Primary & Secondary Education

AI as a subject, or AI across everything?

One of the earliest choices any government faces is structural: do you teach AI as a distinct subject, or do you weave it through everything?

The dedicated subject model creates a visible commitment, makes it easier to train specialist teachers, and produces graduates who can describe what AI is and how it works. Estonia's ProgeTiger programme — which introduced computational thinking as a standalone strand from primary school in 2012 — is the most-studied example. By 2023, over 95% of Estonian schools were participating. Outcomes include measurably higher digital literacy scores and a tech sector that punches well above the country's size.

The cross-curricular model argues that AI literacy is less a technical subject and more a way of thinking, and that separating it creates a false impression that AI is only relevant to computing class. Finland has taken this path, asking teachers across all subjects to integrate AI tools and critical questions into existing learning — a science class might examine how AI models make predictions, a social studies class might interrogate algorithmic bias. Singapore's "AI for Everyone" initiative sits somewhere between: a structured national rollout with core AI concepts taught explicitly, but applied in context across subject areas.

The trade-off is real. Dedicated subjects are easier to implement and measure, but risk creating AI as a niche concern. Cross-curricular integration is more powerful in theory but depends on every teacher being capable and confident — which brings us to the bottleneck discussed below.

Which skills age best?

There is genuine disagreement about what to optimise for. The main camps:

Computational thinking — the ability to decompose problems, recognise patterns, and design systematic solutions — has strong advocates. It transfers across domains and doesn't require knowing any particular language or tool. Critics note that AI now does much of the computational work itself, making some classic CS skills less valuable.

Critical thinking and information literacy — evaluating sources, identifying reasoning errors, recognising manipulation — is increasingly argued as the more urgent priority. As AI generates text, images, and data at scale, the ability to interrogate outputs matters more than the ability to produce them.

Creativity and judgment — roles requiring novel synthesis, aesthetic sensibility, or contextual human judgment appear more durable. Some economists argue these are the skills AI augments rather than replaces. However, they are also the hardest to teach and assess reliably.

"Human skills" — empathy, communication, collaboration — appear consistently in futures research as AI-resistant, particularly in care, teaching, community, and relationship-intensive roles. Whether schools are the right place to explicitly develop these, or whether they emerge from good general education, is itself contested.

The honest position is that no one knows with certainty which skills will be most valuable in 2040. This uncertainty is an argument for a portfolio approach rather than a single bet.

Assessment in an AI-available world

Existing assessments were largely designed for a world where students did their own thinking in controlled conditions. That world is gone.

The options range from prohibition — banning AI tools in assessments — to full integration — redesigning assessments to assume AI availability and test what students can do with it. Neither extreme is obviously right.

Prohibition may be unenforceable and teaches students nothing about working with AI; it also creates a strange asymmetry where the skill of using AI tools is developed in the real world but never assessed. Full integration risks undermining the ability to assess foundational knowledge and reasoning — if a student uses AI to write an essay, what exactly has been demonstrated?

A middle path increasingly discussed internationally emphasises oral and practical assessment — presentations, debates, laboratory work, interviews — where AI assistance is less decisive and human capacity is more visible. NZ's NCEA system has more flexibility for this than many international equivalents. Whether that flexibility is being used, and how consistently, varies significantly across schools.

The teacher bottleneck

No curriculum reform is faster than teachers can implement it. New Zealand has approximately 55,000 practising teachers, most of whom trained before AI tools of this kind existed.

The options for accelerating teacher capability include:

Mandatory professional development — requiring all teachers to complete AI-specific PD. Fast to implement administratively, but quality depends heavily on the PD itself. There is limited evidence on what PD for AI in schools actually changes practice.

Specialist AI teachers or mentors — creating a cohort of specialist educators who support other staff. Builds genuine expertise but creates a two-tier system and doesn't scale easily to smaller or rural schools.

Embedding in initial teacher education — ensuring graduates arrive AI-capable. This is the right long-term fix but does nothing for existing teachers, whose average career is another 15–20 years.

Organic adoption supported by good tools — providing well-resourced platforms and reducing barriers to experimentation, rather than mandating specific approaches. Lower cost, but uneven — the enthusiastic early adopters gain while hesitant teachers wait.

The bottleneck is real whichever path is chosen. Teacher shortages already exist in some regions and subject areas; adding a new competency requirement without resourcing the training creates pressure without capability.

The equity dimension

Any approach to AI in schools interacts directly with existing inequalities. New Zealand's school system exhibits significant variation in resourcing and outcomes by socioeconomic decile (now equity index), geographic location, and ethnicity.

An AI education strategy that works well in well-resourced urban schools and less well elsewhere does not simply fail some students — it widens the gap between them and those who succeed. The digital divide is not only about device access, though that remains a real issue in some rural and low-income communities. It also shows up in teacher capacity, school leadership, internet reliability, and the extent to which whanau can support digital learning at home.

Any option that relies on schools self-organising around AI — without targeted resourcing for lower-equity environments — should be assessed honestly against this risk.