Overview:

The CEOs of Silicon Valley promise that AI will eliminate programming jobs — but the history of technology suggests the opposite. Just as cheaper computers led to more computing, not less, AI-assisted coding is likely to lead to an increase of software demand that keeps programmers busy. The real threat to your job isn't the technology itself; it's a shareholder-driven economy that rewards companies for firing developers regardless of whether it makes technical sense. This article argues that we are heading not toward a world without programmers, but toward one where programmers do the work of ten people while their employers compete to see who can lay off the most — unless we choose a different path.

Introduction: The CEO Predictions

For the past several years, the chief executives of the largest artificial intelligence companies have been telling the world that white-collar jobs are on the verge of extinction. In July 2025, Ford CEO Jim Farley predicted that “artificial intelligence is going to replace literally half of all white-collar workers in the U.S.” (Ma, 2025). Similar pronouncements have come from the leaders of OpenAI, Google, and Microsoft, each competing to paint the most dramatic vision of an AI-transformed economy.

So far, the technology sector itself has been the most visibly affected. According to Layoffs.fyi, which tracks publicly reported job cuts in the technology industry, more than 700,000 tech workers lost their jobs between 2023 and 2025. The 2023 peak saw approximately 429,608 layoffs across 1,193 companies, followed by 152,000 layoffs in 2024 and roughly 122,500 in 2025 (SQ Magazine, 2026). Research from Stanford Digital Economy Lab found that early-career software developers aged 22 to 25 experienced a nearly 20 percent decline in employment from its late 2022 peak by July 2025, suggesting that entry-level programming positions have been particularly affected (Stanford Digital Economy Lab, 2026).

This wave of layoffs has been especially concentrated among the large U.S. technology companies collectively known as FAANG (Meta/Facebook, Apple, Amazon, Netflix, Google) plus Microsoft. Programming departments have borne a disproportionate share of these cuts. From the perspective of investors and shareholders, these layoffs serve as a signal: companies that reduce their engineering headcount are perceived as seriously committed to AI-driven efficiency, regardless of whether the technical reality supports such reductions.

This article argues that, paradoxically, both the CEOs and the investors are wrong about the underlying economics — even though their actions may still produce mass unemployment. The factor that analysts consistently overlook is that the demand for software is not fixed. Rather, it behaves like what economists call an elastic demand: when software becomes cheaper to produce, we find uses for vastly more of it. Understanding this dynamic requires examining both the history of computing and the present reality of how software demand actually works.

The Historical Parallel: Lessons from Computing History

The most frequently cited cautionary tale about underestimating technology is the apocryphal quotation attributed to IBM Chairman Thomas J. Watson Sr. in 1943: “I think there is a world market for maybe five computers.” This quote appears in dozens of technology history books, in educational materials from PBS, and was on the Microsoft website at one point. It is regularly deployed in conference presentations and TED-style talks as evidence that even brilliant business leaders can spectacularly misjudge technological potential.

The problem is that Watson almost certainly never said it. IBM’s own historical documentation directly debunks the quotation. According to an IBM FAQ on company history, the statement appears to be “a misunderstanding of remarks made at IBM’s annual stockholders meeting on April 28, 1953” — a full decade later, and by Watson’s son, Thomas Watson Jr., not Watson Sr. (IBM Historical Archives, cited in GeekHistory, 2026). At that 1953 meeting, Watson Jr. was discussing the IBM 701 Electronic Data Processing Machine, the company’s first production computer designed for scientific calculations. He told stockholders that IBM had developed a paper plan for the machine and pitched it to approximately twenty potential customers. The machine rented for between $12,000 and $18,000 per month — an enormous sum at the time. Watson Jr. reported: “As a result of our trip, on which we expected to get orders for five machines, we came home with orders for 18.”

