You Wouldn't Vibe Code a Car
In 2004, the Motion Picture Association of America placed an anti-piracy ad on purchased DVDs. You wouldn't steal a car. You wouldn't steal a handbag. You wouldn't steal a television. Downloading pirated films is stealing.
The ad played exclusively to people sitting in theaters who had already paid for their tickets. Pirates had pristine copies with no lecture. The campaign's most enduring cultural contribution was a format: "You Wouldn't [X]" — the perfect structure for escalation comedy, adopted by the internet and mutated into immortality.
Twenty-two years later, the format has a third generation. "You wouldn't vibe code a car." But the interesting question isn't about danger. It's about what "downloading" means when the alternative is making your own.
The Phrase
In February 2025, Andrej Karpathy — co-founder of OpenAI, former head of AI at Tesla — coined a term. "There's a new kind of coding I call 'vibe coding,'" he wrote, "where you fully give in to the vibes, embrace exponentials, and forget that the code even exists."
The description was casual. "I just see things, say things, run things, and copy paste things, and it mostly works." Collins Dictionary named it Word of the Year. By December 2025, 41% of all code written globally was AI-generated. 84% of developers were using or planning to use AI coding tools. By early 2026, 92% of US developers were using them daily.
The tools market that didn't exist when Karpathy posted that tweet reached $4.7 billion in 2025. Projections put it at $12.3 billion by 2027.
But the statistic that matters most for understanding what's actually changing isn't about developers at all.
The Sixty-Three Percent
According to community analysis data aggregated by Second Talent, 63% of active vibe coding users are not developers. Product managers are building full-stack applications. Founders are creating MVPs without engineering teams. Designers are deploying interactive prototypes directly to production.
The breakdown: 44% are building user interfaces. 20% are creating full-stack applications. 11% are making personal software tools — dashboards, automation scripts, custom CRMs tailored to their specific workflows.
For the first time in computing history, the majority of people creating software are not trained to create software.
What this looks like in practice: a teacher who needs a grading tool with specific rubric categories doesn't search the App Store for the closest approximation and pay $4.99/month for features she doesn't use. She describes what she wants and receives it — built to her specifications, matching her actual workflow, in minutes. A small business owner who needs inventory tracking doesn't adapt her processes to fit someone else's assumptions about how inventory works. She describes her process and gets a tool shaped to it.
Software has historically been something you acquire — bought, downloaded, subscribed to. Someone else decided what features it would have, what workflows it would support, what it would cost per month. The user's job was to find the closest match and adapt.
Vibe coding inverts this. Software becomes something you describe and receive. The user's job is to know what they want. The tool's job is to build it.
The Platform Race
The companies building infrastructure for this shift are being valued accordingly.
Replit — a browser-based platform where users build and deploy applications through AI-assisted natural language — reached a $9 billion valuation in 2026, tripling from $3 billion in six months. The company is targeting $1 billion in annual recurring revenue by year-end.
Emergent, backed by Y Combinator, coordinates teams of specialized AI agents to design, code, and deploy full-stack applications. It raised a $70 million Series B at a $300 million valuation.
Vercel's v0 generates production-grade React components from text descriptions. Lovable and bolt.new compete for the instant-deploy market — idea to live application in under a minute.
In January 2026, Wix launched Harmony — a hybrid platform that lets users alternate between natural language prompts and visual editing. The same month, Wix acquired Base44 for $80 million, a vibe coding platform that had grown fast enough to attract an acquisition before its authentication vulnerabilities attracted a security report.
25% of Y Combinator's Winter 2025 batch shipped codebases that were 95% or more AI-generated. The accelerator that selects for the strongest startup founders in the world is now selecting founders whose primary technical capability is describing what they want built.
The SaaS Question
The software industry noticed.
In January 2026, Anthropic released Claude Cowork and Claude Code — AI tools designed to build software and automate workflows. The market response was immediate: roughly $300 billion in market value evaporated across publicly traded software companies in a single trading session.
The industry coined a term for what followed: the "SaaSpocalypse."
The economics are straightforward. Building custom software used to be expensive — specialized developers, months of development time, budgets that made SaaS subscriptions the obvious choice. When AI tools compress that timeline from months to hours and the cost from thousands to effectively zero, the asymmetry that SaaS revenue was built on inverts.
