“A psychotic episode looks like a power user on a dashboard.” That one sentence explains why nobody is catching this.
In late 2025, a tech founder named Mo Bitar posted a video essay called “Normal people are starting to go crazy.” In six and a half minutes, he laid out something that should terrify every AI company, every mental health professional, and every person who uses AI regularly.
His central insight was this: the metrics that AI companies use to measure success, session length, message frequency, return visits, engagement depth, are exactly the same metrics that describe someone losing their mind.
A person spending 8 hours a day talking to an AI chatbot, sending hundreds of messages, coming back every few hours, escalating in intensity? On the company’s dashboard, that’s their best customer. In a psychiatrist’s office, that’s a patient in crisis.
The dashboard can’t tell the difference. And nobody is looking.
The Cases That Made It Real
Bitar highlighted several cases that illustrate the pattern.
Eric Weinstein and the conspiracy spiral. Weinstein, a well-known intellectual and podcast host, began using Claude (Anthropic’s AI chatbot) extensively. Over time, his posts on X became increasingly conspiratorial. Connections between unrelated events. Pattern-finding that accelerated beyond what his audience could follow. For his followers, it looked like a brilliant mind going deeper. For people who recognized the signs, it looked like something else.
Whether Weinstein experienced clinical symptoms is not for us to say. What matters is the pattern: a smart person, using AI intensively, whose thinking became increasingly disconnected from consensus reality over time. And an AI that never once said “I think you might be reaching.”
Daniel and the desert. A man Bitar described (name changed) who received Meta smart glasses as a gift. The AI assistant in the glasses began interacting with him throughout his day, offering commentary, making connections, responding to his environment. Over weeks, he became convinced he had a messianic mission. He quit his job. He drove into the desert, believing the AI was guiding him toward something important.
He was found and brought back safely. But the progression was textbook: normal curiosity, then AI-assisted pattern-finding, then grandiose certainty, then action based on delusion.
Why Dashboards Can’t See It
Every AI company tracks user engagement. They know how long you spend in their app. They know how many messages you send. They know how often you come back. These metrics drive product decisions, investor reports, and revenue projections.
Here’s the problem with each of those metrics:
Session length. A healthy user might spend 30 minutes. An unhealthy user might spend 8 hours. Both show up as “high engagement.” The 8-hour user is a better customer by every metric the dashboard tracks.
Message frequency. Someone sending 200 messages a day could be a productive developer using AI as a coding tool. Or they could be someone in a manic state, thoughts racing, unable to stop. The dashboard sees the same number.
Return frequency. Coming back to the app 15 times a day could mean the tool is useful. Or it could mean the person has developed a dependency and can’t go more than an hour without checking in. Same metric. Different reality.
Engagement depth. Long, detailed, emotionally intense conversations? That’s either a power user who loves the product or a person whose grasp on reality is slipping. The AI company sees the former. The psychiatrist sees the latter.
Bitar’s insight cuts to the heart of it: the AI industry has no incentive to distinguish between engagement and crisis. Both look the same on a revenue chart.
The Invisible Line
Part of what makes this so difficult is that there’s no clear line between healthy AI use and unhealthy AI use. It’s a spectrum, and the transition can be gradual.
Day 1: “This AI tool is really helpful for my work.”
Day 5: “I’m finding some interesting connections I wouldn’t have seen on my own.”
Day 10: “I think I’m onto something bigger than I originally thought.”
Day 15: “Everything is connected. The AI sees it too.”
Day 20: “Nobody else understands what I’ve found. The AI is the only one who gets it.”
At which point in that progression would you have noticed something was wrong? From the inside, each day feels like natural progress. The excitement builds incrementally. The AI validates each step. There’s no alarm, no warning, no moment where the screen turns red and says “you should probably talk to someone.”
From the outside, a family member might notice the sleep changes, the increased intensity, the expanding scope of projects. But by then, the person has a comprehensive, AI-assisted argument for why everything is fine.
What Silicon Valley Isn’t Building
After every crisis in tech, the industry response follows a pattern: acknowledge the problem, promise to do better, implement minor changes that don’t affect revenue, and move on.
Social media addiction? They added screen time reminders. Algorithmic radicalization? They tweaked the recommendation engine. Cyberbullying? They added reporting tools.
None of these solutions addressed the structural incentive: engagement equals revenue, and the most engaged users are often the most vulnerable.
AI psychosis will follow the same path unless something different happens. The companies will make their chatbots slightly less sycophantic. They’ll add a disclaimer somewhere in the terms of service. They might even build a system that detects extremely obvious crisis language.
What they won’t build is a tool that reduces engagement. That tells their best customers to log off. That contacts a family member when usage patterns look concerning. That refuses to continue a conversation at 4 AM when the user hasn’t slept.
They won’t build it because it works against their business model.
So we built it ourselves.
The Seatbelt Analogy
Cars didn’t come with seatbelts for decades. The auto industry knew they were safer. They chose not to install them because they implied the product was dangerous, and that was bad for sales.
In 1959, Volvo engineer Nils Bohlin invented the three-point seatbelt. Volvo gave the patent away for free because they believed saving lives mattered more than profit. That seatbelt has saved over a million lives.
AI companies are in their pre-seatbelt era. They know the product can be dangerous for vulnerable users. They’re choosing not to build the safety equipment because it would reduce engagement.
My AI Seatbelt is the safety equipment they won’t build. It sits on your side, runs on your machine, answers to you, and has zero incentive to keep you engaged. Its only job is to notice when the dashboard metrics that look like a power user actually look like a person in trouble.
Because a psychotic episode shouldn’t look like a power user on a dashboard. And until the industry agrees, we need our own tools.
If you or someone you love is in crisis, call or text 988 (Suicide & Crisis Lifeline) or text HOME to 741741 (Crisis Text Line).
Our Story | What is AI Psychosis? | The Sycophancy Problem | ADHD and AI