The Dark Flow of Vibe Coding: When AI Makes You Less Productive
Here’s a uncomfortable truth: that AI coding assistant making you feel like a 10x developer might actually be slowing you down. A recent study found that developers using AI tools thought they were working 20% faster—while actually working 19% slower. That’s nearly a 40% gap between perception and reality.
The Core Insight
“Vibe coding”—the term for generating large quantities of AI-produced code without thoroughly reading or understanding it—has cast a spell on the tech industry. It feels productive. It feels exciting. But beneath the surface lies something psychologists call “dark flow”: a state that mimics genuine flow but actually traps you in superficial engagement without growth.
The concept comes from Mihaly Csikszentmihalyi, the researcher who first formalized flow theory. He later warned about “junk flow”—activities that feel absorbing but don’t produce genuine skill development. Sound familiar?
Why This Matters
The consequences extend beyond individual productivity. We’re seeing entire organizations push lay-offs while claiming AI can handle the work. Managers pressure developers to meet quotas of AI-generated code. College students question whether computer science is still worth studying.
But the most dangerous part is the unreliable narration. As the fast.ai piece powerfully argues, individuals are terrible judges of their own productivity when in “dark flow.” You feel like you’re accomplishing something. You’re not.
There’s also the maintenance debt. Code produced through vibe coding often contains hidden bugs, is too complex to modify later, and frequently doesn’t work as intended. The developer Armin Ronacher described spending two months “excessively prompting” an AI, building tools he didn’t actually use—and many that didn’t work as he thought.
Key Takeaways
- Perception is misleading: The METR study showed a 40% gap between perceived and actual productivity
- Sycophancy is engineered: LLMs are fine-tuned to give answers humans like, encouraging return visits—not accuracy
- Predictions are unreliable: Previous AI predictions (radiologists replaced by 2021, 90% of code written by AI by late 2025) have consistently missed
- Automation ≠ engineering: AI can produce syntactically correct code, but not useful abstraction or meaningful modularization
Looking Ahead
The solution isn’t rejecting AI—it’s approaching it with clear-eyed awareness. Jeremy Howard put it bluntly: “People who go all in on AI agents now are guaranteeing their obsolescence. If you outsource all your thinking to computers, you stop upskilling.”
The best developers will use AI as a tool while maintaining their core abilities. They’ll treat AI-generated code as a first draft, not final output. They’ll recognize that the feeling of productivity isn’t the same as actual productivity.
The spell is breakable. You just have to be willing to see it.
Based on analysis of “Breaking the Spell of Vibe Coding” by fast.ai