What Happens After AI Produces An Answer? The Cost Of Checking It
AIThis post was created with the assistance of artificial intelligence (AI).

🔍 Read the full analysis: What Happens After AI Produces An Answer? The Cost Of Checking It 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

Research cited in a November 16, 2026, article by Georgia Tech cognitive engineer Karen Feigh examines the time and effort people spend checking AI output. Results vary by task: a METR software trial found experienced developers took 19% longer with AI, while other experiments found faster work on tasks suited to the tools. The evidence supports measuring the full workflow, not just how quickly AI produces a draft.

AI can produce work quickly, but checking whether that work is accurate and appropriate can take time and mental effort, Georgia Tech cognitive engineer Karen Feigh argued in an article published November 16 in The Conversation. The studies she discusses report mixed results: a randomized trial found experienced software developers took longer with AI, while experiments in other kinds of work found productivity gains on tasks within AI’s capabilities.

Feigh calls the added effort of evaluating AI output the “verification burden.” She argues that productivity calculations can miss it when they count the time needed to generate a response but not the work required to check, correct or discard it. Her article is an expert analysis drawing on existing research, not a new experiment.

The most direct example comes from a METR randomized trial involving 16 experienced open-source developers and 246 tasks in large software projects. Tasks were randomly assigned to allow or disallow AI. In the study, developers using AI took 19% longer to complete their tasks on average. The result includes work after the initial code generation, rather than measuring only how fast a first draft 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 had to make major changes to clean it up. For 44 tasks with usable screen recordings, researchers estimated that about 9% of AI-allowed working time went to reviewing and cleaning generated output. These figures describe that study and should not be treated as a universal measure of AI review time.

At a glance
reportWhen: Feigh’s article appeared November 16, 2…
The developmentAn expert article and studies it reviews highlight the verification work that can offset AI’s speed gains, with results differing substantially by task and setting.

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

Why AI Speed Can Mislead

The findings matter to employers and workers evaluating whether AI has made a process more productive. Generation speed is not the same as task completion speed: a quick answer may create follow-up work, particularly when the person using it remains responsible for its accuracy. If that review time is left out, estimates of time saved can overstate the practical benefit.

The stakes also vary with the kind of error. A typo may be easy to spot, while an answer that violates a project rule or misunderstands a task can look polished and still be wrong. In an experiment with 2,784 people checking AI-style suggestions, participants caught typos most of the time but achieved only 31% accuracy on errors requiring understanding of the task’s rules. That result points to limits in review performance in that experiment; it does not establish a general error rate for AI users.

For organizations, the practical implication is to track the whole workflow: time spent prompting, checking, testing, revising and correcting, alongside output quality. Review also requires staff who understand the underlying work and have enough time to question confident-looking answers.

Products Worth Considering

Apple Watch SE (2nd Generation) GPS 40mm Smartwatch with Starlight Aluminum Case and Starlight Sport Band (S/M)

Apple Watch SE (2nd Generation) GPS 40mm Smartwatch with Starlight Aluminum Case and Starlight Sport Band (S/M)

Bose QuietComfort Headphones (2nd Gen), Wireless Noise-Cancelling Over-Ear Headphones, Eucalyptus Green

Bose QuietComfort Headphones (2nd Gen), Wireless Noise-Cancelling Over-Ear Headphones, Eucalyptus Green

As an affiliate, we earn on qualifying purchases.

Results Depend on the Task

The METR result is not the only evidence Feigh cites, and it does not show that AI always slows work. In a study of mid-level professional writing tasks, ChatGPT reduced completion time by 40% and increased quality by 18%. Those findings apply to the tasks and participants in that experiment, rather than all writing or workplace use.

A separate preregistered experiment involving 758 knowledge workers and Boston Consulting Group tested tasks at different levels of fit with AI. On 18 tasks judged to be within the tools’ capabilities, participants using AI completed 12.2% more tasks and worked 25.1% faster on average, with higher-quality solutions. On a complex task outside that capability range, AI users were 19% less likely to produce a correct solution than participants without AI.

The studies differ in occupations, tools, tasks and measures, so their numbers should not be combined into a single estimate of AI’s effect. METR’s software trial used Cursor Pro and Anthropic’s Claude 3.5 or 3.7 Sonnet during February to June 2025. Its authors said the findings do not establish that AI is unhelpful in other economically relevant settings or rule out future gains in the tested work.

““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 These Findings?

The studies do not establish a single, general amount of time that workers spend verifying AI output. The METR trial was small and focused on experienced open-source developers working in mature projects; its result may not apply to less experienced developers, other software tasks or newer tools. The writing and consulting experiments used different tasks and outcomes, making direct comparisons difficult.

It is also unclear how review practices and results change as AI systems, workplace processes and worker training evolve. The sources describe particular tools and experiments, not a lasting productivity rate for AI. Feigh’s article connects research to a broader concern about human responsibility, but it does not itself measure how often workers miss errors across workplaces.

Measure the Full Work Cycle

The evidence points to a practical next step for organizations adopting AI: compare assisted and unassisted work using the same task and quality standards, and record time spent checking and repairing output as well as time spent producing it. Results should be reported by task type and capability fit rather than treated as one productivity figure.

People who remain accountable for AI-assisted work will also need time and training to review it. Further research can test whether those measures improve error detection and whether newer tools change the balance between generation and verification. No single follow-up result is specified in the source material; the broader question remains how gains hold up once the entire workflow is counted.

Key Questions

What does “verification burden” mean?

It is the time and mental effort spent checking AI output for accuracy and suitability. Feigh uses the term in her analysis of how review work can offset faster generation.

Did the METR study show that AI always slows developers down?

No. The trial found that 16 experienced open-source developers took 19% longer on average across the 246 tested tasks when AI was allowed. METR cautioned that this does not show AI is unhelpful in other settings or rule out gains in future work.

Why did other studies find productivity gains?

The studies examined different work. In the cited writing experiment, ChatGPT cut task time by 40%; in the consulting experiment, AI users were faster on tasks within the tools’ capabilities. Results depended on task and setting, and one task outside the consulting experiment’s capability range had a worse accuracy result with AI.

What should employers measure when evaluating AI?

They should track time for the complete workflow, including review, testing, corrections and discarded output, while also measuring quality. A fast first draft alone does not show whether the full task was completed faster or correctly.

Source: ThorstenMeyerAI.com

You May Also Like

The Smart Investor’s Tool: Fee-Transparency Calculator For Your Portfolio

A new online tool helps retail investors quantify the lifetime dollar cost of investment fees, promoting better fee awareness and decision-making.

Does AI Writing Wear You Out? These Six Patterns May Be Why

An arXiv preprint compares 14 language models with human writing and identifies six patterns that may make AI text harder to read.

Fair-value appraisals for used GPUs and AI hardware

A new manual valuation method for used AI hardware aims to establish transparent market pricing, helping brokers resolve price disputes and improve resale accuracy.

The Menu: What Ten Answers Reveal

A detailed analysis of how ten jurisdictions respond to automation, AI, and income risks, revealing patterns and political choices in their policies.