Will AI Replace All Programmers? What the Evidence Actually Shows
Somewhere between the third month and the sixth, the code stops being yours. A model writes nine out of every ten lines, and within a year, it writes essentially all of them. That was the timeline Dario Amodei, the CEO of Anthropic, laid out in March 2025 — not for some distant era of artificial general intelligence, but for the ordinary, everyday work of software engineers, on a horizon measured in months rather than decades.
The prediction did exactly what predictions like that always do: it became one of the most repeated sentences in technology. Founders quoted it in pitch decks. News sites ran it as a countdown. And in comment sections and group chats, programmers — the very people who had spent years mastering the craft — traded the sentence back and forth like a diagnosis. As it traveled, the viral version got bolder. AI will replace all programmers. Not some tasks. Not eventually. All of them. Soon.
It is a claim about the future, and the future is the one place a fact-check cannot go in person. What it can do is examine the evidence available right now: what AI is actually doing inside software companies, what the employment data shows, what the most rigorous productivity experiments have found, and — perhaps most revealingly — what the people making the prediction say when they finish their own sentences.
The claim and the men who made it
On March 10, 2025, at a Council on Foreign Relations event, Anthropic’s CEO Dario Amodei said something that would be quoted, clipped, and argued over for the next seventeen months: “I think we’ll be there in three to six months, where AI is writing 90 percent of the code. And then in twelve months, we may be in a world where AI is writing essentially all of the code.”
Note precisely what the sentence claims. Not that AI will help write code. Not that it will assist programmers. That it will write essentially all of the code — first 90 percent, then all — on a timeline that, counting from that March day, lands in early 2026. That is this month. You can now check the prediction against reality yourself.
Amodei kept raising the stakes. In Davos in January 2025, he had said that AI better than humans at almost all tasks was “two or three years” away. Then at a World Economic Forum panel in January 2026, he said AI models might be “6 to 12 months away from when the model is doing most, maybe all of what [software engineers] do end-to-end.” The framing had shifted from code-writing to the whole job: most, maybe all, of what software engineers do.
He was not alone at the microphone. In April 2025, Microsoft CEO Satya Nadella said AI generated 20 to 30 percent of the code in some of Microsoft’s projects. Google CEO Sundar Pichai put the company’s figure above 30 percent of new code. At Meta’s LlamaCon in April 2025, Mark Zuckerberg estimated that AI was doing about half of the company’s development work. Then in January 2026 came the most extreme data point yet: Anthropic announced that more than 90 percent of the code powering new versions of its own flagship model, Claude, was generated by Claude itself. Boris Cherny, the creator of Anthropic’s Claude Code developer tool, disclosed that every line of code he personally shipped in December 2025 was written by the model. Shopify’s CEO Tobias Lütke declared in an April 2025 internal memo that “reflexive AI usage is now a baseline expectation” at the company, and an internal AI agent called River now starts roughly half of Shopify’s code pull requests.
Read in a row, these numbers feel like a countdown to the end of a profession. But notice what every one of them measures: the share of code characters that came out of a model. None of them measures employment. That substitution — from “a share of code” to “a share of jobs” — is where the viral claim quietly does its most important and least examined work.
What AI is actually doing inside software teams
The adoption figures are real, and they are startling. The 2025 Stack Overflow Developer Survey found that 84 percent of developers were using AI tools in their work. JetBrains’ 2026 developer research put it even more bluntly: 90 percent of professional developers surveyed used AI coding agents at least weekly, and 68 percent used them daily. There is no serious version of the future in which programmers go back to writing every line by hand. That era is over.
But adoption is not the same as confidence, and this is where the story gets complicated. Stack Overflow’s 2026 data found that while 84 percent of developers were using or planning to use AI tools, 46 percent distrusted the accuracy of what those tools produced. A SonarSource survey of more than 1,100 developers found that 96 percent did not fully trust AI-generated code for correctness — and yet only 48 percent said they always verify it before committing.
Sit with that gap for a moment: near-universal use alongside near-universal distrust. If AI-generated code were simply correct, programmers would not need to spend their days checking it. The checking has become the job. AI writes the first draft at inhuman speed; the human becomes the reviewer, the editor, the person who stands between the model’s output and production and takes responsibility for what ships. That is not a smaller job than the old one. In many ways it is a harder one — it is easier to write your own careful code than to audit someone else’s fast code for the one line that will take down a database at 3 a.m.
