Spreadsheets were the first vibe coding
In 1979 VisiCalc let people who had never used a computer write programs without calling them that. Forty-seven years of spreadsheet errors, and of the people paid to catch them, say what happens next to the software people now build with AI.
In the compiler essay I made a prediction I liked a lot. When a working tool costs an evening instead of a quarter, I wrote, “people will make software the way they make spreadsheets today: for one problem and one person, thrown away when it’s done.”
I meant it kindly. Then I went to check what spreadsheets have actually done in the forty-seven years since the first one, and the last five words don’t survive. Researchers at Delft surveyed 27 analysts at Robeco, a Dutch asset manager, in 2011. On average a spreadsheet there lived more than five years and was used by 13 different analysts. Those spreadsheets weren’t thrown away. They got handed on.
So I want to do the comparison properly, because I think spreadsheets are the best guide we have to what happens to the software people are now building with AI without reading the code. They were the first time millions of people who weren’t programmers started writing programs. We know how that went: it took thirteen years before anyone was paid to check their work, and twenty before checking it became a field of its own.
The program nobody called a program
VisiCalc shipped in mid-October 1979, for the Apple II, at $99.50. Dan Bricklin and Bob Frankston wrote it. Three months before it shipped, Ben Rosen, the electronics analyst at Morgan Stanley, wrote to his clients about it: “In minutes, people who have never used a computer are writing and using programs.” He had built a dividend discount model in BASIC, which took him 20 or so hours to write and debug. In VisiCalc it took 15 minutes.
So the investor called it programming before it even shipped. The users didn’t, and that was the design working. Simon Peyton Jones, Alan Blackwell and Margaret Burnett, writing in 2003 about how to add functions to Excel, counted it among the great merits of spreadsheets that users “need not think of themselves as doing ‘programming’”. They write formulas, or build a model. Bricklin put it more bluntly in 2010: users “do more and more of the programming themselves, but they don’t know it.”
It sold: about 100,000 copies of VisiCalc by spring 1981, and about 700,000 by 1984. (You’ll see “a million” quoted, but I couldn’t find where that comes from.) Lotus 1-2-3 was the best-selling business program within three months of shipping in 1983, and Excel took over after that. Microsoft counted 30 million Excel users in 1996. By 2021 its researchers were writing “hundreds of millions” of Excel users, against “tens of millions” of developers. (The “750 million” and “a billion” figures that go around don’t trace back to Microsoft either.)
The best count of who was really programming comes from Carnegie Mellon. In 2005 Christopher Scaffidi, Mary Shaw and Brad Myers projected that in 2012 more than 55 million Americans would use spreadsheets or databases at work, against fewer than 3 million professional programmers. Their own guess was that about 25 million of those users were actually programming: writing conditions, linking sheets, building things other people would rely on. That’s still about eight of them for every programmer. Bank regulators got there eventually: by 2013 the Basel Committee was writing about bank units that use spreadsheets “for software development”.
Microsoft’s researchers already called Excel “the world’s most widely used programming language”. In 2021 they added LAMBDA, which lets you define your own functions inside a formula, and announced: “With LAMBDA, Excel has become Turing-complete.”
What went wrong
Raymond Panko at the University of Hawaii spent decades collecting studies of spreadsheet errors, and the number people usually quote from him is that almost every spreadsheet has one: 94%, in his 2015 review. That number is thinner than it sounds. It comes from 85 spreadsheets, and most of the audits behind it are known only from a press release, a vendor’s web page and personal emails. In his 2000 review, all the field audits he had found errors in 24% of 367 spreadsheets, though the older ones weren’t looking very hard.
The number I find more interesting is the rate per cell. Across 14 lab studies, 967 people building spreadsheets on their own got 3.9% of their cells wrong. In four big sets of industrial code inspections, professional programmers made errors in 1.9% to 3.7% of their lines of code. Spreadsheet builders weren’t much worse than programmers. They just had nobody reading their work. A second look, in Panko’s inspection experiments, caught about 60% of the errors.
Everybody makes mistakes at about that rate. What decides the damage is whether anyone reads the work before it runs for five years. A careful audit of 25 working spreadsheets by Powell, Baker and Lawson at Dartmouth found that most errors had little effect, and concluded that “Devastating errors are rare.” Rare across hundreds of millions of users still adds up to a long list.
In January 2012, JPMorgan’s model reviewers looked at a new value-at-risk model for the bank’s Chief Investment Office and wrote the problem down: “The VaR computation is currently done off spreadsheets using a manual process. Thus it is error prone, and not easily scalable.” The same email told them to go ahead with it. The bank’s own task force later found that the spreadsheet divided by the sum of two rates instead of their average, which probably muted the volatility by a factor of two. The office lost $5.8 billion in the first half of 2012, and another $449 million in the next quarter. The spreadsheet didn’t make the trades. It halved the number that was supposed to raise the alarm.
