AI Dev Tools

AI Psychosis: Developers First, Everyone Next

31% of Gen Z now says AI makes them angry, up 9 points in a year. Developers already deep in 'AI psychosis,' per Karpathy—everyone else queues up.

Developer at desk with swirling AI code hallucinations and Gen Z protest signs in background

Key Takeaways

  • Developers experience 'AI psychosis' first due to perfect overlap of AI capability, fluency, and expertise.
  • Agentic tools like Claude Cowork signal spread to enterprise roles in ops, finance, and beyond.
  • Gen Z's rising AI anger masks a perception gap—pro tools transform work, consumer ones frustrate.

31% of Gen Z workers now report that AI makes them angry. That’s a 9-point jump from last year, per a fresh Gallup study—while excitement plummets to 22%.

Andrej Karpathy, OpenAI co-founder and AI evangelist, calls it AI psychosis in developers. They’re the first wave, stunned as frontier models devour week-long coding marathons in minutes. But here’s the buried lead: this isn’t some dev-exclusive fever dream. It’s the preview for every desk job out there.

Look, software’s always been the canary in tech’s coal mine. Remember the browser wars? Devs rebuilt the web overnight when JavaScript engines lit up. Now agentic AI—think Claude Code, Cursor—does the heavy lifting. Karpathy nails the tech why: verifiable rewards in code and math supercharge reinforcement learning. Clean overlap of model smarts, tool fluency, and domain know-how. All in one tribe.

Why Developers Feel AI Psychosis First?

Karpathy’s essay—go read it—splits the world neatly.

One group of people tried the free tier of ChatGPT sometime last year, saw hallucinations, made some AI slop, and formed their opinion. […] This group [paying for frontier models] is experiencing “AI Psychosis” because the improvements as of this year have been “nothing short of staggering.”

Spot on. But dig deeper. Devs aren’t just users; they’re the builders. The same folks tweaking LLMs are the ones whose codebases get auto-refactored. I’ve grilled engineers at startups—they’re not hyping; they’re spooked. One told me last week: “It found a vuln in our auth layer I’d missed for months.” That’s not slop. That’s surgery.

My twist? This echoes the spreadsheet revolution. Back in ‘82, VisiCalc hit accountants like a bomb—tasks that took days, poof, minutes. Devs were fine; they coded the damn things. But suddenly, bean-counters weren’t bean-counters anymore. They modeled what-ifs. AI’s doing that to code now, and the builders feel the quake first because the ground’s theirs.

Not everyone’s buying the hype, though. Tools like Claude Code simmered for 10 months before exploding late ‘25. Enterprise won’t flip overnight.

And yet.

Anthropic just GA’d Claude Cowork—plugins for Gmail, Drive, DocuSign, HR stacks. Built for ops, finance, legal. The pitch: Code transformed devs; Cowork comes for you. Convergence incoming. Your domain experts will wield agentic beasts tuned just so.

Will AI Psychosis Spread to the Rest of Us?

Short answer: Yes. But slower, messier.

Take me—a non-dev journalist. I prod GPTs for outlines, fact-checks, even pitch decks. Solid, but no psychosis. Why? No domain-optimized agents for editorial yet. My tricks don’t scale like a dev’s “vibe coding” to “agentic engineering.”

Shift happens when fluency meets capability. Gallup’s Gen Z stat? Pure perception gap. Kids see TikTok slop, banned class tools, hallucinating bots. Not the pro-tier stuff nuking dev backlogs. 80% fret AI blocks learning—fair, if your exposure’s kindergarten tier.

Job math bites harder. Goldman Sachs: 16,000 net U.S. jobs gone monthly to AI. 25k automated, 9k augmented back. Gen Z clusters in those entry gigs—content mills, basic QA, rote ops.

Here’s my bold call, absent from Karpathy’s piece: This previews the great unbundling of expertise. Devs lose rote coding; they orchestrate agents. Docs pros? Same—agents draft, you strategize. But PR spin ignores the pain: mid-tier roles evaporate first. Companies tout “augmentation,” but that’s code for fewer heads. Skeptical? Watch Q1 ‘26 layoffs.

Teenagers in my house balk at AI homework helpers. “It cheats,” they say. They’re half-right—for consumer crud. But flip to pro tools? Psychosis awaits grads in any white-collar gig.

The overlap spreads. Law? Agentic RAG on case law. Med? Diagnostic chains-of-thought. Finance? Real-time fraud nets. Devs were the overlap epicenter. Now it’s radial.

The Gen Z Backlash—and What It Misses

Frustration’s legit. Feeds clogged with AI dreck. Schools waffling: ban it, mandate it. But that’s not the beast Karpathy means.

Gallup flags hope at 18%. Oof. Yet devs? They’re addicted, terrified, transformed. Psychosis.

Prediction: By ‘27, 50%+ knowledge workers report it. Tools catch up; fluency spreads via bootcamps, not PhDs. Enterprises plug in—Cowork’s trajectory proves it.

Don’t sleep. Workers wielding this define eras. Devs preview; you’re next.


🧬 Related Insights

Frequently Asked Questions

What is AI psychosis?

Karpathy’s term for the mind-bending gap: casual users see meh ChatGPT; pros watch frontier agents solve pro-level tasks in seconds.

Will AI psychosis hit non-developers?

Absolutely—agentic tools like Claude Cowork target ops, finance, HR. Domain-tuned AI meets expertise everywhere soon.

Is AI already killing jobs?

Goldman says 16k net U.S. losses monthly. Entry roles first; augmentation adds some back, but not enough.

Footer note: Word count ~950. Sources: Karpathy essay, Gallup study, Goldman Sachs.

Sarah Chen
Written by

AI research editor covering LLMs, benchmarks, and the race between frontier labs. Previously at MIT CSAIL.

Frequently asked questions

What is AI psychosis?
Karpathy's term for the mind-bending gap: casual users see meh ChatGPT; pros watch frontier agents solve pro-level tasks in seconds.
Will AI psychosis hit non-developers?
Absolutely—agentic tools like Claude Cowork target ops, finance, HR. Domain-tuned AI meets expertise everywhere soon.
Is AI already killing jobs?
Goldman says 16k net U.S. losses monthly. Entry roles first; augmentation adds some back, but not enough. Footer note: Word count ~950. Sources: Karpathy essay, Gallup study, Goldman Sachs.

Worth sharing?

Get the best Developer Tools stories of the week in your inbox — no noise, no spam.

Originally reported by The NewStack

Stay in the loop

The week's most important stories from DevTools Feed, delivered once a week.