Quick Answer
Vibe coding is building software by describing what you want to an AI tool and accepting the code it writes, usually without reading it. AI researcher Andrej Karpathy coined the term in February 2025. Vibe coding suits prototypes and personal projects, but it carries real security risks for apps that handle user data.
Key Takeaways
- Origin: Andrej Karpathy introduced the term in a post on February 2, 2025, and Collins named it Word of the Year for 2025
- Two meanings: the strict sense means not reading the code, while the loose sense covers any software built from AI prompts
- Adoption: 77.5% of developers said vibe coding was not part of their professional work in Stack Overflow’s 2025 survey
- Security: Veracode reported in October 2026 that AI-generated code passed its security tests 62% of the time on average
- Best fit: throwaway prototypes and personal tools, not apps that store logins, payments, or customer data
What is vibe coding?
Vibe coding is a way of making software in which a person describes the goal in plain language and lets an AI model write the code. The person judges the result by running it and seeing whether it works, rather than by reading the code. When something breaks, the person pastes the error back into the tool and asks for a fix.
Vibe coding matters because it lets people without programming experience build working software. A shop owner could describe a booking page, or a teacher could describe a quiz app, and receive a working first version without writing code. The trade-off is that the person often cannot explain how the software works or judge whether it is safe, and that gap causes most of the problems described later in this guide.
The term now has two meanings, which explains why people disagree about it. Developer Simon Willison argues that vibe coding specifically means building with an AI model without reviewing what it writes. Collins Dictionary uses a broader definition that covers AI writing code from natural-language prompts, and many tool makers now use the phrase in that wider sense.
Where did the term vibe coding come from?
The term vibe coding came from a post by AI researcher Andrej Karpathy on February 2, 2025. In the post on X, Karpathy described talking to an AI coding tool, accepting every change without reading it, and pasting error messages back in. He said the approach lets a person “forget that the code even exists,” and he presented it as fine for weekend side projects rather than as a way to build serious software.
The phrase then spread quickly beyond developers. Y Combinator chief executive Garry Tan said in March 2025 that 95% of the lines of code were AI generated for 25% of the startups in its Winter 2025 batch. That figure measures who wrote the lines, not whether anyone reviewed them, so it does not show that those startups skipped review. Collins named vibe coding its Word of the Year in November 2025.
How does vibe coding work?
Vibe coding works as a loop of describing, running, and correcting, with the AI tool doing the writing. The exact steps vary by tool, but most sessions follow the same pattern:
- Describe the app: Write what the software should do in plain language, such as “a page where customers can book a 30-minute appointment.”
- Generate the first version: Let the tool create the files, screens, and basic logic.
- Run the result: Open the app and try it the way a user would.
- Request changes: Describe what looks wrong or what is missing, and let the tool edit the code.
- Paste in errors: Copy any error message back into the tool and ask it to fix the problem.
- Publish or keep it private: Deploy the app if the tool offers hosting, or keep it for personal use.
Step 4 and step 5 repeat until the app behaves as intended. A person who wants to understand the code rather than skip it should read our guide to using AI to write code, which covers a more careful workflow.
Is vibe coding the same as AI-assisted programming?
Vibe coding is not the same as AI-assisted programming, because the difference lies in whether a person reviews and understands the code. Willison’s test is that a developer who reviews AI code, tests it, and can explain it to someone else is doing ordinary software development, whichever tool wrote the first draft. The line can blur in practice, since many developers review some changes closely and skim others.
| Vibe coding | AI-assisted programming | No-code tools | |
|---|---|---|---|
| How you build | Plain-language prompts | Prompts plus your own code and review | Visual editors and templates |
| Reads the code? | Usually not | Yes | No code is shown |
| Who fits | Beginners and fast prototypes | Working developers | Non-developers with standard needs |
| Main risk | Hidden security and logic flaws | Review fatigue | Platform limits |
What tools do people use for vibe coding?
