The phrase “AI psychosis” is intentionally provocative, but the underlying critique is more useful than the label. What Box CEO Aaron Levie is really pointing at is a familiar Silicon Valley failure mode: executives get dazzled by a demo, then mistake possibility for deployment.
<> 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./>
That line lands because it captures the central problem with today’s AI boom. The people signing off on strategy often see a polished prototype, a slick contract generator, or a chatbot that sounds competent in a sandbox. What they do not see is the grind behind the scenes: data cleanup, edge cases, security review, integration failures, human QA, and all the unglamorous plumbing that turns a demo into something a business can trust.
And that matters because the industry is already using AI as a justification for real-world restructuring. Tech layoffs in 2026 are nearing the full-year total from 2025, with 115,430 cuts across 152 companies in the first five months alone, according to Layoffs.fyi as cited by TechCrunch. That does not prove AI is causing every cut, but it does show how quickly narrative becomes policy once CEOs decide the machines are ready.
The uncomfortable part is that the evidence for sweeping AI productivity gains is still mixed. A meta-analysis published in California Management Review found “no robust relationship between AI adoption and aggregate productivity gain,” while a March NBER paper found gains but also described a “productivity paradox” in which perceived gains outpace measured ones. In other words: executives may feel the transformation, but the balance sheet often says otherwise.
That gap is why I think “AI psychosis” is the wrong literal phrase but the right warning signal. It is not a diagnosis; it is a description of organizational overconfidence. The more accurate diagnosis is probably distance: distance from operations, distance from implementation risk, and distance from the employees who have to make brittle AI systems behave like reliable tools.
- Demos are cheap; production is expensive.
- A prototype is not a workflow.
- A smart assistant is not a replacement for judgment.
- A good benchmark is not proof of business value.
That is why developers should be skeptical when leadership talks as if agentic AI is already replacing whole teams. Current systems can accelerate drafts, boilerplate, and narrow tasks, but developers still absorb the failures: hallucinations, bad integrations, flaky outputs, security issues, and the endless edge cases that make software real.
MIT’s recent agent research reinforces that caution: current systems still fall short of human-quality performance in many cases, even if researchers expect models to reach roughly 80%–95% success on most text tasks by 2029 at a minimally sufficient quality level. That sounds impressive until you remember that “minimally sufficient” is not the same thing as dependable, and definitely not the same thing as safe.
So the real story is not that CEOs are losing their minds. It is that too many of them are confusing strategic enthusiasm with operational understanding. In AI, the last mile is the whole game. Ignore it, and you do not get transformation—you get expensive chaos.
