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

🔍 Read the full analysis: A Workplace Guide To Understanding The Cost Of Checking AI’s Work on ThorstenMeyerAI.com

Before you orderOffer from Amazon

Get audio and creator gear 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

A November 16, 2026, article by Georgia Tech cognitive engineer Karen Feigh highlights the time and attention workers spend checking AI-generated work. Studies she discusses show mixed results: AI accelerated some writing and consulting tasks, while a randomized software trial found experienced developers took longer with AI, and a separate experiment found reviewers often missed errors requiring rule-based judgment.

Georgia Tech cognitive engineer Karen Feigh argued in a November 16 article in The Conversation that workers can lose some of AI’s apparent productivity gains to the time and mental effort required to check its output. The research she cites shows that impact depends on the task: some experiments found faster work with AI, while a randomized trial of experienced software developers found they took 19% longer when AI tools were available.

Feigh calls the effort to judge whether AI output is accurate and appropriate the “verification burden.” Her article is an expert analysis, not a new experiment. It draws on published research and describes a workplace challenge: AI can produce fluent, quick answers, but people remain responsible for deciding whether those answers are usable.

The software trial, conducted by the research group METR, randomly assigned 16 experienced open-source developers to work with or without AI on 246 tasks in mature software projects. Across the trial, developers took 19% longer with AI allowed. Beforehand, they predicted AI would reduce their completion time by 24%; after the study, they still estimated that it had saved them 20%.

METR’s observations offer one measure of review work. Developers accepted fewer than 44% of AI code generations, and 75% said they read every line of AI-generated code. In screen recordings from a subset of tasks, about 9% of working time went to reviewing and cleaning AI output. These figures describe this study’s participants and tasks; they are not a general estimate for all AI use at work.

At a glance
reportWhen: Published November 16, 2026; the cited…
The developmentKaren Feigh’s November 16 article in The Conversation draws attention to the workplace time and mental effort required to verify AI-generated output, citing research that finds results 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

Why Review Time Changes the Productivity Equation

For employers, the findings suggest that measuring how quickly AI produces an initial draft can miss work required to deliver a reliable result. Review, testing and correction can take time, and the person checking the output may need enough expertise to spot mistakes that are not obvious from polished wording or functioning-looking code.

The risks also vary with the task. In a study with Boston Consulting Group, 758 knowledge workers completed tasks that researchers classified as within or outside AI’s capabilities. On 18 tasks within that range, participants using AI completed 25.1% more quickly on average and completed 12.2% more tasks. On a separate complex task outside the range, AI users were 19% less likely to produce a correct answer than participants without AI. The results point to the importance of knowing when to check closely or avoid relying on a tool—not to a single productivity effect that applies everywhere.

Verification itself may not be enough if reviewers lack the information or time to assess an answer. In an experiment involving 2,784 people checking AI-style suggestions, participants caught typos most of the time but had 31% accuracy on errors requiring them to understand the task’s rules. That result makes the quality of oversight, not just the presence of a reviewer, relevant to workplace decisions about AI.

Products Worth Considering

Logitech MX Master 4 Ergonomic Wireless Mouse with Haptics – Graphite

Logitech MX Master 4 Ergonomic Wireless Mouse with Haptics – Graphite

As an affiliate, we earn on qualifying purchases.

Research Shows Different Task Outcomes

Feigh’s article connects the verification burden to a longstanding issue in automation: a system may shape decisions in practice while a human remains accountable for the consequences. She points to cognitive engineer David Woods’s 1985 description of a “responsibility-authority double bind.” In workplaces using AI, that tension can arise when a tool supplies an answer but the employee must stand behind it.

The METR trial provides unusually direct evidence about one setting, but its scope is limited. It involved experienced developers working in large, familiar open-source repositories, using Cursor Pro and Anthropic’s Claude 3.5 or 3.7 Sonnet during February to June 2025. METR said the result does not show that current AI systems are useless in other economically relevant settings, and it does not rule out future tools speeding developers’ work in similar tasks.

Other cited studies show why the setting matters. A 2023 study of mid-level professional writing tasks found ChatGPT users completed work 40% faster and produced work rated 18% higher in quality. The consulting experiment found benefits on tasks within AI’s capabilities but worse accuracy on one task outside them. Together, these studies do not establish one universal cost or benefit; they measure different tasks, participants and outcomes.

“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

Where the Review Burden Applies

The studies do not establish how much verification time a typical worker spends across all occupations. METR’s 9% figure came from screen recordings for 44 issues with valid labels, while its survey percentages reflect the participating developers’ reports. Those measures should not be treated as a universal workplace average.

It also remains unclear how the balance will change as AI tools and workplace practices evolve. METR said its findings do not rule out future gains in the software setting it studied. The cited experiments also do not settle how organizations should account for review time, or how reliably workers can identify errors in every domain.

Measure the Full Work Cycle

Feigh’s practical recommendation is to evaluate the whole workflow, rather than counting only the speed of an AI-generated first draft. That means accounting for review, testing, corrections and the time needed to decide whether an output is appropriate for its intended use.

For organizations adopting AI, the evidence supports tracking results by task and checking both speed and quality. Workers who remain accountable for AI-assisted output may need time and training to review it. The research summarized here does not prescribe a single system for doing so; further evidence across more jobs and tasks would be needed to estimate how large the verification burden is in different workplaces.

Key Questions

What does “verification burden” mean?

It is the time and mental effort spent checking whether AI output is accurate and appropriate. Karen Feigh uses the term in her November 16, 2026, article in The Conversation.

Did AI slow down every worker in the studies?

No. METR found experienced developers took 19% longer in its software trial, while other studies found faster work on professional writing and consulting tasks. The findings vary by task and study design.

What did the METR software trial measure?

It compared completion times for 246 real tasks assigned to 16 experienced open-source developers, with AI either allowed or disallowed. The trial found developers took 19% longer on average when AI was allowed.

Why can checking AI output be difficult?

Fluent output can still contain errors that require knowledge of the task to detect. In one experiment with 2,784 reviewers, accuracy was 31% for errors requiring rule-based understanding, although participants caught typos most of the time.

What should workplaces measure when using AI?

Feigh argues that organizations should measure the full work cycle, including review and correction, rather than only how quickly AI produces an initial answer. The cited research does not provide one universal measure for every workplace.

Source: ThorstenMeyerAI.com

You May Also Like

What The Two-Sentence Rule Means For AI Game Clones

Anthropic added a Claude Code instruction against recreating specific games, while allowing original games and genre-based projects.

How AI Transcription Software Can Help With Meetings And Field Work

A product roundup compares 11 transcription tools for meetings, dictation and field recordings, while flagging workflow, privacy and accuracy questions.

14 AI Coding Assistants For Programming Education And Software Development In 2027

A comparison of 14 books on AI-assisted programming, from beginner guides to tool-specific and team-focused workflow resources.

Does Work Shape More Than Our Careers? 8,234 Workers Consider AI

A 8,234-person survey links work to purpose, community and status, while testing whether guaranteed income would persuade people to quit.