Managers run more of their writing past AI than the employees they supervise

The deference figure ^[4] is the survey number closest to judgment. Checking a draft is editing. Sending it because the tool declared it ready hands over the decision to send, and at roughly double the entry-level rate, managers and executives land near one in four ^[1]. The report's two comparisons use different baselines. On the 70% measure, managers sit 24 points above experienced individual contributors ^[2]; the deference measure is set against entry-level staff ^[4]. All of it comes from 705 employed Americans who already use AI for writing ^[2].

The report's authors call the pattern a managerial paradox and suggest it may reflect the volume of high-stakes writing leaders handle, such as performance reviews and messages to direct reports ^[5]. A manager who writes more consequential messages has more reason to check them. The argument that leaders are outsourcing their judgment, and that leadership development has to catch up, draws its evidence from research on other people ^[11].

That research is specific. A UC Irvine and McGraw Hill preprint covering 3.2 million learning interactions found that after ChatGPT launched in 2022, students' scores rose on problems they could paste into a chatbot, while their proctored results away from the computer fell below pre-AI levels ^[7]. "Part of how AI is making us dumber," Adam Green, a cognitive neuroscientist at Georgetown University, said, is "undercutting the development of learning how to think in younger people that now have an alternative." ^[9]

A skeptic would say a manager with years of decisions behind them is a different case from a student who never built the skill. For the math data and for Green's point, that holds. It holds less well for the paper prepared for the Conference on Language Modeling, which found adults used to automating analytical tasks more likely to abandon hard questions they could otherwise solve ^[8]. Sending a document on the tool's say-so shows confidence in the tool. A 2025 study by Microsoft Research and Carnegie Mellon University of 319 knowledge workers found that higher confidence in generative AI went with less critical thinking, and higher self-confidence with more ^[13].

In my view the survey justifies asking how managers use the check. It does not yet justify a remedial programme for leaders. The Middlebury result turned on mode of use: tutoring preserved unaided performance and full drafting eroded it ^[10]. Omni Calculator's measures, as reported, count volume and deference ^[1][4]. They do not separate a manager who asks AI to explain a point from one who lets it rewrite a review.

The remedies on offer are rules, training in scrutinising AI output, and assessing leaders without AI. The base for rules is thin. In KPMG in Canada's survey of 2,239 employees, 40% did not know what AI controls their employer had ^[12]. The AI-free assessments Gartner expects to spread ^[14] measure the judgment a firm wants to keep, at the cost of testing leaders under conditions they no longer work in. This quarter's choice is what a policy counts. A rule that tracks only how often managers use AI will produce usage data next year, and the Middlebury work found that how students used the tool predicted how they did without it ^[10].