Collins Dictionary made "vibe coding" its word of the year for 2025, defining it as "the use of artificial intelligence prompted by natural language to assist with the writing of computer code." In plain terms: you describe what you want, the AI writes it, and you mostly do not read what it wrote.
To a business owner that sounds like the whole problem solved. No developer, no quote, no waiting. Type a paragraph, get a website.
I use AI every day to build sites, so I am not going to tell you it does not work. I am going to tell you where it works, where it stops working, and how to tell which situation you are in before it costs you customers.
The speed is real, and it is not a small effect
GitHub ran a controlled experiment with 95 professional developers, split into two groups, both building the same thing. The group using AI finished 55 percent faster. One hour eleven minutes against two hours forty-one.
That is a real result. But look at the task: one new thing, built from scratch, with no existing site to fit into and nothing already there to break. That is exactly the situation AI is best at, and your first draft of a homepage is that situation. Which is why the first hour of vibe coding feels incredible.
It is a genuinely good way to find out what you want
Most people cannot describe the website they want. They can only react to one. Sitting with an AI tool for an evening and generating four versions of your homepage will teach you more about what you actually want than three meetings ever will.
If you turn up having done that, you are a better client and you get a better site. Bring the mess. That use has almost no downside.
The trouble is the last ten percent, and it is always the last ten percent
Stack Overflow surveyed close to 49,000 developers in 2025. 84 percent use or plan to use AI tools. Only 33 percent say they trust the accuracy of what it produces. 46 percent actively distrust it.
The single biggest frustration, named by 66 percent, was "AI solutions that are almost right, but not quite." 45 percent said debugging AI-generated code took more time than they expected, not less.
Almost right is the dangerous category. Broken is obvious, and obvious gets fixed. Almost right ships. It is the contact form that works on your laptop and quietly fails on an iPhone. It is the mobile menu that traps someone on one page. You never see it happen. You just never hear from that customer.
And remember who is reporting that. People who read code for a living, and it is still their number one complaint. If you are not reading the code, you do not have their ability to catch it either.
Security is the part you cannot check by looking
Veracode tested more than 100 AI models across four programming languages in 2025. 45 percent of the code samples introduced a vulnerability from the OWASP Top 10, the standard industry list of common web security flaws. In JavaScript, the language most websites run on, the failure rate was 43 percent. Against cross-site scripting specifically, the models failed to defend 86 percent of the time.
They also found it is not improving. Newer models got better at writing code that works. Security stayed flat.
A website that looks perfect can be wide open. There is no visual tell. If your site takes a name, an email, a phone number, or a payment, that is your customers' information sitting behind code nobody has read.
The help fades as the thing gets more complicated
METR ran a randomized trial in 2025 with 16 experienced developers working on their own large, mature projects. With AI allowed, they took 19 percent longer. They finished believing they had been 20 percent faster.
To be fair, METR say plainly that this is a snapshot of early-2025 tools in one setting, not proof that AI slows everybody down, and those developers were not building a five page website for a local business.
But the direction is what to notice. The more history a project has, the less the AI knows about it, and the more of the work becomes checking rather than creating. Your website builds up that history too, just slower. Month one is a delight. Month nine, when you want to add online booking to a site you no longer remember the shape of, is a different afternoon entirely.
Ask the ownership questions before you start, not after
Whatever tool you use, get straight answers to these four:
Where does the code actually live, and can you download it? Is the domain registered in your name? If you stop paying for the tool, does the site stay online? Could you hand this to a developer later, or would they have to start over?
The last one is the expensive one. Plenty of vibe coded sites cannot be handed over in any useful state, so the rebuild cost simply arrives later, after you have already paid once.
Where I land
Use AI to explore. Use it for anything with no customer data in it. A one page site for an offer you are still testing is a perfectly reasonable thing to vibe code, and I would rather you did that than pay me to find out the offer does not land.
Once real money or real customer information touches it, somebody should be reading the code.
That is the whole position, and it is why the line on my homepage is "AI-accelerated, not AI-cut-corners." I use the same tools. The difference is that everything gets reviewed by a person, you own the source code, the domain is in your name, and there is someone to call when it breaks at eleven at night.
The AI is not the shortcut. Skipping the review is the shortcut, and that is the bill that turns up later.
Sources
Collins Word of the Year 2025, Collins Dictionary, November 2025 Research: quantifying GitHub Copilot's impact on developer productivity and happiness, GitHub (controlled experiment, 95 developers) 2025 Stack Overflow Developer Survey, AI section (48,938 respondents) 2025 GenAI Code Security Report, Veracode, July 2025 Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity, METR, July 2025