The mangled version — “I think there is a world market for maybe five computers” — appears to have emerged sometime after 1981, with no attested primary source from 1943 or any other year (Shapiro, 2008). The earliest known appearance was in a 1981 book titled Facts and Fallacies: A Book of Definitive Mistakes and Misguided Predictions. It is, in short, an urban legend that became accepted as fact through sheer repetition.

Yet the legend persists because it captures something genuinely true about technology economics. Even if Watson Sr. never made the statement, many informed observers in the 1940s and 1950s genuinely did believe that computers would remain rare and expensive specialty equipment. The IBM 701 was a room-sized machine costing millions in today’s dollars, requiring teams of specialists to operate. The idea that such machines would one day be embedded in automobiles, appliances, and pockets seemed fantastical.

The crucial economic insight, however, is not that experts underestimate technology — it is that the cost of technology determines how much of it we use. Had the cost of computer hardware remained at 1950s levels, we would indeed live in a world with far fewer computers. But hardware costs fell, and as they fell, we found uses for computing in virtually every aspect of modern life. The same dynamic applies to software: the cheaper it becomes to produce, the more software we will demand.

In more recent decades, it has been predicted multiple times that large amount of jobs would disappear. The current wave of CEOs is just the latest wave of that. For example, in 2013, Carl Benedikt Frey and Michael Osborne of Oxford University predicted that computerization would remove 47% of all jobs in the United States. Applying their method to Finland it was predicted that 1/3 of jobs would be be computerized while for Sweden it was 53%. As Anders Ekeland and Ingrid Krüger (2020) pointed out 7 years later, the predictions did not come true so far and that the analysis only focus on the aspect of new technologies and the work they could potentially replace rather than look at the entire picture of what was likely going to happen.

How Much Programming Is Really Needed?

The question of how much programming society truly needs has no simple answer. For a handful of very specific engineering problems, one can specify hard requirements. One might ask, for instance: how many sensors, how much CPU and GPU power, and how many lines of code does an autonomous vehicle need in order to drive more safely than a human driver in 99.99 percent of cases? Or: how much compute does a pacemaker require in order to function under all reasonable conditions during the expected lifetime of a human being? These questions have bounded answers determined by safety engineering.

But for the vast majority of software, there is no such fixed number. We can simply add more and more features to make software more convenient for users and to add functions that no one ever thought possible in the past. This elasticity of demand — the economist’s term for how much more of something people want when it becomes cheaper — is the key to understanding why AI-assisted programming is unlikely to eliminate programming jobs.

The concept has a name in economics: Jevons paradox. In 1865, the English economist William Stanley Jevons observed that as coal-powered steam engines became more efficient — requiring less coal per unit of work — Britain’s total coal consumption actually increased rather than decreased. Efficiency lowered the effective cost of using coal, which expanded its applications, deepened adoption, and increased total consumption (Jevons, 1865). The cheaper the resource, the more uses people find for it.

The same logic applies directly to software. As Jim Rutt has argued in a detailed analysis, AI coding tools do not merely substitute for developers — they lower the cost of turning an idea into working software. When the cost of production falls, demand rarely sits still. Projects once judged too small to justify a team suddenly become feasible. Internal tools that were postponed become worth building. Startups can attempt products that would previously have required twice the capital (Rutt, 2026).

Software has properties that make Jevons paradox apply even more forcefully than it did for coal. First, software is combinatorial: every new piece of software creates integration surfaces that demand more software. A company builds a customer database; now it needs analytics, reporting, API access, mobile views, compliance logging, backup systems, and migration tools. Second, software has near-zero marginal cost of distribution but near-infinite marginal cost of customization, meaning cheaper production creates entirely new categories of products tailored for smaller niches. Third, and most importantly, the physical world remains massively under-digitized (Rutt, 2026).

As Azeem Azhar has noted, the history of computing itself demonstrates this pattern repeatedly. Every previous drop in development cost — from assembly language to high-level languages, from mainframes to personal computers, from desktop to web, from web to mobile — triggered waves of new applications that nobody had anticipated (Azhar, 2025). There is no reason to believe that AI-assisted programming will be the exception that breaks this pattern.

