A preregistered ETH Zurich study of 100 students found that computer-science knowledge, not daily AI habit, predicts who succeeds at vibe coding. People who leaned on language models more often in daily life did worse.

If you’ve been telling yourself that AI finally lets anyone build software, ETH Zurich just handed you a reason to check that assumption. The details matter. The study ran 100 tertiary-level students through tests of computer-science achievement, general reasoning, written communication, and hands-on vibe coding. The paper, by Sverrir Thorgeirsson, Theo B. Weidmann and Zhendong Su, appears in the Proceedings of the 2026 CHI Conference on Human Factors in Computing Systems, held in Barcelona from April 13 to 17.

The result is blunt. Computer-science achievement correlated with vibe-coding performance at r=.39, while writing skill came in at r=.29. Knowledge beat habit. CS achievement stayed significant even after the researchers controlled for domain-general cognitive ability, and in the joint model it carried roughly twice the unique variance of writing skill.

The no-code pitch has a problem

Here’s the part that should sting for anyone selling the no-code dream. Students who reported more exposure to large language models in daily life performed worse on the vibe-coding task, not better. Familiarity with the tool didn’t translate into skill using it for this job.

That lands directly on the pitch you’ve heard from AI coding vendors, bootcamp marketers and no-code evangelists over the past two years: you no longer need to understand loops, state, data structures or failure modes to ship a product. Don’t buy the clean version of that story. The ETH study doesn’t say AI coding tools are useless. It says the person steering them still matters, and the useful steering looks a lot like old-fashioned computer-science knowledge.

The researchers also built the writing test around something concrete. Each participant wrote a 300-to-450-word explanation, in two or three paragraphs, of a technical concept from their own coursework. That’s not a creative-writing exercise. It’s a test of whether you can describe a system precisely enough that someone else, or in this case a model, can act on it correctly.

A vague prompt produces vague code. A prompt that names edge cases, data shapes and failure conditions gives the model a better target. Vibe coding didn’t remove the need for precision. It moved some of that precision from source code into English.

Founders still need the hard skill

That reframes what a lot of solo, non-technical founders have been doing with tools like Cursor, Replit and Lovable. If you’re building an MVP entirely through prompts without having studied how software actually breaks, this study points to the missing skill. You can get further than you could five years ago. That part is real. The ceiling, though, appears to be set by the same knowledge AI was supposed to replace.

There is a second data point worth keeping close. METR, the nonprofit research group, published a randomized controlled trial in July 2025 on 16 experienced open-source developers working across 246 real tasks in repositories they knew well. The developers expected AI tools to cut completion time by 24%. Instead, METR found they took 19% longer.

That’s a useful warning from the other end of the skill spectrum. Novices overestimate what the tool can do alone. Experienced developers make the same mistake in reverse: they overestimate the time AI saves once review, debugging and context enter the work. Confidence keeps outrunning the measured result.

None of this means vibe coding is a dead end. It means the labor didn’t disappear. It moved. For startups hiring or training people to use AI coding agents, the ETH Zurich findings argue against treating vibe coding as a shortcut around fundamentals. Hire for CS knowledge, not prompt count. Someone who understands data structures and failure modes will write a sharper prompt than someone who just opens ChatGPT every day.

The tool amplifies what you already know. It doesn’t hand you what you don’t.

Also read: TAC Is Rewriting Its Own Token Supply To Strand The Hacker Who Drained It • Zcash Tops $1,000 for the First Time as Grayscale’s Privacy Coin ETF Ignites a Rally • Tether Froze $42 Million on a Verbal Request Before Any Warrant Existed