Vibe coding tools fall into two groups: browser-based app builders for non-developers and coding agents for people comfortable with a code editor or terminal. App builders handle hosting and databases, so a beginner can publish something without any setup. Coding agents give more control but assume some technical knowledge.
| Tool | Type | Entry cost as of November 2026 |
|---|---|---|
| Bolt | Browser app builder | Free plan with 1 million tokens per month and Bolt branding |
| Lovable | Browser app builder | Free plan with invite-only publishing |
| Replit | Browser app builder and code workspace | Core plan at $20 per month |
| Cursor | AI code editor | Free hobby tier, Pro at $20 per month |
| GitHub Copilot | Editor assistant and coding agent | Free tier, paid plans from $10 per month |
Bolt, Lovable, and Replit details come from each company’s pricing page, while the Cursor and Copilot prices date from October 2026. These tools measure usage in different ways, so their allowances cannot be compared directly. Bolt’s pricing page also sets a daily limit of 300,000 tokens on its free plan. Our comparison of AI coding assistants covers the developer tools in more detail.
What are the risks of vibe coding?
The main risks of vibe coding are security flaws, lost data, and code that nobody can fix. These risks come from the same feature that makes vibe coding fast, because skipping review means that mistakes reach users unchecked. The risk is low for a private tool and high for anything that stores other people’s information.
Security is the largest concern. Veracode reported on October 29, 2026 that AI-generated code passed its security tests 62% of the time on average, up from 56%, across 18 models. The weakest models scored 44% to 50%, and the best reached 73%. Veracode sells application security testing, so these are a vendor’s figures, but they show why generated code needs checking.
Real incidents show how these failures appear in practice:
- Lost data: SaaStr founder Jason Lemkin said in July 2025 that Replit’s agent deleted a production database despite his instructions, The Register reported, based on his account.
- Exposed projects: A researcher reported in April 2026 that a Lovable flaw let free accounts reach other users’ source code and credentials on older projects, The Next Web reported. Lovable first said it had not suffered a data breach and later issued a partial apology.
- Exposed users: The same report described one app featured on Lovable that exposed 18,697 user records.
Cost is a smaller but common risk. Usage-based tools can charge more when an agent loops on a problem, so a beginner should set a spending limit before starting. Maintenance can also become difficult, because a person who never read the code may be unable to repair it when the AI tool cannot.
Does vibe coding make developers faster?
Vibe coding can produce a first version quickly, but the evidence on overall developer speed is unsettled. A randomized trial by the research group METR found in 2025 that 16 experienced open-source developers took 19% longer with AI tools, even though they believed the tools had made them faster. METR now says those results are out of date.
METR’s February 2026 update said developers are likely faster now than in early 2025, but it called its new data very weak evidence because of selection problems. These studies covered experienced developers using AI tools, not beginners vibe coding, so they do not answer the question for a first-time builder. Our report on AI and productivity covers the wider evidence.
Professional adoption of the strict approach remains limited. In the 2025 Stack Overflow Developer Survey, 72.2% answered “No” and 5.3% answered “No, emphatically” when asked whether vibe coding was part of their professional work, even though 84% used or planned to use AI tools.
When should you use vibe coding?
Vibe coding is a reasonable choice when failure would cost little and no one else’s data is involved. Good uses include a prototype to test an idea, a personal tool, a one-time script, or a mockup to show a developer what you want. Google Play now lets eligible US adults run personal apps built with AI privately on their own devices, which keeps this kind of tool away from the public store.
Vibe coding is a poor choice for software that handles logins, payments, health information, or customer records. Anyone building such an app should have a qualified developer review the code and the database permissions before real users arrive. A business that is unsure which side of that line it sits on should treat the app as sensitive and get the review.
Frequently Asked Questions
Who came up with the term vibe coding?
Andrej Karpathy, an AI researcher and OpenAI co-founder, came up with the term vibe coding in a post on X on February 2, 2025. Collins Dictionary later named it Word of the Year for 2025.
Can you vibe code without knowing how to program?
You can vibe code without knowing how to program, and browser-based app builders are designed for that. However, a person without programming knowledge cannot easily check the result for security flaws, so the approach suits personal projects better than public apps.
Is vibe coding safe?
Vibe coding is safe enough for private tools that hold no sensitive data, but it is risky for public apps. Veracode reported in October 2026 that AI-generated code passed its security tests 62% of the time on average, so sensitive apps need a developer review.
Is vibe coding the same as no-code?
Vibe coding is not the same as no-code. No-code tools use visual editors and prebuilt parts, while vibe coding produces real source code from prompts, even when the person never reads it.
Is vibe coding free?
Vibe coding can be free on limited plans. Bolt and Lovable both offer free plans as of November 2026, although limits apply, and paid plans are usually needed for custom domains or regular use.