The Growth in Demand: Two Illustrative Examples

The following two examples illustrate how software demand expands as programming becomes more accessible. Both are drawn from direct personal experience, but they represent patterns that are playing out across the economy.

The Rabbit Hotel Booking System

I serve as the only technically skilled person on the leadership board of a local 4H farm. One of the services we offer is a boarding facility for rabbits when their owners go on holiday, or for longer periods when owners cannot keep rabbits at home. Historically, bookings were managed entirely on paper. In order to rationalize operations, I was asked whether we could create an electronic booking system.

I began building a WordPress plugin for this purpose. The complexity of requirements, however, grew rapidly. Some cages can hold multiple rabbits — but rabbits should only be housed together if they are already bonded companions. During peak periods such as school holidays, we need to ensure that rabbits occupy no more space than absolutely necessary. Vaccination records must be tracked. Emergency contact numbers must be accessible. Dietary restrictions, medication schedules, and behavioral notes all needed to be incorporated.

Had this same project been proposed just four or five years ago, I would have told the organization that either the complexity must be dramatically reduced or we could not have an IT system at all. The implementation would simply have required too many hours. With AI-assisted development tools, however, the situation changed entirely. I could incorporate substantially more factors and spend less time building a rather complex application.

In the end, we shifted direction: rather than building a full booking system, we created a comprehensive rabbit registration database that holds all information for each rabbit and guinea pig in one place, while continuing to manage bookings manually so that volunteers can craft custom solutions for each animal. Even this change of direction was manageable — what would have been a project-ending setback five years ago became a minor pivot.

How well-written is the code? I have barely glanced at it. It is written in PHP, a language universally known to be painful to write, but it’s required because it is the language WordPress is written in. With something that comes down to being a relatively minor plugin (I can skip all advertisement, trackers, monetization strategy, etc.) it’s not that difficult to get the next LLM to write a modification for it if it turns out that I need something else. I don’t really need to know the code at all.

One might object that this is merely a hobby application, and hobbies can absorb unlimited time. But the underlying pattern — latent demand for software becoming economically viable as development costs fall — applies equally in commercial contexts, as the second example demonstrates.

Reimplementing Microsoft Word

In mid-2025, I was contacted by a small software group in the San Francisco Bay Area. They wanted to recreate all features of Microsoft Word. The official specification of the DOCX file format, known as Office Open XML or ISO/IEC 29500, runs to approximately 5,018 pages in its ISO-published form (ISO/IEC, 2012). The earlier ECMA-376 version grew to approximately 6,000 pages during standardization, and by some estimates the full specification including extensions approaches 7,000 pages (Document Foundation, 2026). For comparison, the competing OpenDocument Format (ODF) achieves the same general purpose in roughly 867 pages (Google, cited during OOXML standardization, 2008).

Creating a word processor that supports the full DOCX specification is a task so enormous that no organization has ever truly attempted it. Even Microsoft’s own online Word processor, Word for the Web, does not render documents identically to its desktop Microsoft Word application. The specification includes not only document structure but detailed descriptions of legacy behaviors, undocumented rendering quirks, and backward-compatibility requirements that essentially encode decades of Microsoft Office development history (Rob Weir, cited in GeekHistory, 2006).

Microsoft has been accused by multiple commentators of making the specification deliberately long and complex in order to maintain competitive moats. Google stated during the ISO standardization process that “the ODF standard, which achieves the same goal, is only 867 pages” and calculated that if ISO were to give OOXML with its 6,546 pages the same level of review that other standards receive, “it would take 18 years to achieve comparable levels of review” (Google, 2008). The Free Software Foundation Europe has characterized OOXML as a “pseudo-standard” that is “compatible only with Microsoft documents, whilst its interoperability with products from other vendors appears to have been made extremely complicated by design” (FSFE, 2024).

