In February 2020, years before anyone was "vibe coding," the writer Robin Sloan published a short essay about a messaging app he built for his family. Four people use it. He described it this way: "It is ruthlessly simple; we love it; no one else will ever use it." And he described himself as "the programming equivalent of a home cook."
Back then, building that app a family could use took a programmer with technical knowledge. Now it takes a paid AI subscription and typing a prompt into an AI app. So it's worth thinking past the individual app and asking the bigger question: what happens when millions of people stop buying software and start cooking up their own?
Some of it is actually good. Software has always been built for the average customer, which means it fits nobody exactly. A bakery owner who builds her own order tracker gets one that matches how her bakery actually works, with her product names, her pickup windows, and none of the forty features she'd never touch. Multiply that across every small business, club, classroom, and family, and you get a lot of people whose tools finally fit them. Sloan's line about this holds up well: things change "when you liberate programming from the requirement to be professional and scalable."
Put for people who depend on it for income, that's another story. But let's focus on the people that need to still pay.
The bill moves to the people least ready to pay it
When you buy software, the vendor carries the maintenance. When you build it, you do yourself, and AI-written code needs plenty of it.
Security is the clearest example. Veracode tested code from more than 100 AI models and found that "AI-generated code introduced risky security flaws in 45% of tests." Worse, "Larger, newer AI models didn't improve security." Professional developers have tools and habits to catch those flaws. A bakery owner usually doesn't, and her homemade order tracker might hold customer phone numbers and addresses.
Developers themselves aren't fully sold either. In Stack Overflow's 2025 survey, 84 percent said they use or plan to use AI tools, but "more developers actively distrust the accuracy of AI tools (46%) than trust it (33%)." The top frustration, named by 66 percent, was "AI solutions that are almost right, but not quite." Almost right is fine for a family photo app. It's a problem for anything that handles money, health, or other people's data.
So one likely outcome is millions of small apps that work well on day one and quietly decay: a library update breaks something, a security hole goes unpatched, the person who built it moves on, and nobody else knows how it works. Claude Code will get things wrong sometimes, and when everyone is their own developer, there's no support line to call.
Everyone's data ends up in a different shape
Bought software comes with a shared format. Every company using the same accounting tool can export the same file and hand it to the same accountant. When everyone builds their own, every business stores its data a little differently. That's invisible until two of those systems need to talk to each other, or the person who built one leaves, or the business gets sold and the buyer gets a mystery app instead of a standard one.
That's why I think the software people keep paying for shifts more than it shrinks. The pieces that hold everything together, like email delivery, payments, logins, hosting, databases, and maps, get more important as the number of homemade apps on top of them grows. I wrote about where that line falls for an individual deciding what to build. At the scale of a whole economy, the same line holds: the thin tools get rebuilt at home, and the infrastructure underneath gets busier.
The developer's job shifts too. Writing code gets cheap. Reviewing it, securing it, and rescuing the homemade app that has grown into something a business depends on gets more valuable. A new kind of work is likely to grow around people who built something with AI and now need someone who actually understands it.
There's also a discovery problem. When anyone can build, everyone builds, and most of it never finds an audience, which I've covered in why almost nobody can get people to find what they build. For home-cooked apps that's fine, since they were never meant to be found. For anyone hoping to turn one into a business, it's the main obstacle.
Sloan's app had four users and no plan to grow. Most of what people build with AI will look like that, and that's probably the healthiest version of all this: software made for a handful of people who know exactly who to ask when it breaks.
Sources
Robin Sloan: An app can be a home-cooked meal - The February 2020 essay about a messaging app built for his family and used by four people, describing himself as the programming equivalent of a home cook and programming freed from being professional and scalable.
Veracode: GenAI Code Security Report (October 2025 update) - Code from more than 100 AI models introducing risky security flaws in 45 percent of tests, and larger, newer models not improving security.
Stack Overflow Developer Survey 2025: AI - 84 percent of developers using or planning to use AI tools, 46 percent distrusting AI accuracy versus 33 percent trusting it, and 66 percent naming almost-right AI solutions as their top frustration.
