The Job Everyone Does That No One Can Admit To

Here is the whole experiment. Two people produce the same output, word for word. The only difference is a single sentence: one of them admits an AI helped. That sentence alone is enough to drag the "lazy" rating down ^[1]. The work was identical. The confession was the crime. The same workplaces that punish the confession insist the check is mandatory. IBM's top talent officer warns against what she calls "cognitive surrender" — workers simply accepting whatever the machine says ^[2]. Companies know AI needs rules-based guardrails before it gets anywhere near a decision, and plenty deploy it without waiting for them ^[3]. And the checking does get done: in one Australian survey, six in ten workers say they've caught the machine making a mistake on the job ^[3]. The fear behind the secrecy isn't paranoia, either. Princeton's Sayash Kapoor found that in roughly nine of every ten cases where a company blames AI for layoffs, there is no working AI ready to do the eliminated jobs — the technology is the excuse, not the cause ^[4]. A worker who hides the tool is reading the room correctly. Ben Thompson of Employment Hero put the two ends of the bind in one breath: a technology we hand people and tell them to use, while they are simultaneously afraid of being fired for using it ^[5]. So the hour goes underground. Across 47 countries, 57 percent of employees use AI without telling anyone at work ^[1]. Half of American millennial and Gen Z workers hide their usage outright ^[6]. Some feel guilty using the very tool they were handed ^[5]. A hidden hour is an unbudgeted one. What firms track is adoption — token consumption on a dashboard, whether you used the tool at all, folded into performance reviews. Developers have a word for the work of keeping these tools in line: "botsitting" ^[7]. Nothing on the dashboard measures whether the checking was any good. Guy Bates, who leads AI research at the University of Auckland, has a way of putting the gap.

the adoption of AI may be rising faster than the confidence in what ‘good’ AI use looks like. — Guy Bates

Nobody is trained for it, either — workers teach themselves on TikTok because the classroom version does not exist ^[5]. And when a government finally wrote these jobs down, the verifier wasn't on the list. China has formalized more than twenty AI professions since 2021 — data collection, annotation, model tuning — and none of them is the person who checks a model's output inside a firm ^[8]. So the hour never shows up in the results. Eighty percent of workers say AI lifts their own productivity; only 37 percent say it reaches their company's operating profits, a number flat year over year ^[9]. Against more than $200 billion in infrastructure spending, the macroeconomic numbers stay modest ^[10]. The researchers who ran the disclosure experiment suggest secrecy "may explain" why individual gains keep failing to appear in national statistics ^[1]. Nobody designed it this way; the bind is what emerges when the mandatory check is also the unmentionable one. The work has a size, at least by self-report: workers say 37 percent of the time AI saves them goes back into checking its output ^[11]. But that is a questionnaire answer. No ledger counts the hour where it actually happens — which is rather the point of hiding it. The picture would break on two observable events: a replication that finds no disclosure penalty, or a firm that formally budgets and titles the verifier while the gap between individual and company results persists. Neither has shown up yet. Which is the strange thing, because the checking hour does have a price — everywhere except inside the firm. At the layer where models are trained, the people who review and correct machine output bill up to $200 an hour ^[12]. At the hiring boundary, recruiters now command a premium for authenticating actual humans against a flood of AI-written applications ^[13]. Out there, the one job everyone quietly does costs $200 an hour. In here it costs nothing — because no one can say they're doing it.