Nevertheless, this small Bay Area startup was able to prototype their project with the assistance of AI coding tools, and then hire developers like myself in other parts of the world to perform the actual implementation work. I worked with them for nine months on this effort. The project would have been economically unthinkable without AI assistance: no venture capital firm would have funded a team to reimplement Microsoft Word from scratch using traditional development costs.

This example illustrates a darker side of the demand expansion as well. Now that Microsoft developers themselves can program faster with AI assistance, there is little preventing the company from extending the DOCX specification further — adding features, increasing complexity, and raising the barrier to entry for potential competitors. The specification of the next DOCX format revision could conceivably balloon to 50,000 pages or more, each new feature designed not for user benefit but as an arms-race weapon against any competitor seeking to achieve full interoperability.

Why You Could Still Be Laid Off

If the demand for software is genuinely elastic — if cheaper development simply leads to more software being created — then why are programmers being laid off in large numbers? The only logical answer I can come up with is that the answer lies not in technology, but rather in the political economy.

Much of the programming world, particularly web development, is centered around the United States, where shareholder value is the legally and culturally dominant corporate priority. The influential investor class, dominated by the wealthier segments of the baby boomer generation, has been told by AI company CEOs that large language models can replace software engineers. These investors expect technology companies to demonstrate commitment to AI by reducing their engineering headcount, regardless of whether such reductions make technical sense.

The result is a form of performative layoffs — job cuts designed to signal corporate strategy to financial markets rather than to improve technical output. Companies that fail to reduce their programming staff risk being perceived as insufficiently committed to AI, potentially depressing their stock prices and reducing their access to capital. The technical reality — that their remaining engineers are now expected to produce the work that previously required significantly more people — is secondary to the market signal.

The consequences are visible in the quality of software infrastructure. As a web developer with extensive interaction with browser development teams, I have observed that browser development appears to be slowing as companies reduce their engineering investment.

In April 2025, Google laid off hundreds of employees from its Platforms and Devices division — the unit that includes the Chrome browser, Android, and Pixel teams — explicitly redirecting resources toward AI investments (PCMag, 2025). A former Chrome product manager told The HR Digest (2025): ‘Middle managers like me were told we’d have chances to pivot internally, but the reality is tougher.’ Microsoft has made similar cuts: its Web Experience Team, which builds the Edge browser, was hit by layoffs in 2023 (Business Insider, 2023), and by May 2025 the company had cut thousands more software engineers across Windows and cloud divisions (CRN, 2025). In July 2026, Microsoft announced another 4,800 layoffs (GeekWire, 2026). Even Apple, which has avoided headline layoffs, has drawn developer ire for chronically under-resourcing Safari — updates ship roughly twice yearly compared to Chrome’s monthly cycle, and web engineers widely describe Safari as ‘the new IE’ for its failure to implement modern standards. Developers have told Apple’s own evangelism team that WebKit has been ‘understaffed/not a priority for Apple’ — a concern the Safari team itself appeared to acknowledge at their 2026 Worldwide Developers Conference when they admitted that ‘there are more good ideas than people to build them.’ (WWDC, 2026)

The result is a growing accumulation of unfixed bugs and unimplemented features in the web platform that developers depend upon. There is a certain irony in the situation: the large language models that web developers now use to work around these browser bugs are partly necessary because the companies building the browsers have fired the people who would have fixed the bugs in the first place.

What Happened to Competition?

If the United States is pursuing a strategy of performative layoffs that undermines its technical infrastructure, one might expect other regions — China and Europe in particular — to capitalize on the opportunity. To some extent, this is happening, but the picture is more complex than a simple competitive displacement.

China’s Alternative Path

China has taken a notably different approach to the relationship between AI adoption and employment. In May 2026, the Hangzhou Intermediate People’s Court ruled that a technology company had illegally fired a quality assurance worker after an AI system automated his job. The court determined that the company’s grounds for termination “did not fall under negative circumstances such as business downsizing or operational difficulties, nor did they meet the legal condition that made it impossible to continue the employment contract” (Fortune, 2026). The employee, identified only as Zhou, had earned an annual salary of approximately 300,000 yuan ($43,900) before being demoted and ultimately terminated after refusing a 40 percent pay cut.

