Does AI Increase Verification Work After First Drafts?

Does AI increase verification work after automating first drafts?

Sometimes AI shifts effort toward checking, correction, and integration, but evidence does not establish a universal increase in verification minutes. A faster first draft is only one stage: workers may still check claims, restore context, correct errors, and prepare material for acceptance. Whether this adds time depends on the task, evidence, error costs, and review design.

The Microsoft Research study “The Impact of Generative AI on Critical Thinking” surveyed 319 knowledge workers, who shared 936 examples of AI use. Participants described critical thinking shifting toward verification, integration, and task stewardship. This indicates a reported change in effort's kind, not timed checking or increased workload. The American Economic Review: Insights field experiment “Shifting Work Patterns with Generative AI” followed 7,137 knowledge workers in 66 firms for six months. In its latter half, tool users spent two fewer hours on email weekly; researchers detected no change in task quantity or composition from individual access. The experiment did not isolate verification time. Compare tasks through acceptance, including review and rework, before deciding whether AI increased workload or saved time. Faster drafting alone establishes neither. This is a workflow question, not a prediction about job loss.

Sources: The Impact of Generative AI on Critical Thinking: Self-Reported Reductions in Cognitive Effort and Confidence Effects From a Survey of Knowledge Workers; Shifting Work Patterns with Generative AI

What counts as verification work once a draft exists?

Verification starts after a draft exists, but it is not a synonym for polishing. It asks whether the material can be relied on for the purpose it is meant to serve: whether its claims are supported, its sources say what the draft attributes to them, its numbers and assumptions hold, and its contents fit the assignment. Proofreading catches spelling, grammar, or formatting defects. Deciding whether a claim belongs in the deliverable is another judgment: a statement may be accurate yet irrelevant, too uncertain, or outside the brief. Consider an example: a worker must send a short report by email summarizing a set of supplied documents for a manager. Before writing, the worker locates the current versions and identifies what question the report must answer. That gathering stage has its own time cost; it is not checking AI output. A drafting tool may then organize the material into a plausible first pass. Verification begins when the worker traces consequential statements back to those documents, confirms that a quoted figure uses the right period or definition, checks whether a conclusion follows from the cited evidence, and notices whether the draft has filled a gap with an assumption. What happens next should be recorded separately. If a figure is attached to the wrong period, correction includes changing it and perhaps revising the conclusion that depended on it. Integration means making the corrected material coherent with the rest of the report and the requested format. Approval or sign-off is the decision that the result meets the sender's standard and may be handed on. If a manager or another reviewer later finds a defect and returns the report, that downstream rework is labor too. Combining these stages under “verification” can hide where effort went and whether a problem originated in drafting, checking, integration, or acceptance. The distinction matters because a draft is an intermediate object; the work being compared is the accepted deliverable. The MIT Economics account of a writing experiment reports faster completion and higher evaluator-rated quality for bounded writing tasks, but notes that these tasks did not require precise factual accuracy or company or customer context, and that researchers did not hire fact-checkers. It therefore helps show what measured drafting and evaluation can capture, while leaving factual checking in a real assignment unresolved. The Microsoft Research study adds a different kind of evidence. Its survey of 319 knowledge workers collected 936 examples, and respondents described critical thinking shifting toward information verification, response integration, and task stewardship. That supports treating these as recognizable stages of AI-assisted work, not claiming that every worker checks every draft or spends longer doing so. A worker can therefore time gathering, drafting, substantive checks, correction, integration, approval, and downstream rework as distinct stages. If drafting contracts while checking remains steady, checking takes a larger share of total time without adding minutes.

Sources: The Impact of Generative AI on Critical Thinking: Self-Reported Reductions in Cognitive Effort and Confidence Effects From a Survey of Knowledge Workers; Study finds ChatGPT boosts worker productivity on some writing tasks

What did workers mean when they said critical thinking shifted?

