Tech CEOs are apparently suffering from AI psychosis, says Box founder Aaron Levie

A theory circulating in the tech industry holds that senior executives, and especially CEOs, are collectively suffering from delusions of grandeur about artificial intelligence. One of their own has now said it out loud: Box founder Aaron Levie.

"Distant from the last mile of work"

In a post on X, Levie argued that "CEOs are uniquely prone to AI psychosis because they're sufficiently distant from the last mile of work that still has to happen to generate most value with AI." Executives, he wrote, play with AI, build a prototype or generate a contract, and then jump to the conclusion that agents can do the work. They are not the ones who review code, hunt for bugs, spot calls to libraries that do not exist, train models on a company's idiosyncratic contract terms, or spend days combing through agreements for hidden clauses.

Levie is not an AI sceptic: he posts positively about the technology to 2.7 million followers and backs AI startups as an angel investor. His advice to fellow CEOs is to use AI "a ton" to see what it can and cannot do, and to "come out the other side with an appreciation for both the upside and the real work."

Layoffs and the productivity paradox

The context is a bruising year for tech employment. Layoffs.fyi counts 115,430 people laid off by 152 tech companies in the first five months of 2026, close to the 124,636 cut by 275 companies in all of 2025. Many of those announcements cite AI. ClickUp CEO Zeb Evans said he cut 22% of staff after deploying about 3,000 AI agents, aiming for what he calls a "100x org."

The evidence for such gains is thin. A meta-analysis published in UC Berkeley's California Management Review found "no robust relationship between AI adoption and aggregate productivity gain." Research published by the National Bureau of Economic Research concluded that AI adoption did improve productivity, but noted a "productivity paradox, in which perceived productivity gains are larger than measured productivity gains."

Where the bottleneck moves

Researchers at MIT, after studying thousands of worker evaluations of labour-market tasks, found that agents are not yet producing human-quality work in many cases. At the current rate of model improvement, they project, models will handle most text-related tasks at average success rates of 80–95% by 2029, at a minimally sufficient quality level. Work published in the Harvard Business Review suggests that when everyone uses AI to produce more, the bottleneck shifts to the executives who must authorise the output — a recipe, the piece argues, for organisational chaos.