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

🔍 Read the full analysis: The Real Effort Of AI At Work May Come After It Delivers 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, 2026, essay by Georgia Tech cognitive engineer Karen Feigh argues that checking AI-generated work can offset its productivity gains. Research she cites shows mixed results: experienced developers in one trial took 19% longer with AI, while other studies found faster work on tasks suited to AI. The evidence does not establish a universal productivity penalty.

Georgia Tech cognitive engineer Karen Feigh argued in a November 16 essay that the effort required to check AI-generated work can offset the time saved producing it, a hidden cost she calls the “verification burden.” Research she cites shows that the effect depends on the task: one trial found experienced software developers took longer with AI, while other experiments found gains when work fell within the tools’ capabilities.

The clearest example comes from a randomized trial by METR, a research group that studied 16 experienced open-source developers working on 246 real tasks in large software projects. Tasks were assigned either to allow or disallow AI tools. In this setting, developers took 19% longer on average when AI was allowed. The trial used tools including Cursor Pro and Anthropic’s Claude 3.5 and 3.7 Sonnet during February to June 2025.

The measured time included work after the AI produced code. METR reported that developers accepted fewer than 44% of AI-generated 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, developers spent about 9% of working time reviewing and cleaning AI output. These findings describe this group and task environment, not all software work.

Results elsewhere point in different directions. In a study with 758 knowledge workers and Boston Consulting Group, participants using AI completed tasks within the tools’ capabilities 25.1% faster on average and completed 12.2% more tasks. On a complex task outside those capabilities, however, AI users were 19% less likely to produce a correct answer. A separate study of professional writing tasks found ChatGPT reduced completion time by 40% and increased quality by 18%.

At a glance
reportWhen: Essay published November 16, 2026; rese…
The developmentKaren Feigh’s essay draws attention to the time and mental effort workers spend verifying AI output, citing research that finds sharply different productivity results across tasks.

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 studies suggest that a fast first draft is not the same as a faster completed job. If workers must inspect, test, correct or rewrite output, those steps belong in any credible measure of workplace productivity. The METR trial is a concrete warning against treating tool use or generation speed as proof that a workflow has improved.

Verification also carries a quality risk. Feigh cites an experiment involving 2,784 people reviewing AI-style suggestions: participants usually caught typos, but correctly identified only 31% of errors requiring them to understand the task’s rules. A polished answer can therefore require more than proofreading. Workers may need enough subject knowledge and time to detect substantive mistakes, while responsibility for the final work remains with them.

For employers, the practical implication is to evaluate the full process, including review and correction, rather than measuring only how quickly AI produces a draft. The evidence supports a task-by-task approach: benefits may be substantial where a tool performs reliably, but a poor fit can bring slower work or incorrect results.

Products Worth Considering

As an affiliate, we earn on qualifying purchases.

Different Tasks, Different Results

Feigh’s article in The Conversation is an expert perspective, not a new experiment. It draws on published and preprint research to explain why AI’s effects on work can be difficult to predict. The studies use different workers, tools, tasks and outcome measures, so their percentages should not be combined into one estimate of AI’s overall workplace effect.

The contrast between the developer trial and the consulting experiment reflects that variation. In METR’s study, participants were experienced developers working inside mature code repositories, where understanding existing systems and checking changes were part of the task. In the Boston Consulting Group experiment, AI helped with a set of realistic knowledge-work tasks within its capabilities, but performance fell on a task deliberately placed outside that range. The findings show uneven gains across task boundaries, not a single result that applies to every job.

Feigh also connects today’s oversight demands to a long-standing automation problem described by cognitive engineer David Woods in 1985: automated systems can exercise practical authority while humans retain responsibility. Her comparison is to supervising a fast, polished intern whose output can be useful but whose confidence does not reliably signal accuracy. That analogy is Feigh’s framing; the measured results come from the individual studies.

“AI’s productivity gains are often countered by time and mental energy spent determining whether the output is accurate and appropriate.”

— Karen Feigh, Georgia Tech cognitive engineer, in The Conversation

How Broad Are the Findings?

The cited research does not establish one productivity effect for AI across workplaces. The developer trial involved 16 experienced contributors, specific software projects and tools available in 2025. Its results may not apply to newer systems, less experienced users, other development work or different ways of integrating AI into a team.

Other limits also matter. The studies measure different tasks and outcomes, and the available source material does not provide enough detail to compare their results on a common basis. It remains unclear how much review time varies by occupation, how workers’ checking skills change with experience, and whether organizations can redesign workflows to reduce the burden without allowing errors to pass through. METR itself cautioned that its result does not show AI is useless in economically relevant settings or rule out future productivity gains in the same environment.

Measure the Whole Workflow

The next step for employers and researchers is to evaluate AI use across the complete task cycle: generation, review, testing, correction and final quality. That means comparing equivalent work with and without AI, tracking both time and accuracy, and identifying which kinds of tasks reliably benefit from assistance.

Feigh’s recommendation is that organizations give workers responsible for AI-assisted output time and training to check it. Further studies will be needed to establish whether newer models or different workplace practices change the results seen in the cited trials. For now, the available evidence supports a measured conclusion: AI can speed up suitable work, but its output still needs oversight, and that oversight can carry a real cost.

Key Questions

What does “verification burden” mean?

It is the time and mental effort spent checking AI output for accuracy and suitability. Feigh argues that this work can reduce or offset the time saved when AI generates a draft or suggestion.

Did AI make developers faster in the METR trial?

No. In the trial of 16 experienced open-source developers completing 246 tasks, participants took 19% longer on average when AI was allowed. The finding applies to that study’s participants, tools and tasks, not every software project.

Do studies show AI always slows down work?

No. Other experiments found faster work on tasks within AI’s capabilities, including a 25.1% reduction in completion time in one consulting experiment and a 40% reduction in time on professional writing tasks in another study. Results vary by task and study design.

Why can AI-assisted work still require careful review?

AI output can contain errors that are difficult to spot through surface proofreading. In one experiment cited by Feigh, reviewers identified only 31% of errors requiring an understanding of task rules, even though they usually caught typos.

What should employers measure when introducing AI?

They should track the full workflow, including time spent checking and correcting output, as well as the accuracy and quality of the completed work. Draft-generation speed alone does not show whether a job was completed more efficiently.

Source: ThorstenMeyerAI.com

You May Also Like

What’s Behind Grok’s Faster Limits For Users?

A headline says Elon Musk explained faster Grok usage-limit hits, but the explanation, affected users and any policy change remain unverified.

What Investors Can Learn From Anthropic’s Possible “Junk” Rating

Blackstone reportedly expects Anthropic to receive a double-B rating after its IPO. No major rating agency has announced a rating.

The Future Of Encryption: AI Mathematics, Quantum Computers And Finance

AI-generated mathematics and quantum computing raise separate questions for encryption. No cryptographic systems have been reported broken.

The Pulse: Anthropic Kills The Coding Interview

An Anthropic hiring manager says their team stopped using coding questions after finding data-structure interviews gave no useful hiring signal.