When AI Helps With Work, Who Checks The Output?
KIDieser Beitrag wurde mit Unterstützung künstlicher Intelligenz (KI) erstellt.

🔍 Read the full analysis: When AI Helps With Work, Who Checks The Output? on ThorstenMeyerAI.com

Buying for a business?Offer from Amazon

Get business pricing on office and shipping supplies

  • 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

A November 16 article by Georgia Tech cognitive engineer Karen Feigh highlights a growing workplace issue: people remain responsible for AI-assisted work, but must spend time checking whether it is accurate and appropriate. Studies show mixed results: AI helped on some writing and consulting tasks, while a randomized trial found experienced developers took longer with AI, and another experiment found people often missed errors requiring task-specific understanding.

Georgia Tech cognitive engineer Karen Feigh argued November 16 that workers using AI often take on a “verification burden”: checking whether fast, polished output is accurate and appropriate while remaining accountable for it. Research she cites finds that AI can speed up some tasks, but in other settings review work, errors, or weaker performance can offset the gains.

Feigh made the case in an article published by The Conversation; it is an expert commentary, not a new study. She describes workers learning to supervise AI like “an extraordinarily fast and polished intern” whose confident answers can contain both good and bad work, with few clear signals about where scrutiny is needed.

One of the most direct examples comes from a randomized trial by METR, a research group that studied 16 experienced open-source developers working on 246 tasks in large software projects. In the trial, developers took 19% longer on average when AI tools were allowed. The study measured task completion time, including work after code review, rather than only how quickly the initial code appeared.

The developers accepted fewer than 44% of AI code suggestions, according to METR. 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. In a subset of tasks with usable screen recordings, about 9% of AI-allowed working time went to reviewing and cleaning output. The result applied to this group and task setting; METR said it did not show that AI tools are unhelpful in other economically relevant settings.

At a glance
reportWhen: Commentary published November 16, 2026;…
The developmentKaren Feigh’s November 16 commentary draws on workplace research to argue that AI productivity gains can be offset by the time and mental effort workers spend verifying its output.

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 Work Shapes AI Productivity

The studies point to a practical measurement problem for employers: faster first drafts are not the same as faster completed work. If employees must check facts, test code, correct omissions, or judge whether an answer fits the task, those steps belong in any assessment of productivity. The METR developers expected AI to reduce task time by 24% before the trial and still estimated a 20% saving afterwards, despite taking longer in the measured tasks. That gap suggests workers’ sense of speed may not reliably capture the full workflow.

Performance also depends on the task. In research conducted with Boston Consulting Group, 758 knowledge workers using AI completed tasks within the tools’ demonstrated capabilities 25.1% faster on average and completed 12.2% more tasks, with higher solution quality. On one complex task selected to fall outside that capability range, however, AI users were 19% less likely to produce a correct solution. These findings make review capacity a matter of work quality as well as efficiency: an answer can look finished while still requiring informed checking.

That responsibility can fall on workers without giving them clear signals about which output is reliable. Feigh connects the issue to a longstanding automation problem: systems may exercise practical authority over a process while a human remains responsible for its outcome. For organizations, the evidence supports measuring end-to-end results and allowing time and training for people expected to verify AI-assisted work. It does not establish that every AI tool or workplace will have the same costs.

Products Worth Considering

As an affiliate, we earn on qualifying purchases.

Evidence Varies by Task and Study

The research Feigh cites does not support a single productivity figure for AI. A 2023 study by Shakked Noy and Whitney Zhang found that ChatGPT reduced completion time by 40% on mid-level professional writing tasks and raised the quality of the work in that experiment by 18%. That result concerned a defined set of writing tasks, not every kind of workplace activity.

The BCG experiment similarly found gains on tasks within the AI’s demonstrated capabilities, alongside worse accuracy on a task outside them. The METR trial involved experienced developers working in mature software projects, using tools including Cursor Pro and Anthropic’s Claude 3.5 and 3.7 Sonnet during the study period in 2025. Its results cannot by themselves predict performance with newer systems, different users, or different coding tasks.

