AIThis post was created with the assistance of artificial intelligence (AI).

🔍 Read the full analysis: What Does Reviewing AI Output Cost At Work? A Look At The Studies on ThorstenMeyerAI.com

Buying for a business?Offer from Amazon

Get business pricing on tech for your team

  • Business-only prices and quantity discounts
  • Tax-exempt purchasing
  • Multiple users, one account, clear invoices
As an affiliate, we earn on qualifying purchases.

TL;DR

Studies of workplace AI show that review can offset speed gains, but the effect varies across tasks. In a randomized trial, experienced developers took 19% longer with AI, while experiments in professional writing and consulting found faster or better results on suitable tasks. The research does not establish one review cost that applies across jobs.

A review of workplace AI research published by cognitive engineer Karen Feigh of Georgia Tech argues that checking AI output can absorb some of the time and mental effort saved by generating it. The studies she cites show sharply different results by task: experienced developers took 19% longer with AI in one randomized trial, while other experiments found faster work when tasks suited the tools.

Feigh described this checking work as a “verification burden” in an article published in The Conversation on November 16, 2026. Her piece is an expert commentary, not a new experiment. It draws on existing studies to argue that productivity assessments should include the time spent reviewing, testing and correcting AI output, rather than counting only how quickly a first draft appears.

The most direct evidence comes from a METR randomized trial involving 16 experienced open-source developers and 246 tasks in large, mature software projects. When AI tools were allowed, participants took 19% longer on average than when they worked without them. The trial ran from February to June 2025 and mainly involved Cursor Pro and Anthropic’s Claude 3.5 and 3.7 Sonnet. Its result is specific to those participants, tools and tasks.

METR reported that developers accepted fewer than 44% of AI code generations; 75% said they read every line of AI-generated code, and 56% said they often had to make major changes. In a subset of 44 tasks with usable screen recordings, about 9% of AI-allowed working time went to reviewing and cleaning AI output, while another 4% went to waiting for the tool. Those figures describe that subset, not a universal share of AI work.

At a glance
reportWhen: Research reviewed November 17, 2026; ci…
The developmentA review of workplace AI studies highlights evidence that checking AI output can consume time and attention, sometimes erasing measured productivity gains.

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 Review Time Changes the Productivity Measure

The findings matter to employers weighing whether AI is making work more productive. A fast draft is not the same as a completed task: workers may need to verify facts, test code, repair errors or adapt material to requirements. If that work is omitted from measurement, organizations can overstate time savings and miss the labor needed to make AI-assisted work usable.

But the evidence does not support treating review as a fixed tax on every AI task. In a study of mid-level professional writing, ChatGPT reduced completion time by 40% and increased quality by 18%. In a Boston Consulting Group experiment, participants using AI completed tasks within the tools’ capabilities 25.1% faster on average and completed 12.2% more tasks. Those results show that review and productivity effects depend on what workers are asked to do, as well as whether the system can handle it.

For workers, managers and policymakers, the practical implication is to measure the whole workflow: generation, review, correction and final quality. People who remain accountable for AI-assisted work need enough time and relevant training to check it, rather than pressure to accept a polished answer at face value.

Products Worth Considering

Bose TV Speaker (2nd Gen) Compact Soundbar with Bluetooth and HDMI Connectivity

Bose TV Speaker (2nd Gen) Compact Soundbar with Bluetooth and HDMI Connectivity

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

As an affiliate, we earn on qualifying purchases.

Different Tasks, Different Study Results

The studies reviewed do not measure one common workplace or one standard definition of productivity. METR studied experienced software developers completing real tasks in established open-source projects. The writing experiment measured performance on mid-level professional writing tasks. The consulting research examined knowledge workers working through a set of realistic business tasks, including one deliberately placed beyond the AI system’s capabilities.

In the Boston Consulting Group experiment, researchers reported that AI users performed better on 18 tasks within the tools’ “jagged” capability frontier. On a complex task outside that frontier, however, AI users were 19% less likely to produce a correct answer than participants without AI. That result underlines a central limit: fluency does not reliably tell a worker whether an output is correct.

A separate experiment involving 2,784 people checking AI-style suggestions found that participants usually caught typos but achieved only 31% accuracy on errors requiring them to understand the task’s rules. The result suggests that obvious surface mistakes may be easier to spot than substantive errors. It does not establish how accurately workers in every profession review AI outputs.

“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 Has Limits

The studies do not establish a single average cost for reviewing AI output across the workforce. They differ in task design, tools, participants and measures, and some test controlled assignments rather than routine work over months or years. The METR trial’s 19% slowdown should not be read as a forecast for every developer, employer or AI system.

It is also unclear how results change as tools improve, workers gain experience, or organizations redesign jobs around AI. The METR researchers said their findings do not rule out productivity gains in other settings or gains from future models. The figures on review time come from a limited subset of recorded tasks, while the BCG result outside the AI capability frontier concerned a particular complex assignment.

Finally, the available evidence does not show that workers consistently detect the errors that matter most. The experiment on suggestion checking found low accuracy for rule-dependent errors, but the supplied results do not establish how that performance translates to specific occupations or real-world consequences.

Measure Work Beyond the First Draft

The studies point to a practical next step for employers: assess AI against the full task, including review time, correction rates and final quality. Teams should track whether work is genuinely completed faster, not only how quickly AI generates an initial answer. Results should be separated by task type, since the cited experiments found both gains and losses.

Organizations also need to give employees who approve or rely on AI output the time and training to check it. The research reviewed here does not set a standard review process or predict when AI will produce net savings across an entire workplace. Further studies of ordinary work over longer periods could help clarify how tool improvements, worker experience and job design affect those outcomes.

Key Questions

Does reviewing AI output always erase its time savings?

No. The studies show different results across tasks. Developers in the METR trial took longer with AI, while professional writing and consulting experiments found faster work on tasks suited to the tools.

What did the METR developer trial find?

In a randomized trial of 16 experienced open-source developers completing 246 tasks, participants took 19% longer on average when AI tools were allowed. The result applies to that study’s participants, tasks and tools.

How much time did developers spend checking 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. That subset is not a universal estimate for software work.

Can workers reliably spot AI errors?

Not always, according to one experiment involving 2,784 people. Participants usually caught typos, but accuracy was 31% for errors requiring understanding of task rules. The study does not establish the error-detection rate for every occupation.

What should employers measure when evaluating workplace AI?

They should include generation, review, correction and final quality, and compare results across task types. Measuring only how fast an AI draft appears can leave out work needed to make it accurate and appropriate.

Source: ThorstenMeyerAI.com

You May Also Like

My Software North Star

A developer on Hacker News shares their guiding philosophy for software development, emphasizing utility, correctness, maintainability, and user love.

Exploring SenseTime’s Top Position In International Vision AI Rankings

SenseTime reports leading global rankings in three vision AI categories, but details and independent verification are not yet available.

Boost Your Mixing With The Best AI-Integrated Studio Headphones Of 2026

Discover the best AI-powered studio headphones of 2026 for mixing, tracking, and mastering. Expert-reviewed picks for every budget and workflow.

How Anthropic’s Message Threads Speed Up Claude Code Without Cutting Token Costs

The New Stack reports message threads are speeding up Claude Code, while token costs remain unchanged. Performance and availability details are not provided.