Why The Cost Of AI At Work Includes Checking Its Answers
AIThis post was created with the assistance of artificial intelligence (AI).

🔍 Read the full analysis: Why The Cost Of AI At Work Includes Checking Its Answers on ThorstenMeyerAI.com

Before you orderOffer from Amazon

Get smart everyday buys 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 article by Georgia Tech cognitive engineer Karen Feigh argues that AI productivity claims should account for the time and mental effort workers spend checking outputs. Research she discusses found mixed results: experienced developers in one randomized trial took 19% longer with AI, while other experiments found faster work on tasks suited to AI. The findings do not establish a universal productivity penalty.

Georgia Tech cognitive engineer Karen Feigh argued in a November 16, 2026, article in The Conversation that workplace AI’s apparent speed can come with a less visible cost: workers must spend time and mental effort checking whether its output is accurate and appropriate. The research she cites shows that results vary by task, including a randomized software-development trial in which experienced developers took 19% longer with AI tools, even as other studies found gains in work within AI’s capabilities.

Feigh calls the effort of reviewing AI output the “verification burden.” Her article is expert commentary, not a new experiment. It draws on existing research to argue that productivity should be measured across the entire workflow, including checking, correcting and testing, rather than by how quickly a tool produces a first draft.

The clearest example comes from a randomized trial by the research group METR. Sixteen experienced open-source developers completed 246 real tasks in large software projects, with each task assigned to allow or disallow AI. In the study, developers took 19% longer when AI was allowed. They had expected AI to make them faster: before the trial, they forecast a 24% time saving, and after it they still estimated a 20% saving.

METR also recorded the review work. Developers accepted fewer than 44% of AI code suggestions; 75% said they read every line of generated code, and 56% reported often needing major changes to clean it up. For a subset of 44 tasks with usable screen recordings, about 9% of working time went to reviewing and cleaning AI output. That figure describes the recorded subset, not every task in the trial.

At a glance
reportWhen: Commentary published November 16, 2026;…
The developmentA new expert commentary highlights research showing that the time workers spend verifying AI output can offset productivity gains, depending on the 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

The Work Hidden After AI Drafts

The findings matter because workplace productivity estimates can be misleading if they count only the time saved on drafting. If workers remain responsible for an answer but cannot assume it is correct, they need time to inspect its claims, test its code or compare it with the task’s requirements. That review is part of the work, even when an AI tool performs the initial generation.

Evidence cited by Feigh points to both gains and risks rather than a single outcome. In a study of professional writing tasks, ChatGPT reduced completion time by 40% and increased measured quality by 18%. In a Boston Consulting Group experiment, AI users completed tasks within the tools’ capabilities 25.1% faster on average. But on a task designed to fall outside those capabilities, they were 19% less likely to give a correct answer than participants without AI.

That contrast suggests the value of AI depends partly on whether workers can recognize when its output needs more scrutiny. The cited studies do not establish that every workplace will see the same results, or that review always cancels out time saved. They do show why organizations need to measure accuracy and correction alongside speed.

De'Longhi Stilosa Compact Manual Espresso Machine with Milk Frother

De'Longhi Stilosa Compact Manual Espresso Machine with Milk Frother

Ninja CREAMi Scoop & Swirl 13-in-1 Ice Cream and Soft Serve Maker, NC701

Ninja CREAMi Scoop & Swirl 13-in-1 Ice Cream and Soft Serve Maker, NC701

As an affiliate, we earn on qualifying purchases.

What the Cited Studies Measured

The METR trial focused on experienced developers working in familiar, mature projects, not on every kind of coding or workplace task. Participants mainly used Cursor Pro and Anthropic’s Claude 3.5 or 3.7 Sonnet during the study period from February to June 2025. The researchers counted completion time, including work after reviewing the code, rather than treating an AI-generated draft as a finished result.