This is why the bluntest version of the prediction keeps tripping over its own evidence. The models are extraordinarily good at producing code. Producing code was never the whole job.
The jobs data says it’s complicated
Start with the official projections, because they are the least emotional numbers in this debate. The U.S. Bureau of Labor Statistics projects employment of software developers, QA analysts, and testers to grow about 10 percent from 2025 to 2035 — much faster than the average for all occupations — with roughly 106,100 openings a year and a median pay near $136,000. That is a growth forecast, published by the federal government, for a period in which AI coding tools are already everywhere.
At the same time, the BLS projects a decline of about 6 to 7 percent over the same period for “computer programmers” — the occupation defined more narrowly around writing code from specifications. The agency explicitly flags AI as a factor, noting that repetitive programming tasks are increasingly being automated.
This is the distinction the viral claim cannot survive: the narrow activity of turning a specification into code is officially expected to shrink; the broader engineering role is expected to grow, adding on the order of 175,000 jobs. The government’s own forecast draws the line the viral claim erases.
The hiring data tells the same story, with one important wrinkle. Indeed’s Hiring Lab found in July 2026 that U.S. software development job postings were up about 15 percent since the February 2025 launch of Anthropic’s Claude Code, even as overall job postings fell 7 percent. Nearly three-quarters of that increase came from senior roles, and senior positions made up 69.3 percent of all software development postings in the first quarter of 2026 — the highest share of any occupation Indeed tracks. Postings were still about 27.5 percent below February 2020 levels, so the market has not fully recovered. But within a weak labor market, software engineering postings have been gaining ground, not losing it.
The wrinkle is at the bottom of the ladder, and it is real. The Stanford Digital Economy Lab, analyzing ADP payroll records, found that employment among workers aged 22 to 25 in the most AI-exposed occupations sat about 19 percent below where it would have been had those workers tracked less-exposed fields — and software development was at the center of the effect, alongside customer support. Crucially, the adjustment came through lower hiring rather than layoffs. A team that once needed six people simply never posts the sixth job. Nobody gets fired; the position just quietly stops existing. Junior work — boilerplate, first-pass tests, documentation, translating a tight specification into code — is exactly where coding assistants are strongest, and it is exactly where the door is closing.
Then there are the layoffs, which supply the viral claim with its most dramatic imagery. The Financial Times counted roughly 140,000 U.S. tech job cuts year-to-date in 2026. Challenger, Gray & Christmas counted about 55,000 U.S. layoffs explicitly citing AI in 2025 — twelve times the 2023 figure. The individual cases read like a roll call of the thesis: Block cut nearly half its staff; Salesforce cut about 4,000 support roles as CEO Marc Benioff said he needed “less heads”; Coinbase cut 14 percent of its workforce, with its CEO saying engineers now ship in days what used to take a team weeks.
But the AI story inside those layoffs is muddier than the headlines suggest. Klarna’s CEO said in May 2025 that the company had focused too much on cost, gotten lower quality as a result, and was hiring humans again. A Robert Half survey of about 2,000 managers in April 2026 found that 32 percent of those who eliminated a role because of AI later rehired for the same or a similar role. Gartner predicted that half of the companies attributing cuts to AI would rehire similar roles by 2027. The Financial Times found that companies which cited AI for their cuts underperformed the Nasdaq by about 10 percent over the following 30 trading days. AI researcher Andrew Ng has a term for the whole phenomenon — “AI washing”: dressing up pandemic-era over-hiring corrections and interest-rate pressure as an AI revolution. A great deal of the 2022–2024 tech bloodletting, on this reading, was a correction for hiring too many people too fast when money was free. AI is the press release, not the cause.
None of this means AI has no effect on employment. It means the effect is specific, lopsided, and mostly invisible in the aggregate numbers: fewer doors opening at the entry level, more leverage for experienced engineers, and a round of layoffs in which AI serves as both accelerant and alibi.
Writing code is not the same job as building software
This is the distinction that collapses in every viral retelling of the claim. Writing code is one part of building a software product. The rest of the work — understanding ambiguous requirements, negotiating with stakeholders who don’t know what they want, choosing an architecture that will still make sense when the product has three times the users, designing data models, managing authentication and permissions, integrating external systems, testing failure cases, securing data, deploying releases, monitoring production, handling incidents, deciding how the software should evolve — does not happen by autocomplete. None of the CEOs’ statistics measure any of it.