Checking output can also be difficult even when reviewers are asked to do it. In an experiment with 2,784 people reviewing AI-style suggestions, participants usually caught typos, but achieved only 31% accuracy on errors that required understanding the rules of the task. That finding concerns the experiment’s particular review exercise; it does not measure error rates across all AI use. Taken together, the studies show why tool performance, task difficulty, and the reviewer’s expertise all matter.

“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 Far the Findings Apply

The evidence does not establish a universal time cost for AI review. The studies examined different tasks, participants, tools, and measures, so their results cannot be combined into one overall productivity estimate. METR’s trial was small and focused on experienced developers working in established repositories; it also used tools available in 2025. Whether newer models or different workflows would change the result remains open.

The supplied research also does not quantify how often workers catch errors in everyday AI-assisted work, how review time varies between occupations, or what training reduces mistakes. The BCG finding on an out-of-capability task is one experimental result, not evidence that AI will reduce accuracy by the same amount across other tasks. Feigh’s article synthesizes prior research and workplace observations; it does not report a new survey or experiment.

Measure the Whole Workflow

The practical next step raised by Feigh’s argument is to evaluate completed work, not just draft speed. Employers adopting AI can track time spent prompting, reviewing, testing, and correcting output alongside accuracy and quality. They can also identify which tasks fall within a tool’s reliable capabilities and give workers who remain accountable the time and training to check results.

No single follow-up study or policy change is specified in Feigh’s commentary. Further research across occupations, tools, and task types would help show when verification costs outweigh speed gains and which review practices work. Until that evidence is available, the cited studies support a qualified conclusion: AI can help on some tasks, but its value depends partly on whether people can identify and correct its mistakes.

Key Questions

What does “verification burden” mean?

It is the time and mental effort spent checking AI output for accuracy and suitability, even when the tool produces a quick, polished answer.

Did AI make developers faster in the METR trial?

No. In METR’s randomized trial of 16 experienced open-source developers and 246 tasks, tasks took 19% longer on average when AI tools were allowed. The result is specific to that study’s participants, tools, and work.

Do studies show that AI always slows workers down?

No. Results differed by task. A professional-writing study found ChatGPT reduced completion time by 40%, and a BCG experiment found faster completion on tasks within AI’s capabilities. The METR result concerned a particular software-development setting.

Why can AI-assisted work still contain errors?

AI output may be incorrect or unsuitable while sounding confident. In one experiment with 2,784 reviewers, accuracy was 31% on errors requiring understanding of task rules, even though typos were usually caught.

What should employers measure when introducing AI?

They should assess the full workflow: time to produce, review, test, and correct work, along with its accuracy and quality. The cited research does not provide one productivity estimate that applies to every workplace.

Source: ThorstenMeyerAI.com

This content is for general information only and is not financial, tax or legal advice. Consult a qualified professional for decisions about your money.
You May Also Like

Your Coding Agent Is an Attack Surface: The Claude Code Security Reckoning

Recent vulnerabilities in Claude Code reveal critical attack surfaces via local configs and MCP integrations, risking token theft and code execution.

Building an AI Trading Bot — Week One: Why a 90 % Win Rate Can Still Lose Money

Initial testing of an AI trading bot reveals that a high win rate does not guarantee profitability. Findings highlight the importance of market-implied probabilities.

Avengers Labs: How Ukraine Turned Its Front Line Into the World’s Scarcest AI Dataset

Ukraine’s Avengers Labs leverages combat drone data to train AI models, creating a unique defense resource that could reshape modern warfare.

AI Signal Monitoring: A Clear Shift Toward Data Center Investment Strategies

New monitoring tools reveal a trend of AI operations resembling data center REITs, impacting small team rollout strategies.