There was a decade, not so long ago, when “learn to code” functioned almost like scripture. Governments built policy around it. Parents who had never touched a keyboard encouraged their children toward it. Across Accra, Lagos, Nairobi and beyond, a generation was told that the surest route out of economic precarity ran through a text editor and a semicolon.
Bootcamps multiplied. Scholarships followed the same script: teach the continent to code, and the continent would code its way into the future.
That promise was not wrong, exactly. It was incomplete. What nobody fully priced in was that the skill itself would be partially automated before the generation trained to supply it had finished graduating.
The tool ate part of the trade
The shift is no longer speculative; it shows up in the numbers. Big technology companies have said openly that a significant share of their own code is now machine-generated. Microsoft and Google have each put the figure at roughly a quarter to a third of new code, depending on the team and the year, and some firms report figures approaching half.
GitHub Copilot alone had grown to several million paying subscribers by early 2026, and platforms such as Cursor, Lovable and Replit have made it possible for someone with no formal training to describe an idea in plain language and watch a working website or app take shape within minutes, a practice now casually called “vibe coding”.
The consequence for the entry point into the profession has been blunt. A Stanford Digital Economy Lab study, led by economist Erik Brynjolfsson, found that employment for early-career workers in AI-exposed fields, software development among them, had fallen by roughly 13% over three years, concentrated heavily among workers in their early twenties, even as employment for experienced practitioners in the same fields held steady or grew.
Separate labour-market analyses in 2026 have pointed in the same direction, with entry-level developer job postings down sharply from their 2023 peak in several trackers, computer science graduate unemployment sitting around 6 to 7% in some US data, and hiring managers admitting, in plainer language than usual, that the traditional first rung of the ladder, junior developers writing boilerplate code, fixing small bugs, and doing the unglamorous grunt work that used to double as an apprenticeship, has largely been absorbed by the machine.
This is worth sitting with, because it inverts the promise. Coding was sold as a hedge against automation. It has become one of automation’s more visible early targets, not the writing of software as a whole, but the specific rung most young, first-time entrants used to stand on to get in.
What has not disappeared
And yet the wider picture refuses to collapse into simple decline. The US Bureau of Labor Statistics continues to project double-digit growth in software development employment through the early 2030s, adding hundreds of thousands of jobs. Global developer headcount, tracked by Evans Data Corporation, reached a record high in 2026.
Demand for engineers who can supervise AI output, integrate large language models into products, manage AI agents, or take responsibility for architecture and judgement calls a model cannot safely make has grown fast enough that some analysts describe a severe shortage rather than a glut. Stack Overflow’s 2026 developer survey listed “AI integration engineer” among its fastest-growing job titles.
The pattern resembles previous technological transitions more than it resembles a straightforward extinction. When high-level programming languages replaced assembly code, or when cloud computing replaced server rooms, the total demand for people who understood software did not vanish, it moved.
What changed was what counted as valuable labour. The same appears to be happening now, compressed into a much shorter timeframe: the premium is shifting away from the ability to produce code and toward the ability to know what code should do, whether it is doing it correctly, and what happens when it is not.
That last point matters more than it sounds. Researchers at MIT’s Computer Science and Artificial Intelligence Laboratory have flagged that AI-generated code can look entirely plausible while quietly failing to do what it claims to do, a problem that becomes more dangerous, not less, as more of the code in circulation is written by systems that cannot be held accountable for their errors. Someone still has to catch that. For now, that someone is a trained human.
The African variable
For a Ghanaian or wider West African reader, this story does not play out identically to the one unfolding in Silicon Valley. The continent entered the coding boom later and from a different starting position. Ghana had an estimated 18,000 professional developers as of recent counts, a fraction of South Africa’s or Egypt’s numbers, and bootcamps such as Codetrain, ALX, and Moringa School, and government-backed efforts like Nigeria’s 3MTT programme were still in the process of building that base when the ground shifted under them.
There is a case that this lateness is not a pure disadvantage. Africa’s tech training ecosystem is younger and smaller than the West’s, which means it carries less institutional weight to defend and can, in principle, retool faster around what the market now actually wants.
AI-native development, not pure syntax memorisation. Programmes describing themselves as training “AI-native” engineers rather than conventional bootcamp graduates have already begun appearing in Lesotho, Botswana and elsewhere. Ghana’s own national AI strategy, still maturing relative to more developed frameworks in South Africa, has at least named talent development and a Responsible AI Office among its stated priorities, even if the distance between the announcement and the operational reality remains wide.
There is also a less comfortable case: that a continent which invested heavily, often at real personal cost to students and families, in exactly the skill now being partially automated risks having built its on-ramp just as the ramp was being redesigned. Both things can be true. Neither cancels the other out.
So, what is the future?
If there is an honest answer, it is not a single skill this time. It is a posture. The people who described the coding boom as the future were not wrong that literacy in how software works would matter. They were wrong, or at least premature, in treating the act of writing syntax as the durable part of that literacy.
The durable part turns out to be closer to what good editors, good engineers and good journalists have always needed: judgement about what is true, what is good enough, what is dangerous if unchecked, and what a tool’s confident-sounding output is quietly getting wrong.
That reframes the question a young person in Accra or Kumasi should be asking. Not “Should I learn to code?” but “What do I want to be able to judge?” A developer who can evaluate whether an AI-generated authentication system is secure is more valuable than one who can type it out from memory.
A journalist, policy analyst, or entrepreneur who understands enough about how these systems work to know where they lie, hallucinate, or quietly fail is more valuable than one who either distrusts the tools entirely or defers to them completely.
The uncomfortable truth in all of this is that nobody, not the companies building these tools, not the economists studying the labour data, not the governments drafting AI strategies, can say with confidence where the floor is.
The data so far describes a redistribution of opportunity, not (yet) its disappearance. Whether that holds is an open question, and open questions are, by definition, not things anyone can responsibly promise an answer to.
What can be said is this: the certainty that made “learn to code” such a persuasive slogan a decade ago was always slightly false advertising. Technology does not hand out permanent hedges. It hands out temporary advantages to whoever adapts fastest to what it has just made obsolete and takes them away just as quickly once someone builds a tool that automates the adapting, too.
The future, on current evidence, belongs less to people who can code and more to people who can think clearly about what code and the machines now writing it should be trusted to do.