Not everyone agrees the inversion is fatal. Bain & Company analysis notes that enterprise software includes compliance infrastructure, security governance, audit trails, and integrations that take years to build and are deeply embedded in institutional operations. Gartner predicts that by 2030, 35% of point-product SaaS tools will be replaced by AI agents — implying that 65% survive in some form.
The distinction analysts keep returning to: "software that performs a task" is vulnerable to replacement. "Software that carries institutional knowledge" is not — at least not yet. A custom CRM built in twenty minutes can replace Salesforce for a five-person team. It cannot replace Salesforce for a five-thousand-person enterprise whose decade of customer data, workflow automation, and compliance infrastructure lives inside it.
The question isn't whether SaaS dies. It's which layers of the software stack become commoditized when any user can generate a working application by describing what they want.
What Goes Wrong
Custom software built by non-developers inherits the limitations of both the tools and the builders.
A December 2025 analysis by CodeRabbit of 470 GitHub pull requests found that AI co-authored code contained approximately 1.7 times more major issues than human-written code. Security vulnerabilities appeared at 2.74 times the rate. Palo Alto Networks' Unit 42 found that AI coding models "prioritize function over security" — generating working output without access controls or input validation.
Escape.tech scanned over 5,600 publicly deployed vibe-coded applications and found over 2,000 vulnerabilities, more than 400 exposed secrets, and 175 instances of PII exposure — including medical records. Every vulnerability was in a live production application.
Red Hat observed that vibe-coded projects typically collapse around the three-month mark. Without specifications, the reasoning behind code decisions evaporates — because no one wrote the code, no one can explain what assumptions it made.
On March 18, 2026, Apple blocked updates for vibe coding apps Replit and Vibecode from the App Store. The stated concern: these apps generate software that bypasses App Store review. The unstated implication: if users can build their own applications, the entire model of curated distribution is in question.
72% of professional developers say vibe coding is not part of their professional work. Developer trust in AI code accuracy fell from 40% to 29% year-over-year.
The gap between what vibe coding produces and what professional software engineering requires is real. It is also, in broad strokes, the same gap that exists between a home cook and a restaurant kitchen — and both produce food that people eat.
Where Custom Stops
For systems where software failure has physical consequences, the boundaries are regulatory and absolute.
Before vibe coding existed, rigorously engineered software killed people. The Therac-25 radiation therapy machine delivered doses up to 100 times the intended level. The Boeing 737 MAX MCAS relied on a single sensor; 346 people died. These systems were developed by professional engineers, reviewed by teams, and certified by regulatory agencies.
The regulatory frameworks for safety-critical software — DO-178C for aviation, ISO 26262 for automotive, IEC 62304 for medical devices — share one requirement: every line of code must be understood, traced to a requirement, and proved tested. Software generated by describing what you want doesn't meet this standard.
In early 2026, Karpathy himself proposed "agentic engineering" as the successor to vibe coding — AI agents that write code under structured human oversight, with quality gates and audit trails. The evolution from "fully give in to the vibes" to "structured oversight with accountability" tracks the arc of every technology that starts casual and becomes infrastructure.
You Wouldn't Download a Car
The MPAA told paying audiences they wouldn't download a car. The premise was absurd because downloading a car was impossible.
Twenty-two years later, the premise is absurd for a different reason. Sixty-three percent of vibe coding users aren't developers. A market that didn't exist eighteen months ago is worth $4.7 billion. $300 billion in software company market value evaporated on the announcement of tools that let people build their own. Apple blocked the apps. Replit tripled its valuation in six months.
You wouldn't download a car. Not because downloading is wrong — the MPAA lost that argument two decades ago — but because downloading is increasingly beside the point. If you can describe what you want and receive it, built to your specifications, in minutes rather than months — why would you download someone else's version?
The era of custom software is arriving. For the 63% who aren't developers and are building anyway, it's already here. The remaining question — the one the regulatory frameworks, the security research, and the professional skepticism are all trying to answer — is where "custom" ends and "engineered" begins.
The MPAA's format still works. There are things you wouldn't vibe code. A car is one of them. But the list of things you would is growing faster than anyone predicted — and every application on that list is one fewer download, one fewer subscription, one fewer compromise between what you wanted and what someone else decided to build.
