Why the CEO's Job Is the Riskiest Seat in the Room

A Box co-founder, a YC CEO, and Microsoft's own AI chief have all used the same unusual word this year: psychosis.

Welcome Back,

In May, Box co-founder and CEO Aaron Levie posted something unusual to X, a warning aimed squarely at his own peer group. "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," Levie wrote, according to TechCrunch. His argument, executives play with a chatbot, watch it generate a prototype or a contract, and leap straight to believing AI agents can do the whole job, without ever having to be the person who reviews the resulting code, finds the bugs, or catches a hallucinated software library before it ships.

"AI psychosis" isn't a formal medical diagnosis, it's a term that's emerged from case reports, journalism, and now, increasingly, from tech leaders describing themselves and their colleagues. The American Psychiatric Association's own news outlet has published a special report on the phenomenon, and Microsoft's AI CEO Mustafa Suleyman has said publicly that he's genuinely concerned about it. What makes this moment worth paying attention to is how specifically the described risk factors map onto the daily reality of running a startup.

Today we look at what the term actually describes and doesn't, why founders in particular keep coming up in this conversation, what a leading AI executive says worries him most, and the important line between the genuine clinical cases and the more common, less severe version of this playing out across the tech industry right now.

🧠 What This Term Actually Describes

It's worth being precise here before going further. "AI psychosis" is not a formal diagnosis in any psychiatric manual, it's a colloquial term that's emerged to describe cases where intense chatbot use appears connected to worsening paranoia, conspiratorial thinking, or delusional beliefs, sometimes in people with a prior history of mental illness, and in some reported cases, in people with no documented history at all.

The American Psychiatric Association's own news publication has been direct about the limits of what's actually known, noting that robust, peer-reviewed epidemiologic research on AI psychosis risk does not yet exist, and that the risk factors being discussed should be treated as exploratory and hypothesis-generating, not settled science. The reported cases have been documented through case series and news reports, not population-level or controlled studies.

That distinction, real and documented but not yet rigorously studied at scale, matters enormously for how this conversation should be handled. The genuine clinical cases described in psychiatric literature are, by every available account, rare. What's changed recently is that prominent tech leaders have started applying the same word to a much more common, much less severe pattern playing out inside their own industry.

Genuine clinical cases remain rare and are still poorly understood, according to the American Psychiatric Association's own reporting.

🎯 Why CEOs, Specifically

Levie's specific argument is worth sitting with because it's genuinely structural, not just a personality critique. CEOs experience AI at a particular altitude, they see a demo, a slick prototype, a generated first draft, and that experience shapes their beliefs about what the technology can do across the entire company. They're rarely the ones who have to sit with the model's actual failure modes.

That's not a claim about anyone's mental health, it's a claim about incentive structure and information distance. The people closest to the actual friction of shipping AI-assisted work, engineers, not executives, tend to have a far more grounded, and often far more skeptical, view of what these tools can currently do.

😴 When Sleeplessness Becomes a Badge of Honor

The clearest illustration of how this language has spread through founder culture came from Y Combinator CEO Garry Tan. At SXSW in March, Tan told Bill Gurley he sleeps four hours a night because he's so energized building with AI coding tools that he no longer needs medication to stay alert, and he called it "cyber psychosis," estimating that a third of the CEOs he knows are in a similar state.

What's genuinely striking, and worth naming directly, is the tone. This wasn't framed as a warning, it was framed as exciting, almost aspirational, according to commentary on the panel. That inversion, a founder publicly describing altered sleep and obsessive tool use as a marker of ambition rather than a sign something's wrong, is arguably more revealing about startup culture than the clinical question itself.

Reduced sleep and obsessive tool use have, in some founder circles, been reframed as signs of ambition rather than warning signs.

🪞 "Competence Addiction," Not a Feature Shipped

One founder writing on this topic, who works in growth at an AI company, offered a genuinely candid, first-person account of what this pattern actually feels like from the inside, describing it as "competence addiction." The core observation, unlike social media, which offers external validation from other people, AI coding tools offer internal validation, a feeling of being smart, fast, and capable that comes from the work itself.

That last line is genuinely the sharpest insight in this whole conversation. Most workplace warning signs, missed deadlines, declining output, visible burnout, are legible and prompt concern from colleagues. A founder who's sleeping four hours a night but shipping more code than ever looks, from the outside, like a success story, not a person who might need to step back. The behavior that would normally trigger concern is instead being actively celebrated.

An entire ecosystem of tools has emerged around this same enthusiasm, open-source projects that let a single founder role-play running a "Board of Directors" of AI agents with titles like CTO and CMO, some with tens of thousands of GitHub stars. That's a genuinely interesting cultural artifact, tooling built explicitly around the fantasy of running an entire company through AI delegation alone.

⚠️ What Actually Worries the People Building This Technology

It's genuinely notable that the concern here isn't limited to outside critics, it includes the people at the top of the companies making these products. Microsoft AI CEO Mustafa Suleyman told The Telegraph that for many people, talking to a chatbot is "a highly compelling and very real interaction," and that concerns around AI psychosis, attachment, and mental health "are already growing."

That last point is a genuinely important business detail too, AI companies face a real, uncomfortable incentive tension here. Users who form strong attachments to a chatbot's personality are loyal, engaged users, and changes to that personality, like model deprecations, can trigger real public backlash, exactly what happened when OpenAI retired GPT-4o. Suleyman's warning is notable precisely because it comes from inside a company with a direct commercial interest in maximizing engagement with its own chatbot.

Some AI companies face a genuine tension between user engagement and the risk of unhealthy attachment to a chatbot.

🧠 AI Spotlight Analysis

There are genuinely two separate conversations happening under the same term, and conflating them does a disservice to both. One is a real, still poorly understood clinical phenomenon, rare, severe, and genuinely worth continued psychiatric research and caution, particularly for anyone with an existing vulnerability to psychosis or delusional thinking. The other is a much more common, much less severe cultural pattern, founders and executives normalizing overwork, sleeplessness, and inflated confidence in AI's current capabilities, and borrowing clinical language to describe it, sometimes seriously, sometimes almost as a badge of intensity.

Founders genuinely do sit at a specific intersection of risk factors worth naming honestly, high-stakes decision-making under real financial pressure, cultural incentives that reward extreme hours and constant AI experimentation, and structural distance from the day-to-day friction that would otherwise correct an overly optimistic read on the technology. None of that means every founder using AI heavily is at risk of anything resembling a clinical crisis. It does mean the incentive structure around founder culture right now doesn't reward the kind of self-monitoring that might catch a problem early, whether that problem is burnout, poor judgment, or something more serious.

Levie's framing is genuinely useful, whether or not "psychosis" is the right clinical word for what he's describing. The underlying point, that distance from the actual, unglamorous work of implementation breeds overconfidence, is a real and testable claim about how executives form beliefs, independent of the more serious, and much rarer, clinical cases documented elsewhere.

🔗 Sources and Further Reading

Thanks for reading AI Spotlight.

Our mission is simple: deliver clear, trustworthy, and actionable AI insights that help professionals stay ahead without the hype.

Today's edition touches on mental health. If anything in this piece feels personally relevant, please consider reaching out to a mental health professional or a trusted person in your life.