The survey’s phrase “shift in critical thinking” describes a change in the activities workers said they used while completing AI-assisted tasks. It does not mean that the study timed a longer review stage. In “The Impact of Generative AI on Critical Thinking: Self-Reported Reductions in Cognitive Effort and Confidence Effects From a Survey of Knowledge Workers,” 319 knowledge workers contributed 936 examples of using generative AI in work tasks. The authors report a shift toward verification, integration, and task stewardship. These categories name possible post-draft work, not its duration. The distinction matters because “verification” here should not be reduced to checking whether a citation exists. A worker might inspect information, compare a response with the task’s requirements, combine it with other material, or decide what to accept and take responsibility for. The study describes these as a qualitative shift, not a universal sequence performed on every task. Integration and stewardship extend beyond checking factual claims to making a response usable and remaining involved in the work. The number 936 counts examples supplied by the 319 participants, not 936 randomly sampled workers, jobs, or independently timed drafts. The paper analyzes these examples qualitatively and examines confidence-related factors quantitatively. It says greater confidence in GenAI was associated with less critical thinking, while greater task-specific self-confidence was associated with more. These are associations, not evidence that confidence caused workers to check more or less, or that the pattern occurs at the same rate across knowledge workers. That boundary changes how to use the result. The examples can guide a worker looking for stages to observe: locating support for a claim, checking whether the answer addresses the request, reconciling it with other inputs, and deciding who owns the final response. It cannot set a time budget, establish whether the same work happened during unaided drafting, or show whether checking or total task time rose. There are ordinary limits to retrospective self-report. Participants described their own recent work rather than having researchers independently time and inspect each workflow. Accounts reveal perceived effort but remain subject to selection and recall effects. This was not a randomized comparison of AI-assisted and human-only task time. The authors’ finding is therefore useful as a map of possible effort shifting, not as a prevalence estimate or causal test. For a practical decision, treat the categories as prompts for a local observation. In one recurring task, note what must be verified, integrated, and owned after the first pass, then record the time and corrections through acceptance. If these stages take fewer minutes because drafting shortened, their greater visibility does not mean checking grew in clock time. The survey identifies what to observe; comparison answers how minutes changed. This is why a reported shift in kind should not be paraphrased as “workers spent more time verifying.” The study supports the former statement; the latter requires timed comparisons that its survey design does not provide. Keeping those claims separate preserves the finding without overstating it.

Sources: The Impact of Generative AI on Critical Thinking: Self-Reported Reductions in Cognitive Effort and Confidence Effects From a Survey of Knowledge Workers

What does real workplace time-use evidence add—and leave open?

Field evidence adds an important correction to a simple story about AI: using a workplace tool does not necessarily add net time. But the available experiment does not answer whether checking a generated first draft takes longer. In “Shifting Work Patterns with Generative AI,” published in American Economic Review: Insights, researchers report a randomized field experiment across 66 firms and 7,137 knowledge workers. Randomly selected workers received access to a generative AI tool integrated into applications they already used for email, meetings, and writing. The study followed the experiment for six months, making it evidence about use in workplace settings rather than only a short laboratory exercise.

The abstract reports that, in the second half of the six-month experiment, 80% of treated workers who used the tool spent two fewer hours on email each week and reduced time working outside regular hours. It complicates any claim that AI invariably creates hidden labor that cancels out faster production. The measured outcome, however, was time spent on email and work outside regular hours. The result does not say that those two hours were saved specifically by automating first drafts, nor that workers spent less time verifying AI-generated text.

The researchers also report that, aside from these individual time savings, they did not detect shifts in the quantity or composition of workers’ tasks resulting from individual-level AI provision. This is a useful boundary on claims about immediate redesign: in this experiment, providing access to individual workers did not produce a detected overall change in the mix or amount of their tasks. It is not evidence that every task stage remained unchanged. A worker might spend less time composing an email and more time checking a particular claim while the broader task still counts as email work. Aggregate measures of task quantity or composition can miss that substitution inside a task.

That distinction keeps the result in proportion. The experiment is counterevidence to the idea that workplace AI use must always increase net time, and it supplies a concrete example of time changing after access to a deployed tool. It does not isolate verification, correction, accuracy, quality, or downstream review. These unanswered questions matter because a reduction in one measured category cannot be translated into a verified time saving for another workflow.

For someone assessing a recurring task, the study therefore changes the question rather than settling it. Do not infer that review work rose from the existence of checking, and do not infer that checking is costless from a reported email-time reduction. The relevant comparison is the full task carried to the same acceptance standard, with drafting, checking, correction, integration, and any downstream review counted. This field experiment does not provide that task-level ledger; its result is a reason to allow for real time gains while measuring whether they survive review in the work at hand.

