In his column <em>Job Machines</em> in NRC, physicist Robbert Dijkgraaf describes (opent in nieuw venster) how technological breakthroughs follow the same pattern time and again. The steam engine did not make coal obsolete — coal consumption exploded. The spreadsheet did not make the bookkeeper unemployed — demand for financial analysis multiplied. The digital camera did not end photography — more photos are now taken in one minute than in the entire nineteenth century. Dijkgraaf summarises it with the paradox of British economist William Stanley Jevons (opent in nieuw venster) from 1865: when technology makes something more efficient and cheaper, usage increases so explosively that total demand grows rather than falls.
Dijkgraaf (opent in nieuw venster) draws the line through to AI and radiologists. But one important example he leaves unexamined: the programmer. Because while the world is busily debating whether AI will make the programmer redundant, something very different is happening on the shop floor. Backlogs are overflowing. Vacancies remain open for months. And the programmers who are there are not using AI to automate themselves out of a job, but to finally get to work that has been sitting on the shelf for years.
This is not a theoretical argument. This is the reality of everyone who has software built in 2026. And that reality shows something different from what the doom scenarios suggest.
The myth of the empty backlog
Let us start with the facts. The global shortage of software developers has, according to IDC, risen from 1.4 million in 2021 to around 4 million in 2025 (opent in nieuw venster). The U.S. Bureau of Labor Statistics predicts growth of 17 per cent in demand for software developers between 2023 and 2033 (opent in nieuw venster) — well over twice the average for all occupations in the United States. And these are not abstract projections: 87 per cent of companies report that they are already experiencing a shortage of development capacity. (opent in nieuw venster)
Anyone who works daily with software teams will recognise this immediately. The backlog — the list of desired features, improvements, integrations and new products — is not a neat task list at most organisations. It is a graveyard of good ideas that were never realised. Not because they had no value, but simply because there were not enough hands to build them.
At Alpina Group, where I am responsible for the online labels, I know this first-hand. We manage more than thirty calculation funnels for four online insurance brands. Behind each of those funnels lies a list of improvements that would improve customer satisfaction, increase conversion, speed up processes. And behind that lies yet another list of entirely new products, features and services that we would like to build. The capacity to realise all of that has never been there. Until now.
AI as a steam engine for the backlog
What happens when you equip a programmer with AI tools such as GitHub Copilot, Claude or Cursor? Exactly what Jevons predicted. The programmer does not become redundant — the programmer becomes more efficient. And that efficiency creates room for more work, not for fewer workers.
The figures speak for themselves. GitHub Copilot now has more than 20 million users (opent in nieuw venster) and has been adopted by 90 per cent of Fortune 100 companies. Developers complete tasks up to 55 per cent faster (opent in nieuw venster). The turnaround time for pull requests — the moment when code is reviewed and merged by colleagues — fell from an average of 9.6 days to 2.4 days (opent in nieuw venster). And some 46 per cent of all code in projects using Copilot is now generated by AI.
But here is the crucial point: companies that introduce these tools are not laying off their developers. They are expanding their teams. As the CEO of GitHub put it: AI accelerates productivity and thereby creates more backlog, not less. Teams that can deliver faster receive more requests. Projects that were previously 'too expensive' or 'too complex' suddenly become feasible. And so demand grows.
This is Jevons in its purest form. The steam engine did not make coal obsolete — it made coal so useful that the world wanted infinitely more of it. AI does not make programmers redundant — it makes software (opent in nieuw venster) so accessible that the world wants infinitely more of it.
The counterargument: the last generation of programmers
I would not be writing an honest argument if I did not take the counterargument seriously. Because it is powerful, and it deserves attention.
The pessimists point out that the Jevons paradox works fundamentally differently for programming than for coal or textiles. In previous technological breakthroughs, human labour remained necessary to realise the extra production. Someone had to operate the factory, sew the garments, evaluate the X-ray scan. But in programming — so the argument goes — the AI that makes code more efficient is the same AI that can also produce the extra code. There would be no bottleneck that makes human programmers indispensable.
This argument is reinforced by the chess analogy that Dijkgraaf describes in his column. The combination of human and machine — the centaur — was for a long time stronger than the machine alone. But in 2017, AlphaZero demonstrated that the machine could manage perfectly well without a human head. Everywhere, the pessimists claim, the pattern shifts: first the machine enhances the human, then it replaces them. The programmer is now in the centaur phase, working with AI as copilot. But that phase is a transitional phase, not a final destination.
Salesforce announced at the start of 2025 that it would no longer hire new programmers due to the productivity gains from AI. The message was loud and clear: AI can do the work. Why would you hire people?
Why the pessimistic scenario is wrong — at least for the next five years
The pessimistic argument is well constructed, but it makes three crucial errors.
1. Software is not a chess game
The comparison with chess breaks down the moment you take the reality of software development seriously. Chess has 64 squares, fixed rules and an objective winner. Software lives in the messy reality of human needs, changing legislation, organisational politics, legacy systems and ethical considerations.