This ruling built on a precedent set by a Beijing arbitration panel in December 2025, which found that a mapping company could not legally terminate an employee solely because AI had replaced his data collection work. The panel ruled that the company’s adoption of AI was a voluntary business choice rather than an uncontrollable event, and that by citing AI replacement as grounds for dismissal, the company had “shifted the cost of the technological transformation to the employee” (NPR, 2026).

Chinese Vice Premier He Lifeng reportedly asked the country’s largest employers in summer 2025 how AI would affect their workforces. Some companies responded that full AI implementation could eliminate 30 percent or more of existing roles, according to people familiar with the discussions (Wall Street Journal, 2026). Yet the Chinese government appears to be prioritizing labor market stability over rapid workforce displacement, even as it pursues state-directed AI development goals.

Chinese technology companies are also pursuing a different competitive strategy. Chinese automakers including BYD, Geely (which owns Volvo), and XPENG are successfully building software-defined vehicles and continuing to maintain and upgrade the software even after selling cars to customers. BYD delivered approximately 200 over-the-air (OTA) software updates in 2025 alone — the highest number among automakers globally — compared to 16 for Tesla in China, 8 for Toyota, and just 5 for Volkswagen (News18a, 2026). XPENG provides OTA updates on a near-monthly basis for its software-defined vehicle architecture.

The stock performance of these Chinese companies suggests that shareholders are not the only priority — or at least not the short-term priority — for corporate decision-makers. The Chinese approach treats software development as a long-term capability to be maintained and expanded, rather than as a cost center to be minimized for quarterly earnings reports.

Many aspects of China’s and the United States’ work culture of working from from 9-to-9, 6 days a week (996) in the tech sector combined with little safety and regional high unemployment make Europe look like the better option starting at age 30. Nevertheless, right now it looks like China is the one prioritizing workers over share holders when it comes to AI.

Europe’s Struggles

Unfortunately, Europe’s response has been less decisive when it comes to dealing with the new realities of AI. Let’s start with the positive parts.

The European Union has adopted an ambitious Open Source Software Strategy with 70 concrete proposals, including a European Open Source Sovereignty Fund, public procurement policies favoring open-source solutions, and investment in digital infrastructure (European Commission, 2026). The strategy covers the full lifecycle from research and development through market uptake and long-term maintenance of critical open-source components.

Working remotely for U.S. companies has become less economically attractive for European software developers as the U.S. technology sector contracts and currency fluctuations reduce the value of dollar-denominated salaries. Some European developers are shifting to local employment or entrepreneurship.

However, in practice, large-scale software procurement in Europe continues to follow a model that mirrors the American approach rather than establishing a genuine alternative. Governments remain the largest software purchasers, and they typically award massive contracts to private companies with extensive management hierarchacies. These companies produce proprietary “products” that governments pay to develop and then continue paying subscription fees to use — even when the resulting systems fail.

The most striking example is Sweden’s disastrous attempt to implement Oracle Health’s Millennium electronic health record system. The system was purchased in 2018 to replace dozens of different IT systems in the southern Swedish health districts of Västra Götaland and Region Skåne.

An external audit by KPMG found that the organization was not adequately prepared for the introduction and that there was a lack of leadership (Heise, 2025). Many of the issues mentioned in the local media after it failed — such as whether Swedish healthcare data could be exported to India or the United States for AI evaluation — could have been answered easily by most knowledgeable technologist, such as myself, but decisions were left to high level officials in government/company without using any of the local expertise.

When the system went live in Västra Götaland in mid-November, it had to be shut down after just three days due to massive protests from medical staff, who felt that patient safety was no longer guaranteed. One doctor described the new system in the Swedish daily Goteborgsposten as a “hamster wheel of hell.” As basic functions failed to work, employees in some departments were forced to resort to pen and paper (Heise, 2025).

