What happened
Startup Fortune reports that a preregistered ETH Zurich study of 100 tertiary-level students found computer-science achievement was associated with better performance on a hands-on vibe-coding task. The reported correlation was r=.39, compared with r=.29 for writing skill. Students who reported more daily exposure to large language models performed worse on the task, although the study does not establish that AI use caused poorer performance.
Startup Fortune says the study was preregistered and involved 100 tertiary-level students completing tests of computer-science achievement, general reasoning, written communication and hands-on vibe coding. It identifies the authors as Sverrir Thorgeirsson, Theo B. Weidmann and Zhendong Su, and says the paper appeared in the Proceedings of the 2026 CHI Conference on Human Factors in Computing Systems in Barcelona from April 13 to 17.
According to Startup Fortune, computer-science achievement correlated with vibe-coding performance at r=.39, while writing skill correlated at r=.29. The report says computer-science achievement remained significant after controlling for general cognitive ability and contributed roughly twice as much unique variance as writing skill in a joint model. It also says participants with more everyday exposure to large language models performed worse on the task. These are reported study findings, not independently verified results in the supplied material.
The article says participants wrote a 300-to-450-word explanation of a technical concept from their coursework. Startup Fortune presents this as evidence that precise communication about systems, data shapes, edge cases and failure conditions may remain important when people direct coding models. It also cites a separate 2025 METR randomized controlled trial involving 16 experienced open-source developers and 246 tasks, in which developers took 19% longer despite expecting a 24% time reduction. That comparison is reported context, not a new result from the ETH Zurich study.
Source details: startupfortune.com ↗
Why it matters
The findings challenge the idea that familiarity with AI tools alone is a reliable substitute for software fundamentals. If the reported relationships hold in broader testing, people using coding agents may still need knowledge of data structures, system behavior, edge cases and failure modes to direct, evaluate and repair generated code. The practical implication is that organizations training staff or hiring for AI-assisted development may need to assess technical understanding rather than prompt frequency.
The report’s central distinction is between using an AI coding tool and understanding the software problem being delegated. A person who can identify relevant inputs, expected behavior and failure modes may be better positioned to write precise instructions and detect incorrect output. The reported association does not prove that computer-science knowledge causes stronger vibe-coding performance, but it does weaken the claim that habitual AI use by itself creates equivalent capability.
For founders, educators and employers, the practical question is whether AI-assisted development changes which skills matter. Based on the report, it may shift some work from typing code to specifying requirements, reviewing generated implementations and debugging failures. That could make foundational knowledge more important at key points in the workflow, even when a model produces much of the initial code. The source does not establish how widely the result generalizes beyond the 100 students tested.
What to watch next
The supplied report does not independently confirm the study’s methods, results or conference publication. Further scrutiny should examine the full paper, task design, participant selection, effect sizes and whether the result replicates with professional developers and nontechnical users. The report provides no coding-tool access or pricing information because the finding concerns user capability, not a new product.
The full study is the key missing source for evaluating the reported claims. Important unknowns include the exact vibe-coding task, the models and tools used, how performance was scored, the students’ prior programming experience, and whether daily AI exposure was measured consistently.
Replication would matter because the sample was made up of tertiary-level students rather than a demonstrated cross-section of professional developers, founders or nontechnical users. Future studies could test whether the same relationships appear across different tools, programming tasks and levels of domain expertise.
Startup Fortune’s cited METR result is relevant as a caution about productivity expectations, but it concerns experienced developers working in familiar repositories and should not be treated as direct confirmation of the ETH Zurich result. No access conditions or pricing are reported because neither study introduces a product.