The Check After The Chat: AI’s Overlooked Cost At Work
AIThis post was created with the assistance of artificial intelligence (AI).

🔍 Read the full analysis: The Check After The Chat: AI’s Overlooked Cost At Work on ThorstenMeyerAI.com

Before you orderOffer from Amazon

Get the latest gadgets 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 essay by Georgia Tech cognitive engineer Karen Feigh argues that workers’ time spent checking AI output can offset productivity gains. Studies she cites show mixed results: a randomized trial found experienced developers took longer with AI, while other experiments found faster work on tasks suited to the tools.

Georgia Tech cognitive engineer Karen Feigh argued in a November 16 essay in The Conversation that workers’ time spent checking AI-generated material can offset its productivity gains. Research she cites shows the effect varies by task: one randomized software-development trial found experienced developers took longer with AI, while other experiments found workers completed suitable writing and consulting tasks faster.

Feigh calls the work of determining whether AI output is accurate and appropriate the “verification burden.” In her account, a polished answer can still contain errors, omissions or outdated information, leaving workers responsible for checking it before use. She says an executive assistant told her she now builds time into her workflow to check AI-generated itineraries and briefings for inaccuracies and bias.

The clearest measured slowdown in the cited material comes from a METR randomized trial involving 16 experienced open-source developers and 246 real tasks. Developers took 19% longer, on average, when allowed to use AI tools. Before the trial, they expected a 24% time saving; afterward, they still estimated that AI had saved them 20% of their time.

METR reported that developers accepted fewer than 44% of AI code suggestions, and 75% said they read every line of generated code. In screen recordings for a subset of 44 issues, about 9% of AI-allowed working time went to reviewing and cleaning the output. The trial used tools including Cursor Pro and Anthropic’s Claude 3.5 and 3.7 Sonnet, and measured task completion time rather than just the time needed to produce an initial draft.

At a glance
reportWhen: Essay published November 16, 2026; rese…
The developmentKaren Feigh’s November 16 essay highlights a growing body of research showing that checking AI output can offset some workplace 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

Checking Can Erase Some Time Savings

The findings matter because workplace productivity is often discussed as the speed of generating a draft or completing an initial step. But the relevant measure is the whole workflow: checking, correcting, testing and taking responsibility for the result can add time after an AI tool produces an answer.

The evidence does not show that AI generally makes workers less productive. A 2023 study of professional writing tasks found that ChatGPT users completed work 40% faster and with 18% higher quality. In a separate experiment with Boston Consulting Group knowledge workers, participants using AI completed tasks within the tool’s capabilities 25.1% faster on average and completed 12.2% more tasks. On a deliberately selected task outside those capabilities, however, AI users were 19% less likely to reach a correct solution.

That contrast has practical implications for employers: speed alone may misstate the benefit if workers spend substantial time reviewing output or if confident errors are hard to spot. It also raises a question about accountability. Workers may remain answerable for decisions even when an AI system supplies part of the work, making time and training for review part of the cost of using the tool.

Products Worth Considering

Anker SOLIX C1000 Gen 2 Portable Power Station with 200W Solar Panel

Anker SOLIX C1000 Gen 2 Portable Power Station with 200W Solar Panel

As an affiliate, we earn on qualifying purchases.

Evidence Varies by Task

Feigh’s article is an expert commentary, not a new experiment. It draws on earlier research and argues that AI can create a mismatch: automated systems may influence work, while the human operator remains responsible for the outcome. The idea has a longer history in automation research; Feigh links it to cognitive engineer David Woods’ 1985 description of a “responsibility-authority double bind.”

The studies measure different work, tools and outcomes, so their results are not directly interchangeable. METR tested experienced developers working in large, mature open-source projects. The consulting experiment examined knowledge workers on tasks designed to fall both inside and outside AI’s capabilities. The writing study focused on mid-level professional writing assignments. Together, they suggest that the effect depends on the task and the system’s fit, rather than establishing one productivity figure for all workplaces.

Another experiment cited by Feigh asked 2,784 people to check AI-style suggestions. Participants caught typos more reliably than errors that required understanding the rules of the task; accuracy on those harder errors was 31%. That result points to a particular difficulty: output can look plausible while still being wrong in ways that require subject knowledge to detect.

“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 Review Burden Falls

The evidence does not establish how much checking time AI adds across workplaces as a whole. The METR trial involved 16 developers and a specific set of software tasks; its result should not be treated as a universal estimate. The screen-recording figure came from a subset of 44 issues, rather than every task in the study.

Results also depend on the tools available at the time and on workers’ experience. METR used particular AI products during February to June 2025, and it said its findings did not rule out future systems improving performance in the same setting. The cited studies do not provide a single estimate of the net cost or benefit across industries, including the money spent on training, error correction and oversight.

It also remains unclear how reliably workers identify subtle errors in day-to-day use. The 31% accuracy result concerned errors requiring an understanding of task rules in one experiment; it cannot establish the rate at which employees miss errors in other settings. Feigh’s essay raises the issue, but does not quantify a company-wide verification burden.

Measure the Full Workday

Feigh’s argument points employers toward tracking total task completion time, including review, corrections and testing, rather than counting only how quickly AI produces a first draft. Evaluations should also distinguish tasks where a tool performs reliably from those that require more human judgment.

Workers who remain accountable for AI-assisted output may need protected time and training to check it. Future studies could clarify how review demands change as tools and workplace practices evolve, and whether workers become better at spotting errors. No single follow-up milestone or new study was announced in Feigh’s essay; the immediate question for organizations is whether measured gains persist after the checking work is counted.

Key Questions

Did the METR study find that AI slows down all workers?

No. It found that 16 experienced open-source developers took 19% longer on average across 246 tasks when allowed to use AI. The result applies to that trial’s participants, tasks and tools, not all jobs.

Why can AI add time to a task?

Workers may need to read, test, correct or reject generated output. In METR’s survey, 75% of developers said they read every line of AI-generated code, and the researchers recorded review and cleanup time on a subset of tasks.

Do other studies show workplace productivity gains?

Yes. The cited writing study found ChatGPT users completed mid-level professional writing tasks 40% faster, with quality 18% higher. A consulting experiment also found faster completion on tasks within AI’s capabilities, while results were worse on a task outside them.

What should employers measure?

They should measure the full workflow, including review and correction, and distinguish performance on tasks suited to AI from performance on tasks outside its capabilities.

Source: ThorstenMeyerAI.com

You May Also Like

The Benefits Of Software-Driven Hospital Spiritual Care Management

Hospital chaplaincy departments are testing new operations software to improve scheduling, visit logging, and compliance, marking a shift towards digital spiritual care management.

2026’S Leading AI Automation Tools For Small Business Efficiency

Discover the 15 leading AI automation tools for small businesses in 2026, designed to boost efficiency, reduce costs, and scale operations effectively.

The Orchestration Layer Arrives: What Anthropic’s Finance Agents Mean for Bloomberg, FactSet, and Wall Street

Anthropic releases AI agent templates and connectors, positioning Claude as an orchestration layer over financial data providers, disrupting Bloomberg’s dominance.

AI Cost Milestone: DeepSeek-V4-Flash-High’s Ninth Point At $0.25 Per Million

DeepSeek-V4-Flash-High achieves a cost milestone at $0.25 per million tokens, driven by post-training improvements, marking a shift in AI model cost-efficiency.