For the past few years, the standard advice I've heard given to educated African youth has been roughly: "Learn to code, or get left behind." I did buy into it partly because I've watched brilliant young students and professionals freeze at words like "Python," "Rust", and partly, to be honest, because I have myself done some freezing of my own. However, in the past two years, I have become uncomfortable with that advice. It increasingly appears we have been assigning the wrong homework to our students. Or the wrong after-work coding tutorials.
Anthropic recently looked at around 400,000 real sessions of people working alongside its AI coding agent, between October 2025 and April 2026. Tucked beneath the charts is a finding I keep turning over: when the AI does the actual building, how well you can code barely predicts whether you succeed. What predicts success is whether you understand the problem well enough to recognize a good answer, to notice when the machine has confidently handed you a bad one.
A few numbers stuck with me. People made about 70% of the decisions about what to build; the AI handled roughly 80% of the how. The people who knew their field deeply were more than twice as likely to end up with something that actually worked, and far less likely to quit in frustration. Give a tool like Codex to a beginner and it does a little; give it an expert's instructions, and that very same tool does several times more. The tool didn't change. The person holding it did.
Here's the part I find quietly thrilling. The hard part of work used to be making the thing function – the syntax, the manual execution, the doing. That part is being automated away. What's left, and what's now scarce, is judgment: knowing what's worth building, and whether what came back is any good. It's the difference between a thesis adviser and the student running the regressions. The adviser rarely touches the numbers, but the whole project sinks if you can't tell a sound argument from a shaky one. And Anthropic carefully noted that coding is simply the first field to show this. Law, finance, medicine, and writing are likely next in line.
I've been running this experiment for years
A few years back, I was teaching an interdisciplinary course at the University of Delaware's Biden School (Energy and Environmental Policy), the kind of class where an engineering reading sits next to a philosophy one and somebody is always slightly out of their depth (often, gloriously, everybody). When ChatGPT arrived around the start of one semester, I did something that, as far as I'm aware, none of my colleagues were trying yet: I encouraged my liberal-arts students to use AI to get through the technical readings, not around them.
I half-expected it to backfire. It didn't. Students who used to vanish during the equation-heavy sessions started turning up with real questions, and their essays got sharper. They weren't outsourcing the thinking. They were, finally, getting a foothold in material that had quietly shut them out for years.
I’m still at it – these days much closer to the work I care about most. With my friend and colleague, Prof. Ogheneruona Diemuodeke (Ruona, to those who know him), I teach an online module on AI, energy, and sustainable development at the University of Port Harcourt’s Energy Technology Institute. This past March, the two of us ran a three-day intensive there on using AI to plan for universal energy access in Nigeria, and on day one, we put it to the 25 energy planners and policy professionals in the room bluntly: you’re not programmers, you’re code pilots – your job is to know what you want; the AI knows the syntax. (I wrote about it here.)
Then we watched energy economists who had never written a line of code direct ChatGPT to reshape real geospatial analyses of Nigerian electrification data. This was happening live, in the room. The leverage came from domain knowledge, theirs and ours. The people who soared were the ones who understood the energy problem in front of them; the code was never the hard part. That room proved something I was beginning to understand: a domain expert can become a powerful AI user in days, not months. The Anthropic study simply put a number on what Delaware, and now Port Harcourt, had already shown me.
The skill worth teaching
So if the advantage now belongs to people who understand their field and can steer AI well, the most useful thing we can hand a learner isn't a programming language. It's the habit of building understanding on purpose. Intentionally. The research even tells us where to aim: the biggest leap happens when someone moves from beginner to merely competent. After that, expertise adds surprisingly little. That early climb is exactly where a good teacher, or a good program, earns its keep.
That conviction is why we built the AI Learner's Journal Kit. The idea is almost embarrassingly simple: after each learning session, a student writes down what they set out to learn, what they tried, how they used AI, where they got stuck, and what they'll do next. That small, deliberate act of reflection is how a beginner becomes competent. It is also how they build the instinct to sense when the AI is leading them astray. It works just as well for a teenager learning to code as for an aspiring energy planner, because the underlying skill is the same.
There's an equity angle here I can't walk past with my sustainable development hat on. Left to its own logic, AI tends to widen gaps: it rewards the people who already have the time, devices, and confidence to experiment. But the same research quietly points to where the leverage is greatest: the jump from beginner to competent. That's not the already-comfortable few. That's the millions of young people in programs like Ghana's One Million Coders and Nigeria's 3MTT who are just getting started. Aim a simple reflection habit at exactly that moment, and you put the scarce, high-value skill: Judgment. Judgment within reach of the people most often left outside the room. That's the bet we're making.
"Learn to code" was always shorthand for something deeper: take a real problem, and reason your way to a solution you can defend. Confidently. AI hasn't retired that skill. It's promoted it to the main event in the ongoing conversation around AI for Sustainable Development.
If you teach, run a coding club, or lead a digital-skills program, I’d be honored if you tried this with your learners. The free 7-day starter lives at agenticppa.com/ghana-ai-coding-journal.
Source: Anthropic, “Agentic coding and persistent returns to expertise” (2026), and the Anthropic Economic Index, March 2026 report.
