The CEO 'AI Psychosis' Problem: Why Vibe Coding a Prototype Doesn't Mean You Can Fire Half Your Staff

There is a specific genre of email making the rounds inside companies right now. A CEO discovers large language models, gets visibly excited, and blasts an all-hands message that boils down to a single demand: everyone must start using these tools immediately, or start looking for work elsewhere. Sometimes it comes packaged with consultants, internal AI hackathons, or office hours. In the worst cases, it arrives with a token-usage leaderboard, which manages to misunderstand the technology on two levels at once. Anyone who has actually learned to use these tools well knows tokens are a scarce resource to be spent carefully, not a vanity metric to be maximized. Rewarding people for burning the most of them is a near-perfect way to teach bad habits.

Techdirt's Mike Masnick, writing about four separate versions of this email forwarded to him over three months, lands on a useful observation. He genuinely believes LLM tools are powerful and important. He also believes the most important word in the whole conversation is willingly. Someone who chooses to pick up these tools as an assistant to their own work can get real leverage out of them. Someone forced into it under threat of termination will never learn to use them well. That distinction is where the CEO email genre quietly falls apart.

Levie's Diagnosis: Distance From the Last Mile

The most precise explanation of what is going wrong comes from someone with no incentive to downplay the technology. Box CEO Aaron Levie, an open AI enthusiast, argues that CEOs are uniquely prone to what people have started calling "AI psychosis" precisely because they sit so far from the last mile of work where value actually gets generated. When an executive plays with an agent, they see the happy path. They build a product prototype and feel a jolt of revelation. What they do not see is the next ten or twenty steps: reviewing the generated code before it touches production, fixing the issues that surface, wiring the new thing into everything that already exists.

Levie's examples cut close. "Look, I generated a contract," the CEO says, without having to verify a single term before it goes to the counterparty, and without reconciling it against every prior agreement. The demo works. The demo always works. The work that turns a demo into something a business can actually rely on is invisible from the corner office, which is exactly why the corner office keeps underestimating it.

Masnick adds a sharp caveat worth repeating: the term "AI psychosis" is genuinely bad, and psychologists and psychiatrists have objected that it is both inaccurate and potentially harmful. The underlying behavior, though, is real and observable. CEOs are going overboard, and Levie's account of why holds up.

Cargo Cult Management

The pattern has an older name. A CEO knows that somewhere in the organization, people are typing at computers and useful things come out the other end. So when that same CEO types at Claude Code and watches useful things appear, it registers as the same activity. It is not. All the steps the employees were quietly handling still have to happen. They have not vanished because the CEO produced a working artifact in an afternoon.

The failure mode is the leap from "I built a thing" to "therefore anyone can build a thing, so why do we need all these people." Making something work is a different problem than making it work well, or well at scale, or well at scale inside a specific regulated environment. The reason a company employs a crowd of people is usually to fill in the small, unglamorous, critical details that never reach the CEO's desk: security, legal compliance, accessibility, edge cases nobody demoed. Agentic coding tools can genuinely help with some of that. They do not replace the judgment of people who know what to look for, which is the entire reason you hired experienced people in the first place.

This is also why Masnick argues the strongest use of these tools is personal and narrow. Building a custom utility to handle one specific task you understand deeply is where the technology shines. Building a mass-market product that strangers will rely on, safely, is a different category of problem, and the gap between the two is where overconfident executives get hurt.

The Layoff Story Doesn't Hold

There is a darkly comic quality to watching an executive go all in on a technology and immediately conclude it justifies cutting half the staff. Companies betting they can shed large numbers of workers because of LLMs are likely to discover their error quickly. The real leverage of these tools, used well and used voluntarily, is helping employees accomplish more. That argues for more capable humans who know how to work productively, not fewer humans.

There is a more cynical version of the layoff announcement, and it deserves naming. A lot of companies that are pointing at AI to explain headcount cuts are simply using it as cover. They over-hired during an easier period, and "AI efficiencies" is a far more flattering story to tell Wall Street than "we made bad staffing decisions." The technology becomes a convenient narrative laundering operation for ordinary management mistakes.

Levie's prescription remains the right one, and it is not anti-AI in the slightest. Executives should learn how the technology works, including, especially, its limitations. If a CEO is convinced the prototype they vibe coded is production-ready, the instructive move is to let them ship it and watch what happens. If they think a machine-generated contract is as sound as one a lawyer reviewed, the legal bills will eventually supply the missing education. The tools are powerful. A leader who believes that power erases the work of the people around them has revealed something about their own judgment, not about the future of employment.