The Job AI Created Is Checking the Machine

Employees who review AI-generated drafts come away more mentally drained than coworkers who did the same task by hand. It isn't the volume of the work; it's the kind. Validating text that sounds confident while hiding errors means making small decisions without a break, never trusting the surface, and Harvard Business Review's researchers measured that this produces higher fatigue and effort than simply doing the task manually. Workday, tracking where the promised savings actually go, found nearly 40% of the time users believe they're saving is spent correcting the AI's output ^[1]. The technology was sold as work removed. A good part of the promised hour off is a checking hour. The finding holds outside the lab. ActivTrak watched 163,638 employees over three years and found no category of work shrank as AI use climbed: chat time rose 145%, email 104%, time in business-management tools 94% ^[2]. ActivTrak's own read of the numbers is blunt.

AI is being used as an additional productivity layer, not a substitute for existing work. — ActivTrak

At Amazon, corporate staff describe extra hours spent correcting agent errors while managers push for speed ^[2]. The checking layer isn't a quirk of one study or one company. It's what the tool does inside an actual workday. Then comes the uncomfortable part. The work being created is precisely the work AI use erodes. MIT researchers, with colleagues at Carnegie Mellon, Oxford and UCLA, found that as little as ten minutes of direct-answer assistance measurably impaired independent problem-solving; people handed solutions made more errors and gave up faster once the help was removed ^[3]. The verification gap makes it worse: in one survey, 65% of American workers don't independently verify AI answers, and only 25% say their company has an effective process for checking output ^[4]. The new job demands sustained critical judgment from people the mechanism itself is dulling, at firms that mostly have no system for it. Jensen Huang has been explicit about why he isn't worried.

I think my cognitive skills are actually advancing, and the reason for that is because I am not asking it to do the thinking for me. — Jensen Huang

He is defending a skill few workers are offered. The MIT team's own distinction cuts closer: direct answers impair, hints and scaffolding don't. One of the study's authors drew the line himself.

Systems that give direct answers may have very different long-term effects from systems that scaffold, coach, or challenge the user. — Michiel Bakker

Huang cross-examines several models and treats prompting as a craft — the scaffolding version of the work. Most workers get the direct-answer version, then are asked to check it. The dissent doesn't blunt the mechanism; it names exactly who escapes it. Employers, meanwhile, grade the adoption itself. Google and Microsoft track employee AI usage with monitoring tools and characterize resisters as slackers, and Microsoft has weighed folding usage metrics into formal reviews ^[5]. IgniteTech's chief executive laid off nearly 80% of staff for not engaging with AI training ^[6]. Visible use became a survival credential — except that workers also learned visible use reads as replaceable. Atlassian found people who disclose AI use are judged lazier at identical output ^[7], and over half of workers hide their AI use from management while 53% fear relying on it makes them look replaceable ^[8]. They are graded on visibly using the tool they are shamed for visibly using, so they perform it and conceal it at once. Where the gains land is just as measurable. The Bank of England finds productivity growth concentrated in AI-adopting sectors, with software and IT consulting's contribution rising tenfold against the pre-pandemic decade ^[9]. University College Dublin's survey of 4,300 workers found only 4.7% of AI users report any earnings increase, and most of them reinvest the saved time into more work rather than shorter hours ^[10]. The firm gets faster; the worker gets more to do for the same pay. Upwork's researchers call what follows "the Great AI Attrition": the most productive AI users, the ones turning a 40% productivity gain, are twice as likely to quit, most of them overloaded ^[11].