Sources: Shifting Work Patterns with Generative AI

Why one role can contain both faster and harder-to-check tasks

An AI-assisted workflow can contain both a quicker step and a more demanding review step. A role average can hide that split: a standard summary may become faster, while a context-sensitive recommendation may be harder to check. Neither an occupation’s exposure label nor a tool’s ability to produce text tells you how much verification a particular worker actually performs. Examine the recurring task, its inputs, acceptance criteria and consequences for error. The randomized field experiment described in “Navigating the Jagged Technological Frontier” makes this variation visible. Researchers worked with Boston Consulting Group to develop realistic consulting tasks and studied 758 consultants, randomly assigning access to GPT-4, GPT-4 with a prompt-engineering overview, or no AI. For 18 tasks judged to fall within the tool’s capability frontier, participants with AI completed 12.2% more tasks and worked 25.1% faster on average, with higher solution quality. On a complex managerial task selected as outside that frontier, AI users were 19% less likely to produce a correct solution. This shows uneven performance in this experiment, not a general law about routine or complex work. The designed tasks also do not capture all workplace complexity. That difference matters for verification because the work needed to trust an output depends on what can be checked. As an editorial inference, structured inputs and clear criteria may make a draft easier to compare against source records or a fixed format. Accessible evidence and a worker able to spot mismatches may bound checking. For an open-ended task with missing context, conflicting evidence or a judgment that must fit a specific organization, a plausible answer can be harder to validate. The worker may need to trace claims, find omissions, correct reasoning and explain fit, adding time after a quick first pass. The study did not measure verification minutes, so this is a mechanism to examine in a workflow, not its finding. A single role may include both patterns. Consider a recurring client update: arranging supplied figures into a standard summary may be a bounded step, while explaining an unexpected result may require checking definitions, dates, exceptions and what the client already knows. AI may help with the first and offer a fluent but incomplete account of the second. This shows why a role average, or the shared format of writing, cannot establish equal review burden. The relevant comparison asks whether each task’s final result meets its own acceptance standard, and what checking and correction it took to get there. The International Labour Organization’s 2026 brief adds a boundary to exposure measures. It says such indicators rely on static descriptions of current tasks, omit adoption and economic-feasibility constraints, and indicate what AI could do rather than what will happen in practice. Exposure can flag tasks to examine; it cannot establish workplace use, workflow design or increased review time. Start with one named task bundle you actually perform, separate steps, and record what sped up, needed more checking and reached acceptance. That local evidence is more useful for a workflow decision than extrapolating from an occupation label.

Sources: Navigating the Jagged Technological Frontier; New ILO brief explains what AI exposure indicators reveal about jobs

What a writing productivity experiment counts as quality—and what it leaves out

A writing experiment can record faster completion and higher evaluator ratings without measuring whether a finished piece is factually sound. In MIT Economics’ account of the experiment, 453 college-educated professionals completed two short writing tasks related to their occupations. The fields included marketing, grant writing, consulting, data analysis, human resources, and management. The assignments included a grant cover letter, an email about organizational restructuring, and an analysis plan using supplied customer data. They lasted 20 to 30 minutes, and participants receiving ChatGPT-3.5 for the second task finished 11 minutes faster than the control group; average quality evaluations were 18 percent higher.

This is evidence about performance under a bounded writing assignment: tool access changed measured task time and ratings in this setting. Evaluators were experienced professionals in the participants’ occupations, judged submissions as if seeing them at work, and did not know which submissions had ChatGPT assistance. This design includes an independent assessment, but “quality” means the evaluators’ rating under the study’s conditions. It does not establish that every claim was correct, every source supported its use, or the text was ready for a specific organization to rely on.

The limits matter because real writing often has acceptance requirements the experiment could not reproduce. MIT Economics notes that the assignments did not require precise factual accuracy or context about a particular company’s goals or a customer’s preferences. Participants were anonymous, so researchers could not expect them to know those local details. The tasks also came with explicit instructions, whereas workplace requests can be more open-ended. And the researchers did not hire fact-checkers to assess accuracy. These boundaries define what the reported gains establish.

For a real deliverable, factual correctness, currentness, source support, alignment with internal policy, fit with a customer’s circumstances, and approval may each be part of the acceptance bar. A general evaluator score can capture qualities such as clarity or persuasiveness while leaving those task-specific conditions untested. The study should therefore not be summarized as showing that AI produces verified workplace writing 18 percent better, or that checking necessarily cancels its speed benefit. Nor should its 18 percent rating difference be treated as an estimate of error reduction: the reported measure was evaluator-assessed output quality, while factual checking was not performed.

The researchers themselves identify fact-checking and prompt writing as reasons real-world speed benefits could be smaller. That caveat is directly relevant to verification, but it is not a measurement of how many minutes checking takes, how often errors occur, or whether checking erases the gain for a particular job. The experiment can show a promising first pass while leaving review outside its measures.

