AI 101 — What is vibe coding?
Vibe coding is writing software by describing what you want in plain language and letting an AI write the code — while you judge the result by whether it works, not by reading every line. Andrej Karpathy, co-founder of OpenAI, named it in a February 2025 post; by November, Collins Dictionary had named "vibe coding" its word of the year for 2025.
Why it matters right now. Code is the one thing AI does that people actually depend on, and the numbers show it: Stack Overflow's 2025 Developer Survey found 84% of developers using or plan to use AI tools, up from 76% a year earlier. The practice has outrun the vocabulary — this week alone, OpenAI's Codex lead pledged 28 days of daily fixes or a full reset of how the tool is built, and Slack launched a product built explicitly for vibe coding with AI agents (we covered the launch). Companies are shipping software that a person described and a model typed, which makes "who checked this?" a real question with real consequences for everyone who uses the result.
The mental model. You stop being the typist and become the director. You state the outcome — "add a login page, make it remember me" — the model writes the changes, you run the thing, and when something breaks you paste the error back and ask again. Karpathy's own description was to "fully give in to the vibes" and "forget that the code even exists." The loop is conversation plus trial and error: the code becomes an output of the chat, like a document, rather than something you constructed statement by statement. Under the hood this is the same machinery as any other AI assistant — often an AI agent that can open files, run commands, and edit its own work until the result compiles.
The analogy. Traditional coding is cooking dinner yourself: you read the recipe, you know what went into the pot, and if the smoke alarm goes off you know why. Vibe coding is ordering from a kitchen you never enter. You describe the dish, you taste it, and you send it back if it's wrong — but you can't tell whether the chef washed their hands. The meal can be excellent. The point is that your confidence comes from the result, not from knowing the process, and that gap is exactly where the risk lives.
Common misconceptions. First: "vibe coding means the AI builds a serious product unattended." It's a spectrum, from accepting autocomplete suggestions you half-read to handing an agent a task and reviewing nothing — most real work sits in between, and professional teams that use these tools daily still review what ships. Second: "no skill required." The typing gets cheaper; the specifying doesn't. Deciding what to build, judging when the output is right, and catching what's missing are still the developer's job — and trust in AI output is falling even as use rises, with Stack Overflow reporting trust at an all-time low in its 2025 results. Third: "if it runs, it's fine." Running and being safe are different tests. A model can confidently reference functions that don't exist — the same failure mode we cover in What is an AI hallucination? — and code nobody read is precisely where the bugs and security holes wait.
Where to learn more. Karpathy's original post remains the clearest statement of the idea; CNN's coverage of the Collins word-of-the-year announcement is the best short explainer for general readers; and Microsoft Research's "Good Vibrations" study looks at how developers actually experience the practice, flow and all.
Related reading: Slack launches Slack Code for vibe-coding with AI agents · What is an AI agent?
Have you shipped code you never actually read? Tell us in the comments.