Genetics has been fighting Excel for a decade. Type a human gene called SEPT1 or MARCH1 into a cell and Excel turns it into a date. A 2016 study found the damage in 19.6% of papers that came with Excel gene lists. The follow-up in 2021 found 30.9%, and called the errors “unabated”. The committee that names human genes spent years telling people which setting to change, and gave up: “we could not possibly reach everybody that uses Excel”. During 2019 it renamed the genes instead, so SEPT1 is now SEPTIN1. Microsoft added a setting to stop the conversion in October 2023.
And in Edinburgh, a new children’s hospital was due to open in July 2019 when someone noticed that critical care was specified for 4 air changes an hour, where the guidance requires 10. The specification was an Excel workbook covering 1,839 rooms. The internal audit traced the error to a row for a generic four-bed room, copied into the critical care rooms, where it had sat since 2012. The opening was stopped days before.
(Two of the most famous cases are weaker than they look. In Reinhart and Rogoff’s paper on public debt, the Excel formula that left out five countries accounted for about 0.3 of the 2.3 points by which the paper overstated its gap; most of the rest came from how they weighted the data. And for England’s 15,841 missing Covid cases in 2020, the official statements only ever said “maximum file size”. The Excel explanation came from unnamed sources.)
The second owner
Back to Robeco, because that’s the part I got wrong. A spreadsheet that lives five years and passes through 13 analysts is a program in production, and its author may have left the building. The thirteenth analyst inherits formulas they didn’t write, assumptions nobody wrote down, and a number the business now depends on.
Enron shows how they travelled. Its email became public after the collapse, and in 2015 Felienne Hermans and Emerson Murphy-Hill went through it: about 100 emails a day went out with a spreadsheet attached. At Robeco, 85% of the analysts said they often passed a spreadsheet on to a colleague. For a lot of the software that ran businesses, that was maintenance: a file in an inbox, with no history and no tests.
The checkers
The fix was people, and it came slowly.
The first lab study of people building spreadsheets came out of IBM in 1987, eight years after VisiCalc, and the first published audit of real ones appeared in an Australian accounting magazine the same year. In 1992, UK tax inspectors at HM Customs & Excise audited 273 spreadsheets that companies had filed with their returns, and found errors big enough to demand more tax in 11% of them. By 1999, twenty years in, the inspector who ran that audit, Ray Butler, co-founded the European Spreadsheet Risks Interest Group with researchers from Greenwich and Cardiff. Its first formal meeting was held at the tax office in London.
Then the auditors arrived. The Sarbanes-Oxley Act of 2002 made American public companies vouch for their controls over financial reporting, and the spreadsheets that fed the numbers were pulled in. In October 2003 Fannie Mae published quarterly results that understated shareholders’ equity by $1.136 billion, and put it down to mistakes in a spreadsheet. Microsoft bought Prodiance, a company that sold spreadsheet controls, in 2011 and shipped its tools with Office 2013, to deal with “critical spreadsheets and databases created by end users without the involvement of IT”. The Basel Committee’s 2013 principles asked banks for “end-user computing policies and procedures”. The UK’s Prudential Regulation Authority brought end-user computing tools into its model risk rules in 2023.
So: eight years to the first study, thirteen to the first paid checkers, twenty to a group of their own, and thirty-three to the tool maker’s controls and a global banking standard. Not every regulator moves the same way, mind. In April 2026, the US banking agencies replaced their model risk guidance with a version that leaves out “simple arithmetic calculations, such as those found within spreadsheets”, and puts generative and agentic AI models outside its scope.
Spreadsheets made jobs on the other side of the ledger too. Planet Money counted 400,000 bookkeeping and accounting clerk jobs gone since 1980, and 600,000 accounting jobs added. That’s the Jevons story again: analysis got cheaper, and people wanted a lot more of it.
Year one of vibe coding
Andrej Karpathy named vibe coding on 2 February 2025: you “fully give in to the vibes, embrace exponentials, and forget that the code even exists.” He accepted every change and no longer read the diffs. People quote the start of that post and drop the end, where he says it’s “not too bad for throwaway weekend projects”. Throwaway, again.
Collins made it the word of the year nine months later. In June 2026 Lovable, one of the app builders, said that “80% of people building are in self-identified non-technical roles”, and that a million new projects were started on it every week. Those are the company’s numbers. A survey Zapier commissioned in May 2026 is smaller and more useful: of 818 American workers who had built a work tool with AI coding tools in the past year, 34% had no formal programming background, and more than half of those said most or all of their tools were still in use.
Even the spreadsheet joined in. Microsoft announced “vibe working” in September 2025: an agent mode in Excel that decides which formulas to use and builds the sheets. A month earlier it had added a =COPILOT() function, so a formula could ask the model for an answer. On 14 September 2026 Microsoft removed it. Old results stay as cached values, but recalculate the cell and it “returns a #NAME? error because it’s no longer recognized.” That’s thirteen months from launch to a broken formula.