For your own task, compare like with like: take representative pieces through the same standard of factual checking, local-context review, correction, and approval, and count that time alongside drafting. A quicker draft and stronger general rating are encouraging signals, not the final productivity measure. The relevant result is a checked, fit-for-purpose deliverable accepted under your actual requirements, including any review shifted to another person.

Sources: Study finds ChatGPT boosts worker productivity on some writing tasks

When faster output still creates review or downstream work

More output is not by itself less work. An AI tool may produce more code, text, or analysis while a worker still has to test it, check fit with the assignment, correct defects, and prepare it for use. Those later steps may be small, substantial, or unnecessary. Output count alone cannot tell us which. To judge a workflow, count work through an accepted deliverable, including review and handoffs. The Bank for International Settlements’ “Generative AI and labour productivity: a field experiment on coding” shows the distinction. In a field experiment at Ant Group after CodeFuse was introduced, researchers compared programmers with access to the coding assistant with comparable programmers without access. The BIS summary reports code output rose 55% for tool users, measured by lines of code. About one third came directly from generated code; the rest was attributed as likely efficiency in other coding tasks. Gains were statistically significant primarily for junior staff. Senior programmers used the tool less frequently, while acceptance of suggestions did not vary by experience. This is evidence of increased measured code production in that setting, not a count of finished, tested software or review minutes. Lines of code can measure output, but the summary does not report whether additional lines met the same acceptance threshold, how much testing or correction they required, or whether another worker later repaired them. It would be a mistake to treat the 55% increase as a 55% saving in labor. That calculation is unsupported. For a real coding task, follow the work until it is usable: clarify the request, produce or modify code, run tests, inspect the result, correct failures, integrate the change, and handle review or rework after handoff. This is a practical measurement rule, not a finding from the BIS experiment. If AI shifts effort from production to testing, the first pass may shorten while a reviewer inherits extra work. If generated output passes existing checks and reduces total effort, there may be a real gain. The study’s output result cannot distinguish these possibilities. The OECD review, “The effects of generative AI on productivity, innovation and entrepreneurship,” explains why one result should not be stretched across every task. Reviewing experiments, it concludes that effects depend on task and user experience, with clearer benefits on well-defined, bounded tasks and a need for human expertise and oversight, especially when complex work requires contextual judgment. It identifies limited evidence about longer-term business effects and workers’ understanding of AI limitations. This synthesis is not a direct measure of review time in the CodeFuse experiment. The practical test is whether more output becomes more accepted work without shifting uncounted checking elsewhere. For a recurring task, record testing, context checks, correction, approval, and downstream rework alongside production time. Compare equivalent work at the same quality bar. If output rises but review expands, count both; if review catches costly errors, its value is not mere overhead. Credit productivity only after including the deliverable and its handoffs.

Sources: Generative AI and labour productivity: a field experiment on coding; The effects of generative AI on productivity, innovation and entrepreneurship

Does checking erase the time saved?

The evidence cannot give a general yes or no. Studies do not jointly time baseline work, drafting, verification, correction, rework, and delivery at the same quality threshold. Checking may take more minutes, the same minutes, or fewer than the work it replaces. Faster production alone does not settle the balance. The difference between a larger share and more minutes matters. If drafting falls from an hour to twenty minutes while checking rises from ten to fifteen, checking takes a bigger portion of the workflow but combined time still falls. If checking rises to fifty minutes, total time may exceed the original hour. These are arithmetic illustrations, not study results. A survey finding that workers describe more verification, integration, or stewardship identifies a shift in reported thinking; it does not establish minutes or the net change. The Microsoft Research study surveyed 319 knowledge workers, who provided 936 work examples, and found critical thinking shifting toward verification and integration. Its self-reports do not measure checking time. Other studies answer different questions. In the bounded writing experiment summarized by MIT Economics, 453 college-educated professionals completed occupation-specific writing tasks; the account reports 40 percent less completion time and 18 percent higher evaluator-rated quality with ChatGPT access. But assignments did not require precise factual accuracy or company and customer context, and researchers did not have fact-checkers assess accuracy. This shows faster performance on those tasks, not time needed to verify an accepted workplace deliverable. The American Economic Review: Insights field experiment across 66 firms and 7,137 knowledge workers found that treated tool users who used it spent two fewer hours on email weekly in the second half of six months. Researchers detected no task quantity or composition shift from individual access. That counters a claim that AI use necessarily adds net time, yet does not isolate checking within a task. The BIS coding field experiment reports code-output growth above 50 percent, varying by programmer seniority; output volume does not tell us how much review or rework followed. A net gain is more plausible when outputs can be checked against accessible, traceable sources, criteria are stable, and the worker has enough subject knowledge to spot errors. A net burden is more plausible when evidence is opaque, mistakes carry serious consequences, expertise is insufficient, or late review pushes correction onto another person. These are conditional mechanisms, not universal cutoffs. Checking also has value beyond minutes: catching an unsupported claim may prevent a costly error, while treating that protection as mere overhead misses part of the work's purpose. Conversely, an elaborate review routine can consume a speed gain if the task is low-stakes and corrections are frequent. The studies do not identify a universal point where either effect wins. Neither shortcut is defensible: drafting time saved does not make checking free, and the presence of checking does not prove productivity vanished. Compare full task time and accepted quality, including corrections and downstream review, for equivalent work. Only that comparison can show whether checking erases the time saved in your workflow.

