Vibe coding: no AI overlords, no 10x engineers. Just code someone was willing to put their name to.

Vibe coding, describing what you want in plain English and letting AI write the app, will happily have a go at anything you ask. It'll look done. It might even run. The bit it won't tell you is whether it's any good, because it doesn't actually know either.

We've had this happen ourselves. Code that reads perfectly fine, does what you asked, and is quietly wrong in a way nobody spots until it matters. Not because the AI's rubbish. Because "looks right" and "is right" turn out to be very different things, and AI's brilliant at the first one.

Most developers now use AI to help write code, us included. Ask those same developers whether they trust what it produces without checking it, though, and the answer's been getting less confident over time, not more, the more people actually use it day to day.

One AI tool, marketed as basically a stand-in software engineer, was tested independently and got the job right about once in seven tries. Not "occasionally wrong." Wouldn't survive its probation.

There's a knock-on effect too. Researchers went through a huge pile of AI-written code and found the normal tidying up, cleaning rough edges, removing duplicated bits, had almost stopped happening. Not because AI's careless. Because it doesn't know it's supposed to go back and tidy, and increasingly nobody's checking that anyone did.

We're not about to tell you vibe coding is bad. We use it constantly, it's a genuinely brilliant first draft machine. First draft's the key phrase there.

What actually decides whether an app works when it matters isn't whether AI wrote the first version of it. It's whether someone who's built things before looked at what came out the other end and was willing to say "actually, no." That's the entire reason nobody junior gets handed your project after the sales call here, human or otherwise.

If you're weighing up how much of a build should lean on AI and how much needs a person who's done this before, worth talking through before anything gets written, not after. Get in touch.

Sources

• 84% developer AI-tool adoption, trust dropping from 43% to 29% over 18 months, code duplication up 48%, refactoring down 60%, 45% of AI-generated code containing vulnerabilities, and a 15% independent-testing success rate for an AI coding agent: Pixelmojo, which itself cites GitClear's analysis for the code-quality figures. One secondary source carrying most of the weight here, not several independent ones.