Did you know a way of programming became word of the year?
In February 2025, Andrej Karpathy, former director of AI at Tesla and one of OpenAI's founding engineers, described a new way of programming in a post: go with the flow, ask the AI for everything and, in his words, "forget that the code even exists." He called it vibe coding.
The term spread so fast that, in November 2025, the Collins dictionary named it word of the year. And it is still part of the conversation: on October 16, 2026, Al Jazeera devoted a segment to whether it is a game-changer or risky business.
What it is (and what it isn't)
Vibe coding relies on language models, the technology behind chatbots: systems that generate text, and code too, from a request written in everyday language. The cycle is simple: you describe what you want ("a page to log my business's orders"), the AI writes the program, you try it and, if something fails, you tell it what happened so it can fix it.
The key lies in what you don't do: read the code. Karpathy said he accepted every change without reviewing it, and noted that the method was fine for "throwaway weekend projects."
That is why programmer Simon Willison proposed a useful distinction: vibe coding is building software with AI without reviewing the code it writes. Using AI to program while understanding every line before using it is something else. His personal rule: don't commit code he couldn't explain to someone else.
A dream from the 1950s
Giving orders to a computer in our own language is not a new idea. In the 1950s, Grace Hopper and her team created FLOW-MATIC, which according to Yale University was the first programming language to use English-like commands. Hopper believed that the people who processed data in companies, who were usually neither mathematicians nor engineers, would be more comfortable with words than with symbols. Her goal, she said in 1980, was "to bring another whole group of people able to use the computer easily."
Much of COBOL came out of that experience: according to the Computer History Museum, many of its specifications were drawn from FLOW-MATIC. By the 1970s, according to Yale, it was the most widely used programming language in the world.
Not everyone shared the enthusiasm. In the late 1970s, computer scientist Edsger Dijkstra wrote "On the foolishness of 'natural language programming'". His argument: formal symbols are not an obstacle but an aid, because they force us to be precise and rule out ambiguities that everyday language lets slip through. Today the machine responds to requests in our language, but the underlying question still stands: who checks that the program does exactly what we wanted?
What the data says
Among professionals, AI is already part of everyday work; vibe coding, not so much. In the 2025 Stack Overflow survey, more than 26,000 developers answered whether vibe coding was part of their work: 72% said no, and another 5% said no emphatically. In the 2026 edition, published this month, 73% of those who use coding assistants or agents use them daily, but only 20% turn to AI to deploy or operate production systems, and 48% trust it when they can easily verify its answers.
What about speed? In 2025, the research organization METR ran an experiment with 16 experienced developers from large open-source projects. With AI tools they took 19% longer to complete their tasks, yet afterwards they believed they had worked 20% faster. The authors warn that the results cannot be generalized, and METR now marks them as out of date. Even so, there is a lesson: feeling fast is not the same as being fast.
A team that includes Kelly Blincoe, of the University of Auckland, reviewed 101 sources with 518 firsthand accounts and described a "speed-quality trade-off paradox": vibe coders quickly get to something that works, but most perceive the result as fast but flawed, and testing is often skipped. The work has not yet been peer reviewed.
The most serious cost shows up in security. At Georgia Tech, the Vibe Security Radar tracks published vulnerabilities that originated in code written by AI tools: in March 2026 alone it found 35, more than in all of 2025. And it is a partial count, because it only detects cases where the tool left its signature.
When it makes sense and when it doesn't
Vibe coding is neither good nor bad: it depends on what is at stake. Based on Willison's criteria, we propose a traffic light:
- Green: prototypes to test an idea, personal tools, games or learning how a program works. If something fails, nothing serious happens.
- Yellow: internal tools your team will use. Before sharing them, have someone experienced review the code.
- Red: anything that handles personal data, payments, passwords or access keys to other services, or that uses pay-per-use services. Pure vibe coding is not enough here: it needs professional review.
If you are going to try it, these questions help:
- What happens if this fails or if someone sees this data?
- Did I set a spending limit on pay-per-use services?
- Did I test it with odd cases, not just the ideal example?
- Could someone explain to me what each part does before I share it?
For a business, the takeaway is concrete: vibe coding is useful for exploring and validating ideas quickly. When customers, data or money are involved, moving to production still requires someone who understands the code.