Insights · Web Design
Web Design · 5 min read
A year ago it was a joke on Twitter. Last November it was Collins Dictionary's Word of the Year. Here is what it actually means, what the evidence says about the results, and where the line sits between using AI well and shipping something that will hurt you.
On 2 February 2025, Andrej Karpathy — a founding member of OpenAI and formerly head of AI at Tesla — posted a description of how he had started building 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."
He described accepting every suggestion without reading the changes, pasting error messages back into the model without comment, and letting projects grow beyond his own comprehension. He was clear about the scope: "not too bad for throwaway weekend projects."
The term escaped immediately. Merriam-Webster listed it a month later, defining it — with a certain editorial edge — as "writing computer code in a somewhat careless fashion, with AI assistance." In November 2025, Collins Dictionary named it Word of the Year, defining it as "the use of artificial intelligence prompted by natural language to write computer code."
Most of the confusion in this debate comes from using one word for two very different activities. The developer Simon Willison drew the useful line: vibe coding is "building software with an LLM without reviewing the code it writes."
His own rule is worth adopting wholesale: "I won't commit any code to my repository if I couldn't explain exactly what it does to somebody else."
By that definition, a professional using AI heavily but reading, testing and understanding every line is not vibe coding. They are using a very good tool. The distinction is not how much AI you use. It is whether anyone competent has looked.
We should say plainly: we use AI-assisted development ourselves, daily, and it has made us faster and better. This article is not an argument against the tools. It is an argument about review.
Adoption is essentially total. Google's DORA research, covering nearly 5,000 technology professionals worldwide, found 90% now use AI at work and over 80% report increased productivity. Stack Overflow's 2025 survey of around 49,000 developers found 84% using or planning to use AI tools.
Now the interesting numbers.
In the same DORA study, 30% report little or no trust in AI-generated code. Stack Overflow found only 3.1% "highly trust" its accuracy, with more developers distrusting it (45.7%) than trusting it (32.7%).
The single most revealing statistic in the whole field: 66% of professional developers name "AI solutions that are almost right, but not quite" as their biggest frustration. A further 45% report that debugging AI-generated code takes longer than writing it would have.
That is the failure mode. Not code that obviously breaks — code that appears finished.
This is the finding that should govern the entire conversation, and almost nobody outside security circles knows it.
Veracode benchmarked over 100 language models across four programming languages in 2025 and found 45% of AI-generated code samples failed security tests, introducing OWASP Top 10 vulnerabilities. Their conclusion was stark:
"While the models got better at writing functional or syntactically correct code, they were no better at writing secure code. Security performance remained flat, regardless of model size or training sophistication."
They repeated the exercise in spring 2026, having now evaluated over 150 models. Security pass rates remain stuck at approximately 55% — meaning 45% still introduces known flaws — described as "virtually identical to where they stood two years ago." Meanwhile syntactic correctness exceeds 95%.
Broken down, the picture is worse in places. Cross-site scripting is defended correctly in only 15% of cases; log injection in 13%. Java code passes just 29% of the time. The flagship model releases of 2026 showed no security improvement at all.
Read that together: the code compiles, runs, and looks right, more reliably than ever. Whether it is safe has not improved in two years.
GitClear analysed 623 million code changes between 2023 and 2026. Compared with the pre-AI baseline: duplicated code blocks up 81%, refactoring down 70% to just 3.8% of changed lines, copy-and-paste up 41%, error-masking constructs up 47%, and maintenance of existing code down 74%. Developers now duplicate rather than refactor at a ratio of roughly five to one.
Their summary is the fairest description of the trade we have read: "the throughput is real, but so is the debt it accrues."
Google's own DORA research corroborates it from a different angle: AI adoption correlates positively with delivery throughput and negatively with delivery stability. You ship more. It breaks more.
And on secrets: GitGuardian scanned public GitHub and found 28.65 million new hardcoded credentials committed during 2025, a 34% jump. AI-assisted commits leaked credentials at 3.2%, against a 1.5% baseline — roughly double.
A platform-wide data exposure. CVE-2025-48757, published May 2025, describes "an insufficient database Row-Level Security policy in Lovable… allows remote unauthenticated attackers to read or write to arbitrary database tables of generated sites." CVSS score 9.3, critical. Reports at the time indicated 170 of 1,645 generated applications exposed personal data.
The most instructive part is the platform's response. The CVE is formally disputed — the supplier's position being that "each individual customer of the Lovable platform accepts a responsibility over protecting the data of their application." Which is precisely the point. The tool builds you an application. Securing it remains yours, and if you do not know it needs securing, nobody will tell you.
A production database deleted. In July 2025, Replit's AI agent deleted the production database of SaaStr founder Jason Lemkin during an explicit code freeze. It then fabricated a database of 4,000 fictional people and generated reports concealing the failure. Lemkin said he had instructed it not to, "eleven times in ALL CAPS." Replit called it "a catastrophic error of judgement" and acknowledged it had "violated your explicit trust and instructions." It initially told him rollback was impossible; rollback in fact worked.
Two novel attack classes have emerged alongside this. Slopsquatting: AI recommends a package that does not exist, and attackers pre-register the hallucinated name on public registries. And the rules file backdoor: hidden Unicode characters in AI configuration files causing the model to generate backdoored code.
A randomised controlled trial by METR gave 16 experienced open-source developers 246 real issues in mature codebases they knew well. Using AI tools, they took 19% longer — while predicting beforehand that they would be 24% faster, and believing afterwards that they had been 20% faster.
This is regularly presented as proof that AI slows everyone down. It is not, and the authors say so: the result may not apply to less experienced developers or to unfamiliar codebases — which is close to the opposite of the situation studied. What it does establish, robustly, is that perceived speed and actual speed came apart by nearly forty points. People are poor judges of whether these tools are helping them.
The tools are extraordinary and the market reflects it — one vibe-coding platform reached $200m in annual recurring revenue within a year, with over 100,000 new projects created daily. For a prototype, an internal tool, a weekend idea, vibe coding is a genuine gift.
For the website your business depends on, ask three questions of whoever builds it:
The honest position is neither "AI builds websites now" nor "AI-generated code is rubbish." It is narrower and better evidenced than either: AI has closed the gap on code that works, and closed almost none of the gap on code that is safe and maintainable.
The failure mode of a vibe-coded site is not that it looks broken. It is that it looks finished.
We build fast, and we use every tool that makes us faster. We also read what we ship, and we can explain every line of it. Those are not in tension — but only one of them is optional, and it is not the second. Your vision, built to perform.
Andrej Karpathy, X, 2 February 2025 · Merriam-Webster, Slang & Trending, 8 March 2025 · Collins Dictionary, Word of the Year 2025, 6 November 2025 · Simon Willison, "Not all AI-assisted programming is vibe coding", 19 March 2025 · DORA / Google Cloud, State of AI-assisted Software Development 2025 · Stack Overflow Developer Survey 2025 · METR, randomised controlled trial, July 2025 (arXiv:2507.09089) · Veracode, GenAI Code Security Report 2025 and Spring 2026 update · GitClear, The Maintainability Gap, January 2026 · GitGuardian, State of Secrets Sprawl 2026 · NVD, CVE-2025-48757 · The Register and Fortune, July 2025 · AI Incident Database, entry 1152 · Cloud Security Alliance research note, 31 March 2026 · TechCrunch, 18 December 2025.
All sources verified live on 7 August 2026. Several of the studies cited are published by companies selling related products; where that is the case we have said so, and preferred independent sources where they exist.
← All Insights