# When AI arrived, you could ask for a website in a sentence

> In the early 2020s, assistants like GitHub Copilot and ChatGPT began writing code and pages from plain words, and the owner still had to check the work

By Orla Finch. Published 2026-10-27.

History of building websites, 8 of 10.

Say it is 2023, and Rosa, who turned to a site builder so she would never have to touch code, hears that she can now ask for a web page in a plain sentence and get one back. [Last time](/blog/frameworks-and-static-sites/) we followed Jess through years of heavy machinery around simple pages. This time the machinery talks back.

The tools behind it are [large language models](https://en.wikipedia.org/wiki/Large_language_model): programs trained on a huge amount of text, which can write it, sum it up and translate it. They are one kind of [generative AI](https://en.wikipedia.org/wiki/Generative_AI), software that makes new text, pictures, sound or code from patterns it learned, usually in answer to a request typed in everyday words. Tools like these spread fast in the AI boom of the early 2020s.

Programmers met them inside their own tools. In 2021, GitHub [unveiled](https://github.blog/news-insights/product-news/introducing-github-copilot-ai-pair-programmer/) [GitHub Copilot](https://en.wikipedia.org/wiki/GitHub_Copilot), which sits in the programs coders write in and [suggests what comes next as they type](https://docs.github.com/en/copilot/about-github-copilot/what-is-github-copilot). Describe a problem in plain words, and it can offer code to solve it, or explain in English what a piece of code does.

Then the public met them. OpenAI released [ChatGPT](https://en.wikipedia.org/wiki/ChatGPT) in November 2022, a chatbot that writes back to whatever you type, and it had 100 million users within two months. It can also give [answers that sound right and are wrong](https://web.archive.org/web/2023/https://openai.com/blog/chatgpt). People call those hallucinations.

By February 2025, building this way had a name. Andrej Karpathy, a co-founder of OpenAI, called it [vibe coding](https://en.wikipedia.org/wiki/Vibe_coding): you describe what you want, the model writes the code, and you steer by what comes out, sometimes without reading the code closely at all. Its fans say it lets people make working software without years of training. Karpathy had quick projects in mind, the kind you throw away after a weekend; some experts see more risk in work that has to last.

For Rosa, that is a real change. Her first site took someone she paid, and her second took a builder and a lot of clicking and dragging. Now the distance between wanting a page and having one can be a sentence.

What she still has to mind is the part the sentence leaves out. Critics of vibe coding point to code its makers do not understand, which is hard to keep up and can hide security holes. When a New York Times reporter tried it, the code he got made up fake reviews for a shop's site.

So for Rosa, and for you: the assistant can write the page, but the page is still yours. Read every word of it before it goes up, and ask what it collects from your customers.
