Should We Still Teach Coding?
The foundation endures. The answer is not “keep coding” or “drop coding” — it is augmentation, not replacement. Here is the evidence, laid out visually. Every figure is sourced; preliminary findings are labeled as such.
The foundation matters more, not less, as the machines get better at the typing.
“Vibe coding” is real — and moving fast
The term was coined by Andrej Karpathy on X on 2 February 2025. Within a year it stopped being a joke and started being a workflow. That is exactly why the foundation matters.
of Y Combinator’s Winter 2025 startups had codebases that were ~95% AI-generated.
Source: Y Combinator, 2025
of AI-generated code samples introduced a security vulnerability.
Source: Veracode analysis, 2025
The cognitive question
Early findings ask what happens to understanding when the machine does the producing. These are preliminary and self-reported — read them as signals, not settled science. They point the same direction: the thinking you offload is the thinking you may not retain.
of the LLM group could not quote from the essay they had just produced.
Source: Kosmyna et al., “Your Brain on ChatGPT,” MIT Media Lab, arXiv 2025 — preprint, not peer-reviewed, n = 54.
Higher confidence in the AI correlated with less critical thinking; higher confidence in oneself correlated with more.
Source: Lee et al., CHI 2025 (Microsoft Research + Carnegie Mellon) — 319 knowledge workers, 936 real examples, self-reported.
Self-reported reduced perceived effort when using GenAI
Share of participants reporting less effort, by cognitive activity (Bloom categories).
Source: Lee et al., CHI 2025 — self-reported. Preliminary.
The durable skill
What outlasts every language is computational thinking (Wing, 2006) built through learning by making — constructionism (Papert, 1980). That foundation is exactly what you cannot delegate to an AI.
SPECIFY
Say precisely what a thing should do — inputs, outputs, states, rules, edge cases.
IMAGINE
See the thing that does not yet exist and picture how it behaves before a line is written.
STRUCTURE
Organize ideas and functionality into something coherent — see how the pieces connect.
The three human axes the essay defends — what you cannot delegate to AI.
What the transfer evidence says
Carrying the base from one language to the next.
Carrying it to distant, unrelated domains.
Source: Scherer et al., 2019 meta-analysis. Bars are illustrative of the direction of findings (strong vs. modest), not measured effect sizes.
The modeling toolkit
The vocabulary that turns a vague intention into something buildable.
The path forward
The calculator lesson
The benefit came when the tool was integrated into instruction. Early, unstructured use could hinder learning. The tool was never the point — when and how you introduce it was.
Source: Hembree & Dessart, 1986; Ellington, 2003.
The pathway ladder
- Ages ~5–10 · K–5Scratch
Learning by making — procedures, variables, loops, conditions, sequencing.
- Ages 8–14 · FreeExperience CS
Raspberry Pi Foundation + Google.org — lesson plans, rubrics, teacher guides.
- K–12 · Rebranded 2 June 2026CodeAI (formerly Code.org)
Keeps computer science inside “digital fluency,” alongside AI science and data science. “Digital fluency” popularized by Mitchel Resnick.
countries host free Code Clubs.
Source: Code Club (Raspberry Pi Foundation), 2025
children on Scratch — 1B+ projects shared.
Source: Scratch, 2024
Keep the making. Keep the structure. Keep the modeling vocabulary.
Add the AI as the executor of well-specified, well-imagined, well-structured human intent.
You know you have to change it. The whole task is changing it well.
← Back to the essay