At Work With AI: The Often Overlooked Cost Of Checking Its Results
AIThis post was created with the assistance of artificial intelligence (AI).

🔍 Read the full analysis: At Work With AI: The Often Overlooked Cost Of Checking Its Results 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 essay by Georgia Tech cognitive engineer Karen Feigh highlights the time and mental effort workers spend checking AI output. Studies show results vary by task: one trial found experienced developers took 19% longer with AI, while other experiments recorded faster work on tasks suited to the tools. The evidence does not establish a universal productivity penalty.

Georgia Tech cognitive engineer Karen Feigh argued in a November 16 essay that the time and mental effort workers spend checking AI output can offset some of its apparent productivity gains. Research she cited shows sharply different results by task: experienced software developers in one randomized trial took longer with AI, while other experiments found faster completion on work better suited to the tools.

Feigh, writing in The Conversation, calls the work of checking whether AI output is accurate and appropriate the “verification burden.” She says workers may need to review polished-looking material for errors, missing steps, outdated sources or bias. Her essay is an expert commentary drawing on earlier research, not a new experiment.

The clearest example of a potential time cost comes from a METR randomized trial involving 16 experienced open-source developers and 246 tasks in large, established software projects. The study reported that developers allowed to use AI took 19% longer to complete tasks on average than developers who were not allowed to use it. Participants had expected a 24% time reduction before the trial and still estimated a 20% reduction afterward.

METR also reported that developers accepted fewer than 44% of AI code suggestions; 75% said they read every line of generated code, and 56% said they often had to make major changes. In screen recordings for a subset of 44 issues, reviewing and cleaning AI output accounted for about 9% of AI-allowed working time. The study included work after code review, rather than measuring only how quickly an initial draft appeared.

At a glance
reportWhen: Essay published November 16, 2026; rese…
The developmentKaren Feigh’s November 16 essay draws attention to research showing that checking AI output can absorb time and complicate claims that AI automatically makes work faster.

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

Checking Time Changes the Productivity Picture

The findings matter to employers and workers evaluating whether AI saves time, because draft speed is not the same as task completion time. A tool can produce an answer quickly while still leaving a person responsible for checking its accuracy, revising it and dealing with errors. If those steps are omitted from an evaluation, estimates of productivity may not reflect the full workflow.

The research also cautions against treating AI assistance as a single, consistent benefit. In a study with Boston Consulting Group, 758 knowledge workers completed realistic tasks. On 18 tasks judged to be within AI’s capabilities, AI users completed 12.2% more tasks and worked 25.1% faster on average, with higher-quality solutions. On a complex task outside those capabilities, however, users of AI were 19% less likely to produce a correct solution than those without it. The different results make task fit and the ability to check output central to assessing impact.

For workers, the burden can include attention as well as minutes: a reviewer must decide what to verify and whether a confident answer is trustworthy. For organizations, assigning responsibility without enough time, expertise or authority to inspect the result could leave important errors undetected. These are implications of the evidence, not findings that every AI-assisted workflow creates the same risk.

Ninja Luxe Café Premier 3-in-1 Espresso Machine, Drip Coffee & Cold Brew

Ninja Luxe Café Premier 3-in-1 Espresso Machine, Drip Coffee & Cold Brew

As an affiliate, we earn on qualifying purchases.

Studies Find Different Results by Task

Feigh describes AI at work as resembling “supervising an extraordinarily fast and polished intern.” In her account, the output can be useful but may combine good and bad work in a similarly confident voice, while the human remains accountable. She also points to a longstanding concern in automation research: machines can exercise practical influence over decisions even when responsibility stays with people.

The METR trial tested a specific group and setting: experienced developers working in mature repositories, mainly using Cursor Pro and Anthropic’s Claude 3.5 or 3.7 Sonnet during February to June 2025. Its authors said the results did not show that current AI is unhelpful in all economically relevant work and did not rule out future tools speeding up developers in the same setting.

Other cited research points to potential gains. An experiment on mid-level professional writing tasks found that ChatGPT reduced completion time by 40% and increased quality by 18%. Those figures describe that study’s tasks and participants, not a general estimate for all workplace uses. Taken together, the studies show why results from one occupation or task cannot automatically predict performance elsewhere.

“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

How Often Checks Catch Errors

The cited studies do not establish a single average verification cost across occupations, tools or tasks. The METR result applies to a small sample of experienced developers working on particular repositories with tools available in 2025. The BCG experiment and writing study used different participants and work, so their results are not directly interchangeable.

Feigh also cites an experiment involving 2,784 people checking AI-style suggestions. Participants caught typos most of the time, but achieved only 31% accuracy on errors requiring an understanding of the task’s rules. That result raises concern about harder-to-spot mistakes, but the supplied material does not establish how often such errors arise in ordinary workplaces or how well review performs in every profession. It remains unclear how newer tools, better training and different review practices change the overall time cost.

Measure the Full Work Cycle

The practical next step suggested by the evidence is to measure the whole workflow: time spent prompting, reviewing, testing, correcting and completing a task, as well as the quality of the final result. Employers should not treat the speed of an initial draft as proof that the entire job is faster. Comparisons also need to specify the task, the tool and the workers involved.

Feigh’s essay argues that people who remain accountable for AI-assisted work need time and training to check it. Further research across roles and tasks will be needed to show how often verification cancels out initial speed gains, when it improves results, and how those patterns change as tools develop. Until then, productivity claims should be tied to the specific work measured rather than generalized to all workplace AI.

Key Questions

What does “verification burden” mean?

It is Karen Feigh’s term for the time and mental effort spent checking whether AI output is accurate and appropriate before relying on it.

Did AI make developers slower in the METR trial?

Yes. In a randomized trial involving 16 experienced developers and 246 tasks, METR reported that the AI-allowed group took 19% longer on average. The result describes that study’s setting, not all software work.

Do studies show that AI always reduces productivity?

No. Results varied by task. The cited BCG experiment found faster completion on tasks within AI’s capabilities, while the METR developer trial found longer completion times. The evidence does not support a single outcome for every use.

Why can checking AI output be difficult?

Some errors require understanding the task’s rules rather than spotting a typo. In an experiment cited by Feigh involving 2,784 reviewers, accuracy was 31% for errors requiring that kind of understanding.

How should organizations assess AI productivity?

They should track the complete task cycle, including review, testing and correction, alongside speed and quality. The cited evidence also supports giving accountable workers enough time and training to inspect output.

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

15 AI Recorders To Turn Meeting Conversations Into Clearer Notes

A comparison of 15 AI meeting recorders, including their stated strengths and the app, privacy and transcription details buyers should verify.

Open-Weight Kimi K3.1 Pro Nears The Frontier Before Its Weights Release

Artificial Analysis ranks Kimi K3.1 Pro fourth overall. Moonshot says its weights will follow by Nov. 4, but has not published a licence.

How ByteDance’s Methodical AI Strategy Is Changing Industry Dynamics

ByteDance’s ‘slow first, fast afterwards’ AI strategy aims to build foundational capacity before rapid deployment, potentially impacting industry competition.

A Deep Dive Into AI In A Canada-EU Partnership

An analysis of the current AI models from Canada and Europe, highlighting strengths, licensing differences, and implications for collaboration.