The consulting experiment, led by Fabrizio Dell’Acqua, Ethan Mollick, Karim Lakhani and colleagues, involved 758 knowledge workers in a preregistered study conducted with Boston Consulting Group. On 18 tasks within the defined AI capability frontier, participants using AI completed 12.2% more tasks and finished 25.1% faster on average, with improved solution quality. On one complex task outside that frontier, AI users were less likely to produce a correct solution.

Feigh also cites an experiment involving 2,784 people reviewing AI-style suggestions. Participants usually caught typos, but accuracy was 31% when errors required understanding the task’s rules. This points to a distinction between spotting surface mistakes and judging whether an answer is substantively right.

“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 Productivity Cost Varies

The studies do not establish a single verification cost for AI across jobs. The METR trial was small and limited to experienced developers working on specific repositories with particular tools. Its review-time estimate came from only 44 tasks with valid labeled recordings. Results could differ by task, worker experience, software and model.

Nor does the 19% slowdown show that AI is generally unhelpful. METR said its findings do not rule out useful applications in other economically relevant settings or future systems that may improve performance on the tested tasks. The consulting and writing studies cited by Feigh measured different work and found gains in defined settings.

It is also unclear how consistently workers can detect errors that require domain knowledge. The reviewer experiment’s 31% accuracy figure concerns errors that depended on understanding task rules; it should not be treated as a general error-detection rate for all AI outputs. The evidence supports measuring verification, but does not yet specify one reliable method or standard for doing so across workplaces.

Measure the Full Work Cycle

Feigh’s practical recommendation is for employers to track the complete process: generation, checking, correction and final quality. That means comparing AI-assisted work with an appropriate baseline and recording whether speed gains come with additional review time or errors. A fast first draft alone is not evidence of a faster completed task.

Organizations also need to decide who is accountable for the final output and give those workers enough time and training to assess it. The cited research does not announce a new workplace policy or regulatory requirement. The next evidence will come from further studies and from employers’ own task-level measurements, which can show where AI helps, where it needs close supervision and whether the finished work improves.

Key Questions

What does “verification burden” mean?

It is the time and mental effort spent checking AI output for accuracy and suitability, and correcting it when needed. Feigh uses the term in her commentary to describe work that may be missed when productivity is measured only by drafting speed.

Did AI make the developers in the METR trial faster?

No. In that randomized trial, 16 experienced developers took 19% longer on 246 tasks when AI tools were allowed. The result applies to that study’s participants, tasks and tools; it is not a finding about every developer or workplace.

Do the studies show that AI always slows workers down?

No. Other cited experiments found faster work on tasks that fit AI’s capabilities, including professional writing and consulting tasks. The reported results varied by task, and the studies do not establish one productivity effect for all uses of AI.

Why can checking AI output be difficult?

Some errors are easy to spot, such as typos, while others require knowledge of the task and its rules. In an experiment cited by Feigh, reviewers’ accuracy was 31% on errors requiring that kind of understanding.

How should employers judge whether workplace AI saves time?

Measure the whole workflow, including generation, review, corrections, testing and final quality. The cited research does not prescribe a universal measurement system, but it shows why draft speed alone can give an incomplete picture.

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

Évian and the Fallout: What Europe Actually Wants From Amodei, Hassabis, and Altman

Europe pushes for reliable access, sovereignty, and safety standards from Amodei, Hassabis, and Alt at the G7 AI summit in Évian.

How To Review Cursor After A Tampered Claude Code Update

Cursor says a tampered Claude Code extension was available for about 11 hours. Affected users should check for an SSH key and follow cleanup guidance.

Jeff Bezos’ family office backed five AI startups in June

Jeff Bezos’ family office reportedly backed five AI startups in June, highlighting increased private sector interest in AI development.

How Expanded Claude Access Fits Into Harvard Students’ Academic Lives

A Harvard Crimson headline reports students are using expanded Claude access for classes and personal projects, but details and effects remain unknown.