The most rigorous test of the productivity story so far makes this concrete. In 2025, the research organization METR ran a randomized controlled trial: sixteen experienced open-source developers completing 246 real issues in repositories they had maintained for years. Each task was randomly assigned to be done with AI assistance allowed or switched off — no cherry-picking easy tasks to make the tool look good. Before starting, the developers forecast that AI would make them about 24 percent faster. The measured result: the early-2025 tools made them 19 percent slower. And after completing the work — after living through the slowdown and observing their own output — the developers still believed AI had made them roughly 20 percent faster.
That last number is the one that should stay with you. It was not a forecasting error; it was a perception error that survived direct experience. The developers had access to their own outcomes and still could not feel a nearly 40-percentage-point gap between expectation and reality. Felt velocity is not a measurement.
METR reran the experiment in February 2026 with 57 developers and more than 800 issues across 143 repositories. The central estimate for the returning group flipped to an 18 percent speedup — but with a confidence interval spanning a 38 percent speedup to a 9 percent slowdown, a range so wide that METR itself cautioned against leaning on the number. More telling than the flip was why the rerun nearly failed as an experiment: developers increasingly refused to work without AI at all, even at generous pay, and between 30 and 50 percent of participants withheld the tasks they expected AI to accelerate — quietly steering the “no AI” condition toward harder work.
The honest reading of all of this is narrower than either side’s slogan. On greenfield work, small changes, tests, and boilerplate, AI saves real time: a 2023 Microsoft Research study found that developers using GitHub Copilot completed a bounded coding task 55.8 percent faster than a control group. On large, unfamiliar, heavily context-dependent changes — the maintenance-heavy work that senior engineers are actually paid to do, in codebases with years of implicit context a model cannot see — the leverage can cost you time. The tool narrows the gap between less and more experienced developers on well-specified tasks, and it widens the value of the judgment that no benchmark measures.
And here is the final irony of the viral claim, delivered by its own author. When Amodei announced that AI was writing over 90 percent of the code behind Claude, he was asked what that meant for his engineers. His answer: “If Claude is writing 90% of the code, what that means, usually, is you need just as many software engineers.” The same man who predicted AI would write essentially all the code also said the company writing the code would keep all of its programmers — to edit the code, to write the hardest 10 percent, to supervise the models. The prediction and the retraction arrived in the same interview.
The Verdict
The claim that AI will replace all programmers within a couple of years is Unproven — and the words “all” and “couple of years” are doing most of the heavy lifting.
Here is what the evidence supports. AI already generates a large and growing share of code inside the world’s most AI-forward companies — over 90 percent at Anthropic by its own January 2026 account. Routine coding tasks are being automated now, not later. Employment of 22-to-25-year-olds in the most AI-exposed occupations is running about 19 percent below where history would have predicted, because companies are hiring fewer juniors for work the models can draft. And the occupation narrowly defined as “computer programmer” is officially projected to shrink about 6 to 7 percent over the next decade.
Here is what the evidence does not support. The profession is not disappearing. The same Bureau of Labor Statistics projects software developer employment growing 10 percent through 2035. Software development job postings are up since the first wave of coding agents, tilted heavily toward senior roles. The most rigorous productivity experiment ever run on AI coding tools found the gains were neither automatic nor even visible on the hardest, most context-dependent work — and that developers systematically overestimated how much faster the tools made them. The wave of tech layoffs was driven substantially by over-hiring corrections and interest-rate pressure, with AI serving as both accelerant and alibi — and a meaningful share of the “AI layoffs” have already been quietly reversed.
“Replace all programmers” is a statement about the future, and futures are not facts. But as of September 2026, the evidence says AI is transforming the job — automating the typing, shifting the value toward judgment, review, and architecture, and quietly narrowing the door for junior developers — rather than eliminating the people who do it. That transformation is still worth worrying about, especially if you are 22 and looking for your first role. It is just not the same as the viral claim.
Amodei’s original sentence has now outlived its own deadline. “Three to six months” from March 2025 expired sometime last summer, and programmers, as a profession, are still here — posting job ads, reviewing pull requests, and arguing about the prediction in the comments of the articles that quote it. The next time the sentence comes traveling past you, watch which verb it carries. “AI writes most of the code” is a fact about the present. “AI will replace all programmers” is a story about the future. Facts deserve belief. Stories deserve scrutiny. That is the whole job of a fact-check.