burden”, drawing on studies that range from faster professional writing to a randomized software trial in which developers took longer with AI tools.

The clearest evidence of a time cost comes from a randomized trial by METR, a research group focused on AI capabilities and their effects. Sixteen experienced open-source developers completed 246 real tasks in large software projects. Tasks were randomly assigned to conditions allowing or disallowing AI tools. In the study, developers using AI took an average of 19% longer to complete tasks.

The trial measured completion time through the work after code review, not just how quickly an initial draft appeared. Developers accepted fewer than 44% of AI-generated code suggestions, according to METR. 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. On a subset of 44 tasks with usable screen recordings, about 9% of AI-allowed work time went to reviewing and cleaning generated output; another 4% went to waiting for the AI.

Other studies report gains under different conditions. A study of mid-level professional writing tasks found that ChatGPT users completed work 40% faster and produced work rated 18% higher in quality. In a preregistered experiment involving 758 knowledge workers and Boston Consulting Group, participants using AI completed tasks within the system’s capabilities 25.1% faster on average and completed 12.2% more tasks. On one complex task outside those capabilities, however, AI users were 19% less likely to produce a correct solution.

At a glance
reportWhen: Feigh’s article was published November…
The developmentA November 16 article by Georgia Tech cognitive engineer Karen Feigh brings together research showing that AI’s time savings can be offset by verification work, with results varying by task and setting.
Crypto market snapshot
Fear & Greed Index
12/100 — Extreme Fear
Bitcoin BTC$64,991▲ 1.2%
Ethereum ETH$1,822▲ 1.2%
Tether USDT$0.9992▼ 0.0%
BNB BNB$593.76▲ 1.1%
USDC USDC$0.9996▼ 0.0%
XRP XRP$0.9709▲ 2.1%
Solana SOL$73.6▲ 4.0%
TRON TRX$0.3055▲ 0.7%
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 Changes the Productivity Picture

These findings show why measuring only the time needed to produce an AI-assisted first draft can give an incomplete account of workplace productivity. Workers may still need to check facts, test code, revise weak answers and decide whether a polished response is suitable for its intended use. That work can add time, and an error that survives review can carry consequences beyond a missed efficiency target.

The studies do not establish that AI slows work overall. Some show faster or better results, and METR cautions that its developer trial does not mean current AI systems are unhelpful in other economically relevant settings. The practical lesson is narrower: the effect depends on the task and the full workflow. Employers evaluating AI may need to count review and correction alongside drafting speed, and account for who has the time and expertise to verify outputs.

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 UltraGear 32GX870A-B 32-inch 4K OLED Dual-Mode Gaming Monitor

LG UltraGear 32GX870A-B 32-inch 4K OLED Dual-Mode Gaming Monitor

As an affiliate, we earn on qualifying purchases.

Different Tasks, Different Results

Feigh’s article is an expert commentary, not a new experiment. She links research on verification to a broader workplace problem: AI systems can generate plausible answers quickly while leaving people responsible for deciding whether those answers are accurate and appropriate. Feigh describes the arrangement as similar to supervising a fast, polished intern whose work can be useful but also requires scrutiny.

The studies she cites measure different jobs, tools and outcomes, so their results are not directly interchangeable. METR’s trial involved experienced developers working in mature open-source repositories with tools including Cursor Pro and versions of Anthropic’s Claude Sonnet available during the study period. The consulting experiment tested defined tasks and distinguished those within AI’s demonstrated abilities from one outside them. The writing study measured time and quality on a particular set of professional writing assignments. Taken together, they show variation, not one universal productivity rate.

METR also found a gap between expectations and measured performance: participating developers expected AI to reduce task time by 24% before the trial and estimated a 20% reduction afterward, despite taking longer on average in the AI condition. Those figures describe participants’ expectations and later perceptions in that experiment; they are not estimates for all workers or current AI systems.

“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 studies do not provide a single estimate of how much time workers spend reviewing AI output across the economy. Their samples, tasks, tools and measures differ, and the METR trial involved only 16 experienced developers working on particular open-source projects. Its reported 9% review-and-cleaning figure came from a subset of tasks with usable recordings, not every task in the trial.

It is also unclear how results would change with different AI models, user experience, workplace procedures or higher-stakes tasks. The experiments identify gains and failures in specified settings; they do not determine how often workers catch consequential errors across routine employment. Feigh’s commentary connects the evidence but is not itself a new study, and the research cited does not quantify a general financial cost for verification.

Measure the Full Work Cycle

The research reviewed here supports evaluating AI across the complete task: generation, checking, correction and final delivery. Organizations can compare that total time and the quality of completed work with workflows that do not use AI, while recording where errors are found and how much rework they cause. Such measurements would help distinguish tasks where AI saves time from those where review absorbs the gain or leaves users exposed to mistakes.

For now, the evidence points to a continuing human role in checking many AI-assisted outputs, especially where correctness matters. Feigh argues that workers who remain accountable need time and training to verify what systems produce. Further studies across more occupations, task types and tools will be needed to establish how common the verification burden is and how its cost changes as AI systems and workplace practices evolve.

Key Questions

Did the METR study find that AI always slows workers down?

No. The randomized trial found that 16 experienced developers took 19% longer on average when AI tools were allowed for the 246 tasks studied. METR said the result does not show that AI is unhelpful in other settings.

How much time did developers spend reviewing AI code?

In a subset of 44 tasks with usable screen recordings, developers spent about 9% of AI-allowed working time reviewing and cleaning AI output. This figure applies to that subset, not necessarily to all tasks or software work generally.

Do other studies find productivity gains from AI?

Yes. A professional writing study reported work completed 40% faster with ChatGPT, while an experiment with Boston Consulting Group found AI users completed tasks within the system’s capabilities 25.1% faster on average. Results changed on a task outside the AI’s capabilities, where users were less likely to be correct.

What does “verification burden” mean?

Feigh uses the term for the time and mental effort people spend checking whether AI output is accurate and appropriate, then correcting or rejecting it when needed.

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

Ultimate Guide To The Best AI Workflow Automation Tools In 2026

Discover the top AI workflow automation tools in 2026, including features, benefits, and how to choose the right solution for your business needs.

Want Better AI Agent Decisions? Try These Four Rules

An October 27 article proposes four rules to keep AI agents from creating more review and management work than they save.

Prevent Eye Strain With A Webcam Blink-Rate Tracker For Better Screen Wellness

A new webcam-based app is being tested to monitor blink rate and promote breaks, helping remote workers combat digital eye strain during long screen hours.

Thrymvault: A System Around Your Content

Thrymvault introduces a private, self-hosted platform that consolidates content creation, management, AI workflows, and client portals into one integrated system.