Sources: The Impact of Generative AI on Critical Thinking: Self-Reported Reductions in Cognitive Effort and Confidence Effects From a Survey of Knowledge Workers; Study finds ChatGPT boosts worker productivity on some writing tasks; Shifting Work Patterns with Generative AI; Generative AI and labour productivity: a field experiment on coding

How can you measure your own task fairly?

To find out whether AI changes the work of a recurring task, compare human-only and AI-assisted attempts through the same acceptance bar. Time the job through checking, correction, integration, handoff, and rework. A shorter drafting stage is not a saving if the accepted deliverable takes longer overall; conversely, checking can occupy a larger share of a shorter workflow while using fewer minutes. This is a practical decision aid, not a validated measurement instrument. Choose one task that recurs often enough to compare, such as preparing a routine internal summary from source documents. Use comparable cases and inputs. Before starting, write down what “accepted” means. For a summary, require claims to trace to supplied sources, figures and names to be correct, and context to fit the reader’s need. Use the same bar for both approaches. Record the date, tool or workflow used, so you do not confuse a changed case mix with a changed process. For each attempt, record minutes by stage, not just the moment the first draft appears. Separate production, verification, correction, integration, and downstream review. Note errors found and whether correction changed acceptance. Count reviewer time: shifted work remains part of the process. Repeat across ordinary cases; one unusual example is a weak basis for changing a workflow. You need not run a formal experiment: a small, consistent log can reveal whether the review burden is stable or highly case-dependent. Then compare like with like. If assisted attempts meet the same bar with less total time, and errors remain manageable at the task’s stakes, retaining AI for that use is reasonable. If total time falls while checking takes a larger share, the workflow may still be faster; report both figures rather than calling the checking increase a net workload increase. If quality falls, required corrections erase the time difference, or errors are costly and hard to detect, narrow AI to a first pass or stop using it for that task. If cases differ materially, collect another comparable batch. The boundary matters because an exposure indicator describes potential task change, not actual adoption or an individual workplace outcome. The International Labour Organization’s “New ILO brief explains what AI exposure indicators reveal about jobs” says such measures rely on static task descriptions and do not account for economic feasibility or adoption constraints; it recommends combining them with evidence of actual labor-market developments. The OECD’s “The effects of generative AI on productivity, innovation and entrepreneurship” reviews experimental findings but identifies workers’ handling of AI-generated errors and longer-term effects as gaps. Measure one recurring task before drawing conclusions about your role or career.

Sources: New ILO brief explains what AI exposure indicators reveal about jobs; The effects of generative AI on productivity, innovation and entrepreneurship

What should you do next with the result?

Use the measured result to make a workflow decision. Retain AI for this task when accepted work improves under your constraints. If drafting is faster but checking varies, limit AI to a first pass and keep review explicit. Stop for this task when you cannot reliably check consequential claims or correction repeatedly consumes the gain. These are choices about one task, not conclusions about your occupation. The verdict is conditional: evidence does not establish a universal increase in verification minutes or net saving. The ILO’s “New ILO brief explains what AI exposure indicators reveal about jobs” calls exposure measures early signals. They rely on static task descriptions, omit adoption and economic feasibility, and indicate what AI could do rather than what will happen in practice. An exposure score cannot substitute for observing your workflow. Reassess if the tool, source material, task requirements, or consequences change, since these can alter what needs checking and what counts as acceptable. Choose one task and apply the retain, limit, or stop rule to its observed result. If you need a broader map of duties facing possible change, AI Proof Work’s free task checker offers task-level change-pressure signals, not a validated displacement probability. Use it for orientation; let your measured task outcome guide this workflow decision.

Sources: New ILO brief explains what AI exposure indicators reveal about jobs

Questions readers ask