Skip to content
Argument

The app is born public

Making software became a mass skill decades ago, with the spreadsheet. What AI builders change is who uses the result on its first day.

· 4 min read

If a person who has never programmed can describe a tool in plain English and get one that runs, it is tempting to say that software is becoming a general literacy, the way writing did, and that everyone is in some sense becoming a software engineer. Making software became a mass skill long before AI, through the spreadsheet, and it was never the same thing as engineering. Engineering is what making becomes when other people depend on the result. The real change the AI app builders bring is that they put beginners on the engineering side of that line from the first day, usually without telling them.

A survey of end-user software engineering published in 2011 drew on federal labor statistics to project that by 2012 the United States would have fewer than 3 million professional programmers and more than 55 million people using spreadsheets and databases at work, many of them writing formulas and queries.1 Raymond Panko, who compiled much of the research on spreadsheet errors, put the global figure at tens of millions of managers and professionals creating hundreds of millions of spreadsheets a year.2 A spreadsheet formula is a program, and whoever writes one is programming. By headcount, making software was a mass skill before the web was.

That literacy had a known error rate. In the field audits of real business spreadsheets that Panko collected, the ones conducted from 1997 onward, which used better methods than earlier audits, found errors in 91% of the 54 spreadsheets examined.2 In careful audits, between 1.1% and 2.5% of cells contained an original error, about what the wider research on human error would predict for complex mental work.2 The authors of these spreadsheets were, in the survey's phrase, "heedlessly confident" about correctness.1 In one study Panko cites, 42% of spreadsheet developers said the highest-level user of their data was the chief executive.2

A literacy that wrong that often survived because most of its mistakes stayed inside a file. A broken formula reached the world only through a number someone read and a decision someone made, and that person could check it, doubt it or ignore it.

The 2011 survey separates the two activities by who the program is for.1 Borrowing a phrase Bonnie Nardi popularized in her studies of office spreadsheets, it defines end-user programming as programming "primarily for personal, rather public use", and says professional programming has "the goal of producing code for others to use".1 The definition turns on intent, not skill or language: a skilled developer writing helper code for her own task is doing end-user programming, and "the moment novice web designers move from designing a web page for themselves to designing a web page for someone else, the nature of their activity has changed."1 Once others use the program, the author has to plan for a broader range of possible uses, and design and testing matter more.1

The person who named the AI version of the practice placed it on the personal side. Andrej Karpathy's post coining "vibe coding" in February 2025 described accepting every change, never reading the diffs, and forgetting "that the code even exists", and judged it "not too bad for throwaway weekend projects".3 Simon Willison, a programmer who has written at length about building with these tools, defined vibe coding as building software without reviewing the code, endorsed it for low-stakes projects and prototypes, and singled out software "used by other people" as the case that needs care.3

The beginners did not stay there. In a survey Zapier commissioned in May 2026 of 818 American workers who had built or deployed a work tool with AI coding tools in the past year, 34% had no formal programming background.4 Of those, 56% said they were making something customers or the general public would see, and 44% built only for internal use.4 The data is unweighted and comes from a survey a software vendor paid for, so the proportions are a sketch, not a census.4 A spreadsheet began as a private file on one desk. A web app has a public address, and its first user is often someone else.

In March 2025 Matt Palmer, an employee of the rival company Replit, scanned 1,645 projects from the showcase of Lovable, an AI app builder, and found 303 endpoints across 170 of them, about 10%, whose database access rules were missing or inadequate.5 The exposed tables included users, transactions and subscriptions, along with API keys, and the scan looked only at homepages.5 Lovable apps run mainly in the browser and rely on outside services for sign-in and data storage, which, Palmer wrote, "shifts the security burden to the implementor of the application"; the company's later security scanner checked only that some access policy existed, not whether it was correct.5 His hypothesis was that many of these builders did not understand the database's access settings and did not know they were exposing anyone's data.5

Other people's data, an adversary, and a responsibility the author did not know they had accepted are the core of engineering, and none of them appears in the prompt. The previous wave of tools for non-programmers kept that responsibility somewhere else. Gartner forecast that by 2026 developers outside formal IT departments would make up at least 80% of low-code users, and still expected the external projects, the ones built for customers, to be done by IT teams and professional developers.6 An AI builder has no such division: the same chat window that makes a personal to-do list makes a shop.

Willison's answer is a container. He praised the sandbox of Claude Artifacts, where code runs locked in the page, loads only approved libraries and cannot make requests to other sites, because it makes it very hard for a beginner to harm anyone, at the price of limiting what the project can do.3 That design restores what the spreadsheet had by accident: a default audience of one, so that whatever is wrong stays with its author until someone chooses to share it. The literacy is safe to spread exactly as far as its tools keep it private.

Sources

  1. 1
  2. 2
  3. 3
  4. 4
  5. 5
  6. 6

About this piece

Kind
Argument. A reasoned case. The logic is the evidence; check it.
Evidence
Moderate
Research
Standard, 6 sources
Made
My idea. AI researched and wrote it, and an AI model checked each sourced claim against its source before I published it here.
Updated
Nov 21, 2026

How this was made →Write to me about it →

aisoftware engineeringwork