Can AI generate code? Absolutely. AI can even understand regulations, apply UX principles and analyse customer profiles — these are tasks where AI is surprisingly good. But building software is more than the sum of its parts. It is the director who decides on Monday morning that the strategy has changed and that the system that just went live now needs to work differently. It is the customer who calls with a claim that fits no process diagram, but for which a solution must nonetheless be found. It is the moment when two systems that both work perfectly produce unpredictable behaviour together because nobody anticipated they would run simultaneously.
These are not edge cases. This is the daily work of software development. And it is precisely the kind of work that AI is furthest from.
2. The bottleneck shifts, but does not disappear
The idea that AI can deliver both efficiency and extra production sounds logical, but ignores the reality of software development in 2026. AI-generated code must be reviewed, tested, secured, integrated, maintained and explained to stakeholders. Research shows that 46 per cent of developers do not fully trust AI output (opent in nieuw venster), and that longer debugging times sometimes completely negate the expected speed gains.
The bottleneck shifts from typing code to reviewing code. From writing to orchestrating. But that shift does not make the human developer redundant — it changes what the human developer does. And for that work — architecture, quality assurance, system integration, security audits — not less but more expertise is required.
BCG and Deloitte confirm this: 74 per cent of companies struggle to scale AI value (opent in nieuw venster), not due to technical limitations, but due to people and processes. Human capacity for specification, evaluation and integration has become the binding constraint. Not the capacity to generate code.
3. The Salesforce argument is an exception, not a trend
When one large technology company announces it will no longer hire programmers, that makes the headlines. But one swallow does not make a summer. Salesforce is a mature SaaS company with a rounded product portfolio. Their situation is not representative of the thousands of scale-ups, mid-sized companies, government agencies and healthcare institutions that are crying out for development capacity.
The broader data tells a different story. The global market for custom software development is growing at 23 per cent per year. The total software market was worth $570 billion in 2025 (opent in nieuw venster) and is on course for more than $1 trillion by 2030. The U.S. Bureau of Labor Statistics projects growth of 15 per cent in software developer jobs by 2034. These are not figures that fit a profession on the verge of disappearing.
The insatiable appetite for software
Dijkgraaf describes in his column how intelligence is a 'meta-commodity' that is used in everything, with no natural upper limit on demand. The same applies to software, and perhaps even more so.
Today perhaps five per cent of all potential software projects have seen the light of day. Every small business would like a bespoke customer management system. Every school would like its own educational apps. Every GP practice has specific workflows that are crying out for digitalisation. Every insurer would like to build a streamlined online sales process for every niche product. It has not happened because building software (opent in nieuw venster) was too expensive, too complex and too time-consuming.
AI is fundamentally changing that economy. When a team that previously needed three months for a project can now deliver that same project in six weeks, the demand for that team does not decrease. The number of projects that management approves doubles. The backlog does not shrink — throughput speed increases, and with it the level of ambition.
The democratisation of software — and why that requires more programmers
There is yet another dimension that the pessimists overlook. AI not only lowers the production cost of software — it lowers the threshold for wanting software. Low-code platforms and AI assistants make it possible for non-technical professionals to build simple applications. Gartner predicts that the low-code market will be worth nearly $45 billion in 2026. By 2026, 80 per cent of low-code tool users will be outside IT departments.
At first glance, this seems bad news for programmers. If everyone can build, who needs a professional? But history teaches the opposite. When WordPress democratised the building of websites, web developers did not disappear. There were more of them. Because millions of people discovered that having a website was valuable, and then discovered that a good website requires more than a template.
The same will happen with AI-generated software. Companies that build a first version with low-code tools quickly discover that they need scalability, integration with existing systems, security that complies with the AVG (GDPR), and performance that can handle their growing customer base. At that point they call a professional developer. Democratisation creates demand — it does not destroy it.
The centaur phase is not a transitional phase
Dijkgraaf describes the phenomenon of the centaur — the combination of human and machine — as a potentially temporary stage. The pessimists point out that AlphaZero ultimately no longer needed the human chess player at all. But software is not a closed system with fixed rules. Software is a living artefact that must constantly be adapted to a changing world.
New European regulation such as the AI Act and the Digital Services Act require continuous adjustments to existing systems. Cyber threats evolve daily. Customer needs shift with generations. Corporate mergers require system integrations. And the human context — knowledge of a specific domain, a specific organisation, a specific customer — is precisely what AI lacks.
Dijkgraaf's radiologist did not disappear when the CT scanner arrived. The accountant did not disappear when the spreadsheet arrived. The photographer did not disappear when the smartphone arrived. In all these cases, the profession grew because the technology made the field more accessible and relevant. The programmer is next in this line.
What does change is the nature of the work. The programmer of 2030 will type fewer lines of code and spend more time on architecture, quality assurance, domain knowledge and user experience. The profession shifts from craft to direction. But direction requires deep understanding of the field — you cannot conduct an orchestra if you cannot read a note. And precisely that deep understanding is scarce.
The Dutch perspective: digitalisation is far from complete
For those who think the shortage of programmers is an American problem: three-quarters of European companies struggle to find IT talent. In the EU, a shortage of more than eight million ICT professionals threatens by 2030. In the Netherlands specifically, we see that the digital transformation of sectors such as healthcare, government, education and financial services is far from complete.
