Give our children a licence to think: What AI literacy should look like

WE DO not let teenagers drive without a licence. Even though many can drive a car, driving is consequential enough to require structured preparation: knowledge of risks, demonstration of competence and graduated exposure.

The same logic ought to apply to artificial intelligence.

A chatbot in the hands of a child is not a calculator. It is a powerful machine that can, after only a short period of unstructured use, measurably impair the child’s ability to solve problems and reduce its willingness to persist with a difficult task. That is consequential and hence, warrants a structured response.

AI’s cognitive impact on children

Recent empirical evidence shows that this cognitive erosion is already happening.

Two large-scale controlled experiments, one published in the Proceedings of the National Academy of Sciences and the other conducted across Carnegie Mellon, Oxford, Massachusetts Institute of Technology and University of California, Los Angeles, have converged on the same conclusion: brief exposure to AI assistance measurably reduces children’s ability to solve problems on their own.

In the first, nearly a thousand high school math students who used ChatGPT scored 17 per cent worse on unassisted exams than classmates who never had AI access. In the second, roughly 10 minutes of AI-assisted problem-solving was enough to reduce persistence and impair unassisted performance across arithmetic and reading comprehension.

Both studies found that the damage was concentrated among participants who used AI to obtain direct answers. Those who used it for hints or partial guidance showed no significant decline.

The distinction is between cognitive outsourcing and cognitive co-creation. When a student asks a chatbot to solve a problem, the student skips the productive struggle that makes the learning stick. When a student asks for a clue and then works the rest out alone, learning proceeds.

In other words, the problem is the pattern of use and not the tool itself. And patterns of use, unlike tools, can be taught.

A silent erosion

What makes this dangerous is that the damage is invisible at the point of contact.

A child using AI to complete homework will produce better output.

Grades may rise. Parents will see evidence of competence where competence is eroding.

The analogy is like the frog who (allegedly) leaps out if suddenly put in boiling water, but if the temperature is raised gradually, it stays until it is too late.

Similarly, AI assistance produces immediate, visible gains in output while incrementally and quietly degrading the underlying capacity for persistence, independent reasoning and self-correction. By the time the habit is entrenched, the brain-bound muscle for learning has already atrophied.

The cumulative effect of sustained outsourcing could undermine the very abilities AI was designed to support in the first place.

This finding is significant. The mechanism by which such cognitive erosion occurs matters because it points to a policy response.

We must act now

Singapore has more reason than any other country to act on this finding, and more capacity to act well.

Its students ranked first in the world across all Programme for International Student Assessment domains. In mathematics, their 2022 score is the highest any country has recorded in any subject since the assessment began.

That lead represents decades of investment in the human capacity to think carefully, persist with hard problems and reason independently. Those seem to be the same capacities that unstructured AI use seems to erode as evidenced by these studies.

The country’s EdTech Masterplan 2030 already commits the Ministry of Education to develop AI literacy across all schools, with an explicit focus on helping students understand what AI can and cannot do.

The infrastructure, intent and need are clearly in place.

What must be added is a curriculum that addresses the specific cognitive risks these studies have identified: that students will use AI to skip the thinking rather than to deepen it.

A durable AI literacy framework is essential

We have spent the past year developing such a response: a three-stage AI literacy curriculum designed for children from age five to 18.

The first stage addresses what AI is and whether it can be trusted. Children learn that a chatbot does not think – it matches patterns in text. It is confident because of its fluency, even when wrong, as it has no concern for truth.

The second stage addresses the question that these studies make urgent: am I learning, or am I skipping learning? Students are taught to distinguish between AI as tutor and AI as shortcut, between using a tool to strengthen thinking and using a tool to replace it.

Aristotle would have recognised the stakes: intellectual capacities, like virtues, are formed by exercising them, not by watching a machine do the work.

The third stage addresses civic responsibility. The information environment is a shared commons, and what any individual can rationally outsource may, once everyone does, erode the shared stock of reasoning a society depends on.

When individuals outsource their thinking and judgment to machines, the effects do not stay individual.

A population that cannot evaluate what it reads degrades the epistemic infrastructure – the system of shared knowledge – on which democratic governance depends. Our epistemic autonomy is jeopardised, but so is the epistemic authority of others on which we often rely.

This three-stage curriculum is assessed through evidence-based tasks, not rote examinations:

• A student earning a “Learner’s Permit” keeps a verification journal.
• A student earning a “Provisional Licence” produces a “Knowledge Source Audit” and defends a self-developed argument against AI-generated objections.
• A student earning a “Full Licence” maps an information environment, evaluates its health and produces original communication designed to teach younger students.

The framework is built to be durable. The content, as underpinned by cutting-edge research, updates annually to track the speed at which AI capabilities and manipulation tactics evolve.

What earns a Learner’s Permit in 2026 may be a baseline expectation by 2030. The architecture stays, while the difficulty calibrates upwards.

More than an infrastructure issue

In our view, AI literacy belongs within existing digital literacy requirements, either as an integrated component or as a standalone credential.

The precedent exists: financial literacy mandates now sit within the curriculum in multiple jurisdictions, because policymakers recognised that financial decisions are too consequential to leave to informal learning.

AI literacy is upstream of financial literacy, media literacy and nearly every other competence that a knowledge economy requires.

A child who cannot distinguish between genuine understanding and fluent pattern-matching from a machine is not equipped for any subsequent intellectual task any more than, in pre-AI days, a child who could not tell the difference between working a problem out and copying the answer from the back of the book was equipped for the next one.

Most countries will treat AI’s effects on children as an infrastructure problem to be solved by filters, usage caps and content moderation. Singapore is positioned to treat it also as an education problem.

That is the harder and more important response. The evidence from two independent research teams establishes that brief, unstructured AI use begins the erosion of cognitive performance and persistence.

A well-designed curriculum, informed by research, embedded early and updated continuously, prevents it.

The window for acting before the habit becomes the norm is not large. Singapore has the institutional capacity, the educational infrastructure and the national interest to move first. It should.

Jesper Kallestrup holds the Chair in Philosophy at the University of Aberdeen. J Adam Carter is professor of philosophy at the University of Glasgow. Sriram Raghavan, a research scholar with Kallestrup, is an adviser and investor in technology firms