The big difference from 1979 is where the program lives. A VisiCalc model sat on one desk and travelled on a floppy disk, later by email. Its errors hurt the company that owned it, and they usually took years to surface. A vibe-coded app is on the public internet from the first day, with a login page and a database, and its users are strangers. So the bill came in weeks.
Six weeks after Karpathy’s post, a researcher found an app built with Lovable whose database anyone could read. By the time it was disclosed, a scan of 1,645 apps from Lovable’s own showcase had found 170 exposing data: names, phone numbers, API keys, payment details. Lovable disputed the vulnerability report, on the grounds that “each individual customer of the Lovable platform accepts a responsibility over protecting the data of their application.” In January 2026 Moltbook, a social network for AI agents, left its production database open to anyone who read the key out of its own web page: 1.5 million API tokens and 35,000 email addresses. Its founder had said “I didn’t write a single line of code for @moltbook.”
(Not every famous story holds up. Replit’s agent did delete a production database during a code freeze in July 2025, but it was an investor’s test build, and the rollback worked. And the Tea app breach that summer came from an older storage system, not vibe coding.)
The cleanup became a job within seven months. In September 2025, 404 Media reported that “freelance developers and entire companies are making a business out of fixing shoddy vibe coded software.” Spreadsheets waited thirteen years for their first paid checkers.
- Someone measures the errors
Spreadsheets: 1987, year 8 IBM’s lab study of people building spreadsheets, and the first audit of real ones.
Vibe coding: April 2025, month 2 A scan of 1,645 apps built with Lovable finds 170 exposing their data.
- Someone is paid to check
Spreadsheets: 1992, year 13 UK tax inspectors audit 273 spreadsheets filed with tax returns.
Vibe coding: September 2025, month 7 404 Media reports developers and agencies paid to fix vibe-coded apps.
- The tool maker adds checks
Spreadsheets: 2012, year 33 Microsoft ships spreadsheet controls in Office, after buying a spreadsheet-audit company.
Vibe coding: April 2025, month 3 Lovable adds a security scan to its app builder.
- A bank regulator names it
Spreadsheets: 2013, year 33 The Basel Committee asks banks that build with spreadsheets for end-user computing controls.
Vibe coding: Not yet. Canada’s bank regulator asked banks in July 2026 to check AI-written code before it goes live, not who wrote it.
Both tracks start at year zero: VisiCalc shipping in October 1979, and Karpathy’s post in February 2025. Sources: Panko’s reviews, HM Customs & Excise, Microsoft, the Basel Committee (BCBS 239), the CVE-2025-48757 timeline, 404 Media and Canada’s OSFI.
Three guesses
Here’s what I’d bet, and what would prove me wrong.
Most of it gets thrown away, and the rest outlives its builder. Karpathy said throwaway, and most of it will be. Gartner predicts, as one vendor summarised its June report, that 60% of vibe-coded apps will be retired by 2028, and that 30% of software engineering teams will be strained by maintaining the apps business users built. The Robeco numbers say what happens to the other 40%: they run for years and change hands, and the person who inherits one is a developer, paid to read code nobody read the first time. GitClear’s 2026 report already shows code being written and then left alone: the share of changed lines that move existing code around, which is what refactoring looks like, fell from 13% in 2023 to 3.8% this year. I’m wrong if the second owner can simply regenerate the app from a description and never needs to read it, and fixing vibe-coded apps stops being a line of work by 2028.
The builders will outnumber the developers, as they did with spreadsheets. Spreadsheet users who were really programming outnumbered programmers about eight to one in the United States. SlashData counts about 47 million developers in the world. If Lovable’s 80% holds as these tools spread, most people building software will not be developers. I’m wrong if, by 2028, the people building with these tools turn out to be mostly developers, the way Y Combinator’s startups with 95% AI-written code turned out to be run by founders its partner called “highly technical”.
The checkers arrive in years, not decades. Spreadsheets waited thirteen years for paid checkers and thirty-three for a bank regulator to name them. Vibe coding got paid fixers in seven months, and in July 2026 Canada’s bank regulator told banks to validate AI-generated code “prior to production deployment”. Nobody has named the thing itself yet: software built with AI by people outside IT, running part of the business. I bet a bank regulator or an audit standard setter does it by the end of 2028, the way the Basel Committee named spreadsheets. The UK looks likeliest, since its rules already cover end-user computing; the US just went the other way. I’m wrong if 2029 starts with vibe-coded apps covered only by general lines about AI-written code.
Back to my sentence
So I’d write the compiler essay’s line differently now. People will make software the way they make spreadsheets: for one problem and one person. Then the problem will turn out to matter, someone else will inherit it, and somebody will be paid to find out what it does.
Have you inherited a spreadsheet, or an app, from someone who’d already left? Tell me in the comments what was in it.
A thought to add?
A good question, another perspective, or something this made you think about. There’s room for it here.
Loading the conversation…