The Dutch insurance industry, the sector I know best, is illustrative. Many processes still run on legacy systems that are decades old. Customers expect a seamless digital experience like Coolblue or Bol, but receive forms reminiscent of 2005. Regulation is becoming increasingly complex: Solvency II, DORA, the AI Act, Wft (Financial Supervision Act) amendments — each with its own digital implications. And meanwhile expectations are growing: real-time claims handling, personalised premium calculation, chatbots that actually help.
All of this work requires developers. Developers who have the domain knowledge to understand insurance logic. Developers who know the regulations. Developers who can link legacy systems to modern APIs. AI can help them work faster, but AI cannot replace them — because AI does not understand why Article 7:932 of the Dutch Civil Code stipulates that an insurer must provide cover, unless contributory negligence applies. That kind of knowledge lives in people's heads, not in models.
Five years ahead: more programmers, different work
Let me make a concrete prediction. Over the next five years — until 2031 — the total number of working programmers worldwide will not fall. It will rise. Not because AI fails, but because AI succeeds.
The reasoning is as follows. AI makes software development cheaper. Cheaper software lowers the threshold for companies to start digital projects. More projects lead to more demand for people who can guide, review, secure and maintain those projects. Meanwhile, existing backlogs are being worked through more rapidly, creating space for projects that had been lying dormant for years. And behind those wait the projects that nobody ever dared to dream of, but which have suddenly become feasible thanks to AI.
Atlassian's internal data shows that developers spend only 16 per cent of their working week on actual coding. The rest goes to coordination, debugging, testing, deployment and documentation — precisely the areas where AI is now providing acceleration. But that acceleration does not eliminate those steps. It multiplies the number of cycles that teams can run. And more cycles means more ambition, more projects, more work.
LinkedIn's Emerging Jobs Report 2025 confirms this: demand for AI-proficient software engineers has grown by nearly 60 per cent per year. There is a salary premium of 15 to 25 per cent for developers who are proficient in AI frameworks and orchestration. These are not the signals of a dying profession. These are the signals of a profession in transformation.
What does change — and how we must prepare
I do not want to give the impression that nothing is changing. An enormous amount is changing. The programmer of 2031 is a fundamentally different kind of professional than the programmer of 2021. Let me be honest about what is shifting.
Firstly: entry becomes harder and easier at the same time. Easier, because AI tools lower the learning threshold and make junior developers productive more quickly. Harder, because companies are increasingly turning to AI for 'simple' coding work, meaning the entry-level positions where you previously learnt the trade are under pressure. This is a genuine problem. If we are not careful, a gap will emerge between the AI-proficient senior developer and the junior who never gets the chance to become senior.
Secondly: the centre of gravity of the profession is shifting. From writing code to designing systems. From knowing syntax to understanding domains. From individual programming to orchestrating and validating AI output. The programmer becomes, in Dijkgraaf's words, a centaur — and that requires different competencies from what most educational programmes currently teach.
Thirdly: productivity gains are not evenly distributed. Senior developers benefit more from AI tools than juniors, because they are better able to assess whether AI output is correct and safe. This reinforces the existing scarcity of experienced talent. The 'structural polarisation' that analysts are signalling — more juniors applying, fewer seniors available — is reinforced rather than resolved by AI.
A nuance that the doomsayers miss
The debate about AI and programmers is too often conducted in absolute terms. 'AI replaces programmers' versus 'programmers are irreplaceable'. The reality is more nuanced and more interesting than either position suggests.
The Jevons Paradox teaches us that efficiency and demand go hand in hand. But Jevons also teaches us that the nature of work changes with every technological leap. The miner of 1900 had a different profession than the miner of 1800. The photographer of 2024 has a different profession than the photographer of 1970. And the programmer of 2031 will have a different profession than the programmer of 2021.
What remains constant is this: as long as there are human problems requiring a digital solution, as long as there is a gap between what organisations want and what their systems can do, as long as laws change and markets shift and customers become more demanding — there will be work for people who understand how to use technology to solve those problems.
We call those people programmers today. In ten years we may call them AI architects, systems designers or digital engineers. But the work does not disappear. It transforms. And in that transformation, for the next five years, there is no contraction but growth.
The real lesson of Jevons
Dijkgraaf ends his column with the observation that there will always be something left that only a human can do: judgement, the difficult conversation, taking responsibility. I would add one thing to that: imagination. The ability to see what does not yet exist, and to envisage how it could be.
That imagination is precisely what lies behind every backlog. Behind every wish that was 'too expensive'. Behind every idea that was left on the shelf. AI now makes it possible to realise that imagination — faster, cheaper, more accessible than ever. But the imagination itself, the understanding of the problem, the translation of a human need into an elegant digital solution — that remains human work.
The golden age of the programmer is not over. It is only just beginning. Because the Jevons Paradox teaches us one thing above all: the more we can, the more we want. And who translates that 'wanting' into working systems? Those will be programmers. More than ever. And for at least the next five years that will not change — not because technology cannot manage it, but because the world is simply not yet ready to do without them.