The most-shared vibe-coding study this year had the wrong country in the caption, and nobody checked before it hit 700,000 views.

A study on vibe coding went viral this year with the country wrong in the caption. Almost 700,000 people saw it before anyone checked.

The post read: "China published the most uncomfortable paper on vibe coding." It described a controlled study of 100 developers, tracking computer science background, writing skill and general reasoning, to see who actually succeeds when they build software by describing it to an AI instead of writing it.

The study is real. China had nothing to do with it. It's from ETH Zurich, in Switzerland — published on the university's own newsroom on April 30, 2026, authored by Sverrir Thorgeirsson, Theo Weidmann and professor Zhendong Su, and accepted to CHI 2026, the top conference in human-computer interaction, held in Barcelona that April (DOI: 10.1145/3772318.3791666). None of that is hidden. It just wasn't in the post, and 700,000 people didn't go look.

What the study actually found

A hundred students who'd taken at least one intro CS course worked through tasks in a real vibe-coding setup. Two things predicted whether they shipped something that worked. First, and by far the strongest: how much computer science they already knew — the effect held even after controlling for general cognitive ability, so it isn't just "smart people do better at everything." CS background contributed roughly twice the predictive power of writing skill on its own. Second: general writing skill also correlated with success, because describing what you want clearly turns out to be its own technical skill — "writing the prompts becomes a form of coding in itself." And the finding nobody predicted: people who used LLMs the most, daily and heavily, did worse at both essay writing and the vibe-coding tasks than people who used them less. The researchers don't have a confirmed explanation for that one yet.

It isn't really about prompting

Read past the headline and the study is about debugging, not prompting. The people who succeeded weren't the ones who wrote cleverer instructions — they were the ones who could look at what came back and notice the one place it quietly did the wrong thing. That's not a skill you can prompt your way into. It either shows up when something breaks, or it doesn't, and if it doesn't, you don't find out until later.

The misattributed post is the same failure, one level up

Catching "this says China, but the source is a Swiss university" takes the same move as catching a broken edge case in generated code — stopping to check the thing underneath instead of trusting the confident summary in front of you. Most people don't, for the same reason most non-technical vibe coders can't tell when an AI quietly got their app wrong: verifying takes a skill the summary was specifically designed to let you skip.

Which is the actual lesson, for anyone building or using AI-generated software: the gap between "works" and "silently broken" comes down to whether someone involved can see what's underneath. A smarter model or a better prompt doesn't close that gap. Keeping the thing small enough — one purpose, one screen, nothing accumulating underneath it — that you don't need a CS background just to tell whether it did what you asked, does.

Sources: ETH Zurich newsroom, "What skills do people need to successfully program with AI?", published 30 April 2026. Thorgeirsson, Weidmann & Su, "Computer Science Achievement and Writing Skills Predict Vibe Coding Proficiency," CHI 2026, DOI 10.1145/3772318.3791666.