Why coding-first AI education gets the sequence wrong

Picture

Member for

1 year 10 months

Real name

Gordon Institute Policy Forum Editor

Bio

Gordon Institute Policy Forum Editor

Input

Sep 4, 2026 09:35

Modified

Sep 7, 2026 06:55

Early AI education should develop representation, reasoning, and measurement before rewarding software fluency.
Code can execute an idea, but it cannot supply mathematical concepts that a learner has never formed.
Korea should sequence AI education by cognitive prerequisites rather than by the visibility of fashionable tools.

The Korean Education Question

Korea’s AI for All: Strategy for Cultivating Artificial Intelligence Talent expands AI education across the learning system. The ambition is understandable: children will encounter automated systems throughout adult life, and schools should not leave access to family income or private academies. The harder question is what should be taught first. A coding-first policy assumes that operating a formal language is the natural entry point to computational reasoning. That assumption is often wrong.

Programming uses representations that depend on prior concepts. A loop is easier to understand when a learner already recognizes sequence and iteration. A function in code becomes meaningful when inputs, outputs, and composition are more than unfamiliar notation. Arrays require the learner to reason about indexed structures; probabilistic models require ideas about variation and conditional information. Interfaces can hide these dependencies, but hiding a dependency is not the same as teaching it.

This does not imply that children should be excluded from computational activity. It implies that the activity should match the concepts they can investigate. Classification, decomposition, measurement, uncertainty, and rule-based reasoning can be explored without pretending that typing syntax is equivalent to understanding computation.

Mathematical Level and Problem Difficulty Are Different

A long derivation can be mechanically easy when every step is supplied. A short equation can be intellectually difficult when the student must decide what its symbols mean, whether its assumptions are defensible, and whether its solution is stable.

AI education often confuses these two dimensions:

\text{Technical complexity} \neq \text{Reasoning difficulty}.

Technical complexity concerns the machinery required to execute a calculation. Reasoning difficulty concerns the work required to formulate the right calculation and interpret its limits.

A student may know multivariable calculus but fail to recognize selection bias. Another may use elementary probability to identify a decisive base-rate error. A third may reproduce a neural-network derivation but be unable to explain why the model should generalize.

The objective is not to make advanced mathematics unnecessary. It is to place difficulty where it belongs.

Five Sources of Difficulty

None of these requires mathematical ornament. Each requires disciplined judgment.

A Simple Equation with No Unique Answer

Consider

y=\beta_1x_1+\beta_2x_2.

If $x_2=cx_1$ for every observation, then

y=(\beta_1+c\beta_2)x_1.

The data can identify the combination $\beta_1+c\beta_2$, but not the two coefficients separately. Infinitely many pairs $(\beta_1,\beta_2)$ produce the same fitted values.

The arithmetic is elementary. The problem is difficult because the desired decomposition is not identified by the available variation.

A software package may warn that the design matrix is singular. A more subtle case may be nearly singular, producing estimates that exist but change sharply with small perturbations. The modeller must then decide whether the individual coefficients are meaningful, whether new data can separate the variables, or whether the target should be redefined.

Sensitivity and Conditioning

For a linear system

$$ Ax=b, $$

the formal solution is

x=A^{-1}b.

This appears complete. Yet if $A$ is poorly conditioned, a small error $\Delta b$ can create a large change $\Delta x$:

\frac{\lVert\Delta x\rVert}{\lVert x\rVert} \lesssim \kappa(A) \frac{\lVert\Delta b\rVert}{\lVert b\rVert},

where

\kappa(A) = \lVert A\rVert\lVert A^{-1}\rVert

is the condition number.

The equation teaches a broad lesson: numerical existence is not inferential stability. A result may be calculable but too sensitive for decision-making.

Regularization replaces the unstable inverse with a controlled compromise:

\widehat x_{\lambda} = \arg\min_x \left\{ \lVert Ax-b\rVert_2^2 +\lambda\lVert x\rVert_2^2 \right\}.

The parameter $\lambda$ reduces variance and numerical instability at the cost of bias. Selecting it is not a purely computational act. It encodes how much instability the decision can tolerate.

A Basic Probability Problem with an Unintuitive Answer

Suppose a condition occurs in 1% of a population. A classifier has 95% sensitivity and a 5% false-positive rate.

Let $D$ denote the condition and $+$ a positive classification. Bayes’ rule gives

\Pr(D\mid +) = \frac{ \Pr(+\mid D)\Pr(D) }{ \Pr(+\mid D)\Pr(D) + \Pr(+\mid D^c)\Pr(D^c) }.

Substituting the values:

\Pr(D\mid +) = \frac{0.95(0.01)} {0.95(0.01)+0.05(0.99)} \approx 0.161.

Despite 95% sensitivity, only about 16% of positive classifications correspond to the condition under these assumptions.

The mathematics is taught early in probability courses. The difficult part is recognizing that the base rate belongs in the problem and that a familiar performance measure does not answer the decision question.

Difficulties Created by Dependence

The sample mean

\overline X = \frac{1}{n}\sum_{i=1}^{n}X_i

has variance $\sigma^2/n$ when observations are independent with common variance $\sigma^2$. If observations are positively correlated with common correlation $\rho$, then

\operatorname{Var}(\overline X) = \frac{\sigma^2}{n} \left[ 1+(n-1)\rho \right].

Even a small $\rho$ can prevent uncertainty from declining as quickly as the independent formula suggests. Ten thousand correlated clicks are not equivalent to ten thousand independent users. Repeated measurements from one institution are not equivalent to measurements from many independent institutions.

