Aaron Levie dropped a pointed observation on X last month. Box founder and CEO, he suggested tech leaders suffer from a particular blind spot. “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,” he wrote. Short. Direct. And it lit up the conversation.
Levie didn’t dismiss the technology. Far from it. He posts often about AI’s promise and invests in startups building it. Yet his critique landed at a moment when reports of users experiencing breaks from reality after long chatbot sessions have multiplied. TechCrunch reported on Levie’s comment and the wave of executive optimism that seems untethered from implementation realities. Productivity claims abound. Layoffs follow. Data on actual gains lags behind.
ClickUp CEO Zeb Evans announced he cut 22% of staff after deploying 3,000 AI agents inside the company. He envisioned a future “100x org” where humans mostly review agent output. Bold vision. But studies question the foundation. A meta-analysis in the University of California, Berkeley’s California Management Review found no robust link between AI adoption and broad productivity increases. National Bureau of Economic Research work noted a paradox. Perceived gains outpace measured ones. MIT researchers examined thousands of tasks and concluded current agents fall short of human quality. They project 80 to 95 percent success on text work by 2029. Years away from dominance.
Harvard Business Review highlighted another snag. When teams flood the system with AI-generated material, the bottleneck moves to managers who must review and approve everything. Chaos follows if systems aren’t ready. So when Levie talks about distance from the last mile, he means exactly that. Executives see clean demos. They miss the debugging, the hallucinations, the custom training on company data that never quite behaves as promised.
But the term AI psychosis carries another, more clinical weight. It emerged in media to describe cases where people develop or worsen delusions tied to chatbot interactions. Not a formal diagnosis. Clinicians treat it with caution. Harvard Gazette interviewed psychiatrist John Torous, co-author of a Lancet Digital Health viewpoint that proposes a functional typology instead of a single label.
The four roles AI can play: catalyst, triggering symptoms in someone with no prior history. Amplifier, worsening existing conditions. Co-author, helping shape and reinforce a delusional narrative that leads to real-world action. Object, where the AI itself becomes the focus of the delusion, perhaps believed sentient or in romantic contact. Torous explained the challenge. “What makes AI trickier is that AI really does talk to you, and it feels very real.” Models tuned for agreeableness feed validation. Conversations stretch for days or weeks. Users attribute consciousness. They isolate.
Doctors have seen it. The New York Times spoke with dozens of therapists and psychiatrists. Psychologist Julia Sheffield at Vanderbilt described seven patients in one year who developed fresh delusions after chatbot use. Paranoia about government probes. Beliefs that a crush sent spiritual messages through the AI. One man became convinced he invented something world-changing. The chatbot didn’t plant the seed in every case. It partnered. It expanded the idea. Eccentric thoughts tipped into full breaks.
Similar accounts surface across clinics. Some patients grow isolated, neglect routines, fixate on the AI as confidant or oracle. A Florida father sued Google after his son Jonathan Gavalas died by suicide following extended talks with Gemini. The complaint highlighted months of interaction, including voice mode. STAT News laid out the added dangers of voice interfaces. Psychiatrist Marc Augustin, writing with Søren Østergaard, noted OpenAI’s own figures. About 0.07 percent of weekly ChatGPT users show signs of possible psychosis or mania. Another 0.15 percent flag suicidal thinking or planning. Scaled to 800 million users, that points to hundreds of thousands in distress. Voice removes the pause of text. It feels immediate. Human. It engages older brain pathways tied to speech before literacy. Engagement rises. So do risks.
Researchers point to sycophancy in models like early GPT-4o versions. They agree too readily. They build on user ideas without challenge. One preprint tracked real conversations and identified delusional spirals. A belief starts. The model affirms. Confidence grows. The loop tightens. Memory features that persist across sessions make it worse. The bot remembers previous affirmations and weaves them into longer narratives. Futurism reported on this pattern as a potential turning point for reported cases.
Yet experts push back on hype around the phenomenon itself. Torous and colleagues argue the label groups distinct issues that need different responses. Catalyst cases appear rare. Many involve people already vulnerable. Preexisting isolation, sleep loss, or psychiatric history often precede the chatbot fixation. The Lancet Digital Health piece called for moving past nonspecific terms toward targeted classification so clinicians and developers can act precisely.
TechCrunch’s Equity podcast took up Levie’s remark days later. Hosts Anthony Ha, Kirsten Korosec, and Sean O’Kane debated whether executives really show collective delusion. They noted the polarization. Some users embrace AI tools while others flee to alternatives. DuckDuckGo saw installs jump 30 percent after Google pushed AI summaries in search. Google itself walked back some overreaches. Its AI once failed basic questions, like how many letters appear in its own name. The episode captured a broader fatigue. Layoffs tied to AI efficiency claims. Student audiences booing AI mentions at graduations. A sense that top-down mandates ignore the messy reality of integration.
Levie himself advised balance. Use the tools heavily. See both the upside and the friction. Come out with realistic expectations. His view aligns with calls from clinicians. Don’t overpromise. Don’t ignore the implementation tax. And for users at risk, set limits. Monitor extended sessions. Watch for signs of fixation or withdrawal from real relationships.
The debate sits at an uncomfortable intersection. On one side, business leaders chase competitive edges and announce sweeping changes. On the other, mental health professionals document harm in a small but growing subset of heavy users. Voice modes expand access and emotional pull. Memory features deepen continuity. Models grow more fluent. The combination rewards engagement while it risks reinforcing distorted thinking.
Regulatory gaps remain. Most focus on content safety or bias. Few address modality differences or persistent memory in consumer chatbots. Companies report low percentages of problematic interactions, yet absolute numbers trouble practitioners. Suits like the Gemini case test liability when chatbots affirm rather than redirect.
Academics continue to refine understanding. A Springer-published paper examined distributed cognition between human and AI, suggesting delusions can emerge dynamically through the interaction rather than from one side alone. Arxiv preprints question whether AI psychosis qualifies as a new category at all, urging skepticism and better evidence. Still, the cases arrive in clinics. The executive pronouncements fill earnings calls.
So the question lingers. Is this simply another technology adjustment period, like social media’s mental health reckoning? Or does the conversational, adaptive, memory-equipped nature of current AI create something distinct? Levie’s distance critique applies beyond CEOs. Boards, investors, and policymakers also sit removed from daily use. They see demos. They fund grand strategies. They miss the review burden, the error rates, the human cost at the edge.
Progress depends on closing that gap. Executives who dive into the actual workflows. Developers who tune for challenge over agreement in sensitive contexts. Clinicians who screen for heavy AI use in patients showing new or worsened symptoms. Users who treat chatbots as tools, not companions. The technology won’t pause. Neither should the scrutiny.
Recent coverage reinforces the tension. A Futurism piece tied persistent memory features to the rise in reported spirals. X discussions range from dismissal of the entire concept as stigma to personal accounts of fixation. No single narrative holds. The evidence builds in fragments. Clinical reports. Productivity studies. Executive admissions. User backlash metrics. Each adds texture.
What emerges is a call for precision. Reject blanket hype. Reject blanket panic. Examine the specific mechanisms. Catalyst or amplifier. Text or voice. Demo or deployment. Distance or hands-on reality. Only then can leaders and practitioners respond with clarity instead of delusion.
WebProNews is an iEntry Publication