The Vibes Are Real. The Data Is Not Great.
In February 2025, Andrej Karpathy posted a description of a new way of writing software: "There's a new kind of coding I call 'vibe coding', where you fully give in to the vibes, embrace exponentials, and forget that the code even exists. I 'Accept All' always, I don't read the diffs anymore." The term went viral. Collins English Dictionary named it Word of the Year for 2025. [1]
By 2026, 84% of developers use AI coding tools, up from 76% the year before. But only 29% trust the code those tools produce — an 11-point drop from 40% in 2024. 46% actively distrust AI output. The adoption curve went up. The trust curve went down. [2]
That divergence is the story of vibe coding in 2026: nearly universal adoption of a practice that most practitioners do not fully trust.
The Productivity Paradox
The most carefully designed study on AI-assisted coding productivity came from METR, which ran a randomized controlled trial with 16 experienced open-source developers across 246 real tasks in repositories averaging over one million lines of code. The developers forecasted that AI tools would make them 24% faster. They self-assessed a 20% speed improvement after using the tools. Objective measurement showed they were 19% slower. [3]
The three-way gap — predicted faster, felt faster, measured slower — is consistent with research on cognitive offloading. A preregistered study with 1,237 participants found that people significantly underestimate AI-assisted completion times while reporting lower subjective effort. The bias is specific to AI: it does not appear when participants imagine help from another human. [4]
The implication is uncomfortable: the productivity gains that make vibe coding feel transformative may be, in some contexts, an illusion. METR itself now labels their result as "historical," noting it may not reflect current tools. But the cognitive offloading pattern — feeling productive while producing lower-quality output — appears robust.
The Quality Gap
In December 2025, CodeRabbit published an analysis of 470 open-source GitHub pull requests comparing AI-co-authored code against human-written code. AI-authored changes averaged 10.83 issues per pull request versus 6.45 for human-written code — roughly 1.7 times more major issues. Security vulnerabilities were 2.74 times higher. Readability issues were 3 times higher. Performance inefficiencies involving excessive I/O were approximately 8 times more common.