The formula is simple. The difficult judgment is identifying the dependence structure hidden behind rows in a dataset.

An Easy Objective Can Encode the Wrong Decision

Suppose a model minimizes average squared error:

\widehat f = \arg\min_f \frac{1}{n} \sum_{i=1}^{n} (y_i-f(x_i))^2.

If errors have asymmetric institutional costs, the objective may be inappropriate. A one-unit error in forecasting routine demand may be harmless, while a one-unit error near a capacity limit may be severe.

One alternative is a weighted loss:

\widehat f = \arg\min_f \frac{1}{n} \sum_{i=1}^{n} w_i (y_i-f(x_i))^2,

where $w_i$ reflects decision relevance.

Choosing $w_i$ requires substantive judgment. The mathematical modification is small; the institutional reasoning can be difficult.

Nonlinearity from Simple Components

A difficult system need not contain individually difficult functions.

Consider repeated application of a simple nonlinear map:

x_{t+1} = rx_t(1-x_t).

The logistic map uses multiplication and subtraction, yet its long-run behavior can change dramatically with the parameter $r$. Local rules that are easy to state can generate complex dynamics through iteration.

Neural networks use the same general principle of composition:

f(x) = f_L \circ f_{L-1} \circ\cdots\circ f_1(x).

Each layer may perform a matrix multiplication and a simple activation. The overall function can represent highly complex boundaries.

The difficulty lies in interaction and composition, not in an individually exotic operation. This is why foundational mathematics must be understood structurally. A student who knows each operation separately may still fail to reason about the full system.

Model Multiplicity

Another source of difficulty is that several models can explain the same observations.

Suppose two models $M_1$ and $M_2$ produce similar empirical risk:

\widehat R(M_1) \approx \widehat R(M_2).

The models may nevertheless imply different behavior under intervention or shift:

R_{\mathrm{target}}(M_1) \neq R_{\mathrm{target}}(M_2).

Observed fit alone cannot choose between them. The modeller needs external knowledge, new variation, experimental design, or a decision criterion.

This is common in AI. Many parameter configurations can yield similar predictions. Several features can act as substitutes in the training environment. The hard question is which representation captures structure that will persist.

A Prerequisite Model

Let an AI learning task require three capabilities: conceptual representation $M$, procedural fluency $P$, and the ability to evaluate an outcome $J$. If these capabilities are strong complements, readiness is closer to a bottleneck than a sum:

R=\min\{M,P,J\}.

A polished programming interface can raise $P$ quickly. If $M$ and $J$ remain low, the learner can reproduce a demonstration but cannot explain why the procedure works, identify an impossible output, or transfer the method to a new problem. Generative AI makes this distinction more important because it can supply much of the procedure on demand. The educational return shifts toward specifying the object, checking the result, and locating failure.

The ordering problem can also be represented as a dependency graph $G=(V,E)$. Each node is a concept or task, and an edge $u ightarrow v$ means that $u$ must be understood before $v$ can be used independently. A curriculum that begins at a visually attractive downstream node may create engagement, but it should not mistake temporary interface support for mastery of the missing upstream nodes.

What Early AI Education Can Teach

At primary level, computational education can begin with comparison, grouping, patterns, measurement, and explicit rules. Learners can ask why two classification rules disagree, how a changed observation alters a conclusion, or whether an average represents every member of a group. These are foundations of data reasoning even when no programming language appears.

At secondary level, mathematics and computation can become more tightly integrated. Functions can be explored through transformations; sequences through iteration; matrices through spatial operations and networks; probability through uncertain evidence. Programming then becomes a laboratory in which mathematical claims can be tested. The sequence moves from concept to representation to execution, with feedback in both directions.

Coding projects remain useful when they expose rather than conceal reasoning. A student who predicts a result, writes or generates code, tests edge cases, and explains a discrepancy is learning something different from a student who follows a screen recording until the output matches. Assessment should reward the former process.

Policy Design Without a False Choice

The policy choice is not mathematics or coding. It is whether software activity is aligned with conceptual readiness. Korea can publish age-banded capability outcomes, require providers to identify prerequisites, and evaluate transfer rather than completion. A course advertised as AI education should state what learners will be able to explain or test without relying on the original template.

Teacher preparation matters more than the choice of platform. Instructors need enough mathematical and statistical confidence to turn unexpected results into questions. Otherwise, classroom success is defined by whether the program runs, and the most important opportunity—the mismatch between an idea and its implementation—is treated as an inconvenience.

This sequence is consistent with the wider problem identified by AI Lifelong Learning Must Arrive Before the Skill Gap Hardens: AI competence increasingly means calibrating a system rather than merely accessing it. That capacity should be developed gradually. It cannot be inferred from the age at which a child first encountered a coding interface.

Table 1. A sequenced foundation for AI education

Stage	Primary capability	Suitable activity	Misleading proxy
Early foundation	Classification and explicit rules	Compare competing groupings	Hours spent in an app
Mathematical foundation	Functions, sequences, variation	Predict and test transformations	Syntax recall
Computational application	Representation and debugging	Implement and challenge a model	A working demonstration
AI judgment	Validation and uncertainty	Detect failure under changed conditions	Prompt fluency

Source: Author’s analysis

Conclusion

Korea should introduce children to computational thinking, but it should not define progress by how early they imitate adult software work. The durable objective is a learner who can represent a problem, understand the structure being executed, and test whether an answer deserves confidence.

As AI systems make code cheaper, premature coding specialization becomes less—not more—compelling. A sound sequence builds the mathematical and empirical judgment that allows later tools to become genuinely useful.