The 300% vibe-coding statistic, traced

The trail starts at a table. In a report published on 4 February 2025, GitClear classified 211 million changed lines of code by what each change did. Lines that were copied and pasted rose from 8.3% of changed lines in 2020 to 12.3% in 2024. Lines that were moved, which the report reads as refactoring, fell from 24.1% to 9.5%.2

Those two pairs give the post's second and third numbers. 8.3 to 12.3 is a rise of 48%, and 24.1 to 9.5 is a fall of 60%. Neither percentage is printed in the report. Both are arithmetic on its table.

The arithmetic changed two things on the way. The window is 2020 to 2024, about four years, not 18 months. And the subject is every changed line in GitClear's dataset, read against the years in which AI assistants spread. A search of the report's text for "AI-authored", "distinguish", "identify" and "proxy" found no line-by-line split between code written by AI and code written by people.

The report's own appendix gives the 2020 copy/paste share as 8.86%, not 8.3%. From that figure the rise is about 39%.2

The post credits Checkmarx, an application-security company. Searches of Checkmarx's site on 24 September 2026 did not find the claim. They covered its homepage, a report page, its blog index for August and September 2026, and four of its posts on AI-written code from November 2025 to July 2026. That is a finding about those searches, not proof that no Checkmarx document contains it.

No primary source for "300% in 18 months" turned up. The closest material is two blog timelines.

Codebridge published one on 3 February 2026. It says unmanaged AI-generated code "can drive maintenance costs to four times traditional levels" and describes "Months 16–18 (The Wall)." It links GitClear for its 48% and 60%, and gives no source for the four times.3

The other, published by Tian Pan on 17 April 2026, says "Months 12–18: Maintenance costs have grown to four times the level of a comparable traditional codebase," also without a source.4

A cost four times higher is 300% more. Whether the post's number came from timelines like these is an inference. The post links nothing that would settle it.

GitClear's report uses the words "long-term maintenance cost" once, in a recommendation: "Unless managers insist on finding metrics that approximate 'long-term maintenance cost'…"2 It doesn't measure one.

GitClear sells developer analytics, and the report states that all of its data "is available to GitClear customers with a basic subscription."2 The sample is 211 million changed lines authored between January 2020 and December 2024. About two-thirds comes from private companies that opted in to anonymised data sharing and one-third from open-source projects. Each changed line is classified as added, deleted, updated, moved, copy/pasted or find/replaced, and the shares are compared by year.2

The report's own findings are narrower than the post. "2024 marked the first year GitClear has ever measured where the number of 'Copy/Pasted' lines exceeded the count of 'Moved' lines." And: "we recorded an 8-fold increase in the frequency of code blocks with 5+ duplicated lines during 2024."2

A randomised controlled trial by METR had 16 experienced open-source developers complete 246 tasks in projects they knew well, with AI tools allowed or disallowed at random. The developers forecast that AI would cut completion time by 24%. The measured result: "allowing AI actually increases completion time by 19%." The authors note that "the influence of experimental artifacts cannot be entirely ruled out."5

A study of 806 GitHub repositories that adopted the Cursor editor, compared with 1,380 matched repositories that did not, found "a 281.3% increase in lines added in the first month." The abstract calls that increase "transient." Over the same period, "Static analysis warnings increase by 30%" and "code complexity increases by 41.6%." The sample covers only repositories that committed Cursor configuration files.6

A study of "302.6k verified AI-authored commits from 6,299 GitHub repositories" found that "more than 15% of commits from every AI coding assistant introduce at least one issue," and that "22.7% of tracked AI-introduced issues still survive at the latest version of the repository." Code smells made up 89.3% of the issues.7

Veracode, which sells application-security testing, tested "over 100 large language models across Java, Python, C#, and JavaScript" and reported that "45% of code samples failed security tests." It also reported: "Security performance remained flat, regardless of model size or training sophistication."8

None of these is a maintenance-cost multiple. Each one measures something narrower: time on task, warnings and complexity, issues that survive, failed security tests.

I don't have a computer-science degree. The pipeline behind this site and my game is built with AI agents. What it had to add to stay maintainable weren't features.

• Checks that can fail. Two automated checks once passed while inspecting nothing. One skipped all of its input because an escape sequence had collapsed into a raw control byte. The other matched no font declarations on two projects. Every check now has a sample it must fail on. The note A check that has never failed is not tested covers one of them.
• Line endings. One read, replace and write pass produced 5,599 changed lines for about 30 real changes across 24 files, because the writes rewrote every line ending. Source writes now go through one helper that keeps each file's own style, and a pre-commit check inspects the diff.
• Rollback. For the first 23 passes the repository wasn't under git, so every rollback the documentation promised was fiction. It has been under git since 25 July 2026.
• Scoped commits. One commit swept 28 files from another project into its own history. Commits are now scoped to one project's folder.

• Whether vibe coding is bad in general. The studies above measure specific outcomes in specific samples at specific dates.
• A maintenance-cost multiple for AI-written code. No primary measurement of one was found.
• Whether Checkmarx published the claim somewhere these searches didn't reach.
• This build's own maintenance cost. It hasn't been measured.