The costs have been staggering. Region Skåne invested over 270 million euros in the Millennium system. Region Västra Götaland invested over 335 million euros. The total exceeds 600 million euros for a system that, in Västra Götaland’s case, operated for three days before being abandoned. In May 2026, Sweden’s social affairs minister Jakob Forssmed announced a formal government investigation into the procurement of both Millennium and Cambio’s Cosmic systems, noting that “if you look around in Sweden, you can see that this has not been [the technological leap forward that was promised]” (Euractiv, 2026).

The Swedish experience illustrates the core problem: even when governments invest enormous sums in software, the contracting model — massive payments to large vendors for proprietary systems from the United States — produces failure after failure. A more rational approach would involve directly employing open-source developers to create modular systems that can be reused across government agencies without each region making the same catastrophic investment. It would also increase the experience and knowledge among local developers. The technology and the developers exists; the political will to change procurement models does not.

So What Should Be Done?

Once one recognizes that we are rarely working close to any natural limit on how much software society needs, and that most software demand depends on how difficult or expensive it is to write, the fear of AI taking programming jobs begins to seem misplaced. Programmers continue to spend 40 or more hours weekly in front of their screens even though individual productivity has risen dramatically. The relevant question is not whether AI will eliminate programming work, but how society will distribute the gains from that increased productivity.

There are essentially two paths. The first is what we appear to be following: programmers produce the output that once required ten people, most of it goes to waste in meaningless competition between large companies, some firms lay off staff to impress shareholders while losing the capability to handle bug reports, and individual workers see no benefit from their own increased efficiency.

The second path involves making deliberate choices about how to use productivity gains. Some software development genuinely serves human needs: medical systems that improve patient care, educational tools that expand access to learning, infrastructure that reduces environmental harm. If AI-assisted programming makes such projects economically viable for the first time, that is a genuine social benefit worth pursuing.

Beyond that, this should be the moment to reduce working hours. The evidence from multiple countries supports this direction. Sweden’s most recent four-day workweek pilot, conducted in 2024-2025 with eleven organizations including care homes, social services, energy providers, and NGOs, found that measured stress levels fell by 19 percent, sleep problems were halved, and participants slept an average of 24 minutes longer per night. Perceived productivity rose by 13 percent and work capacity increased by 10 percent. Ten of the eleven participating organizations chose to retain the shorter working hours (The Better News, 2026).

These findings are consistent with international research. Between 2015 and 2019, Iceland conducted one of the largest global trials of reduced working hours, involving over 2,500 workers. Productivity remained the same or improved in the majority of workplaces, while worker well-being increased substantially. In the United Kingdom, a 2023 pilot study with over 600 companies found that 93 percent of participating organizations chose to continue the four-day week (PaceLab, 2026).

The technology sector, of all industries, should be capable of leading this transition. Software engineering is knowledge work that depends on concentration, creativity, and problem-solving — all capacities that deteriorate with exhaustion and improve with rest. A programmer working 32 hours per week with AI assistance can likely produce more valuable output than the same programmer working 40 hours without assistance, while enjoying substantially better quality of life.

What stands in the way is not technical feasibility but political choice. The current system prioritizes shareholder returns over worker welfare, quarterly earnings over long-term capability, and performative cost-cutting over genuine efficiency. Changing this requires not better technology but better policy: labor protections against AI-driven dismissal like those emerging in China, procurement reform that directs public investment toward reusable open-source systems rather than proprietary failures, and a broader social movement to reclaim productivity gains for working people.

AI will not take your coding job. But the economic system that deploys AI might leave you unemployed anyway — unless we choose a different path.

I would like to thank Anders Ekeland for commenting on an earlier draft of this article, as well as several tech workers who contributed to my understanding of the field and who must remain anonymous.