Karpathy warns of “AI psychosis” among developers — everyone else could be next

A terse jab that landed on Reddit

It has been reported that Andrej Karpathy, a prominent AI researcher, used the phrase “AI psychosis” to describe a growing tendency among developers to anthropomorphize and overtrust models. The comment surfaced in a Reddit thread and quickly drew attention — and ire. Short, sharp, and a little unnerving. Developers treating code like a conversation partner? That’s what Karpathy warned about.

What he meant — and why it matters

Allegedly, the worry is that the habit of talking to models as if they had intentions or understanding isn’t harmless play; it reshapes how people design, test and rely on systems. When engineers start assigning agency to a tool, design choices creep from rigorous engineering into a zone of intuition and ritual. That’s not just an academic worry. Automation bias, misinformation and fragile safety practices can follow. Think HAL, but with API keys and a lot more Stack Overflow chatter.

The bigger picture: guardrails, education and a little humility

Reddit responses ranged from agreement to eye-rolls, which is a nice microcosm of the larger debate: are we training humans to trust machines too fast? The obvious fix is not to ban curiosity — it's better education, clearer interfaces, and stronger guardrails so users understand what models actually do. Will public conversation catch up before the rest of us start whispering to our phones like they’re friends? Maybe. But Karpathy’s warning is a useful prod: this isn’t just about clever code. It’s about how we think.

Sources: reddit