Singapore Is Teaching Students to Use AI Without Outsourcing Their Thinking

That creates a more difficult educational task than simply providing devices: students must learn how to question an AI system’s output, identify errors and decide when using it would undermine the purpose of an assignment.

The real risk is cognitive offloading

Cognitive offloading is not automatically harmful. People routinely use external tools to reduce unnecessary mental load. A calculator, reference book or spellchecker can free attention for more demanding work.

The problem begins when a student delegates the core learning operation before attempting it independently. That might mean asking AI to form the argument, interpret the evidence, solve an unfamiliar problem or write the explanation before the student has developed a provisional answer.

This distinction matters because a polished submission can conceal weak learning. AI may improve speed, fluency and visible accuracy while leaving the student without the mental model needed to:

• detect a plausible but false claim;
• adapt an idea to a new problem;
• explain why an argument is persuasive;
• identify missing evidence; or
• defend the work without the tool.

This is why educators refer to productive struggle: the effort of recalling, attempting, making mistakes, receiving feedback and revising. Removing every difficult step may make a task easier while removing the very practice through which reasoning develops.

What the evidence says—and what it does not say

The evidence supports concern, but not a simple claim that AI always weakens critical thinking.

A scoping review of generative AI in higher education describes benefits for writing, feedback, problem-solving and research while also identifying concerns about cognitive offloading, dependence and learner agency. A broader systematic review of 89 studies found that effects were conditional rather than uniform: it reported positive outcomes in 40.4% of studies and mixed or conditional effects in 23.6%. In that review, over-reliance, reduced analytical autonomy and cognitive offloading were among the leading reported risks.

Other reviews reach a similar caution. A synthesis of research on classroom learning warns that higher-order thinking—including analysis, reasoning and creativity—can be compromised when general-purpose AI bypasses the struggle required to acquire skills. It recommends establishing foundational knowledge first, teaching AI literacy and scaffolding the way tools are introduced.

At the same time, the research base is still developing. Reviews note inconsistent definitions of critical thinking and differences in subjects, ages, tasks and kinds of AI assistance. Much of the evidence therefore supports a conditional conclusion rather than a universal one:

AI is more likely to support learning when it acts as a scaffold, feedback partner or challenge to student reasoning—and more likely to undermine learning when it replaces foundational interpretation and analysis.

The most important gap may be between task performance and durable learning. A student can complete an assignment more effectively without gaining the knowledge or judgment that the assignment was designed to develop.

From plagiarism detection to learning design

After ChatGPT’s public release in 2022, much of the initial institutional response focused on cheating, plagiarism and detection. Singapore’s university policies now show a more differentiated model. AI use may be allowed in some coursework, but the rules remain course-specific, and students remain responsible for the accuracy, integrity and attribution of submitted work.

NUS guidance states that students should acknowledge AI use, check expectations with instructors and remain responsible for the quality and integrity of their work. Its policy identifies both direct AI output and undeclared paraphrasing of AI output as improper forms of plagiarism.

The limits of detection are also becoming clearer. NTU has said it will discontinue its institutional AI detector from 2027, citing unreliability and a lack of empirical validity for proving academic misconduct. The university plans to place greater emphasis on documenting work processes and standardised AI-use disclosures.

That points toward a different assessment philosophy. Instead of trying to infer authorship from a final text alone, educators can examine how a student reached a conclusion through:

• unaided first attempts;
• iterative drafts and revision histories;
• annotated sources;
• reflective process notes;
• oral explanations or defences;
• local, discipline-specific or personal evidence; and
• tasks that require students to apply concepts in unfamiliar situations.

The goal is not to make assessment “AI-proof.” It is to make the learning being assessed visible.

What responsible AI use looks like

A practical rule for students is simple: use AI after, alongside or against your own thinking—not instead of it.

For students

1. Start unaided. Make an outline, attempt the problem, identify relevant sources or state a provisional position before opening an AI tool.
2. Use AI as a thinking partner. Ask it to generate counterarguments, identify assumptions, suggest search terms, create practice questions or critique a draft.
3. Treat output as a claim, not evidence. Check facts, quotations, citations, calculations and interpretations against credible primary or authoritative sources.
4. Keep an AI-use record. Note the tool, prompts, significant output, what you accepted or rejected and how you verified it. NUS guidance already requires acknowledgement and personal responsibility, while some departments require more detailed declarations.
5. Retain independent command. A student should be able to explain and defend every major claim without relying on the tool.

For educators and institutions

Schools and universities can make expectations clearer by assigning each task a defined level of AI permission:

• Prohibited: the task measures unaided recall, calculation, writing or another foundational skill.
• Permitted with disclosure: AI may support research, feedback or editing, but the student must document its use and verify the result.
• Expected: students are assessed on their ability to interrogate, improve or challenge AI output.

A strong workflow can sequence these stages: first attempt unaided, receive guided AI feedback, revise independently and reflect on where the AI was correct, incomplete or wrong.

Teachers should also distinguish prompt-writing from understanding. A sophisticated prompt is not proof that a student understands a subject. A well-supported decision to reject a plausible AI answer may be stronger evidence of learning.

For younger learners, MOE’s stated position is that AI use should benefit learning, be purposeful, age- and developmentally appropriate, safe and responsible. Recent reporting also describes supervised, phased introduction of MOE-developed tools from Primary 4 and expanded AI training for teachers. Children therefore need guided classroom use and adult support rather than open-ended private access.

Access is not the same as AI literacy

Singapore’s centralised digital infrastructure can help reduce unequal access to devices and approved learning tools. MOE says schools receive support for digital infrastructure, devices and software, with students able to use SLS across the school system.

But equal access does not guarantee equal critical use. Students whose families have more time, confidence or educational experience may receive more help questioning an answer, checking a source or recognising when an explanation is superficial. Without explicit teaching, those differences can become a new form of educational inequality.

Schools can narrow that gap by making expert behaviour visible and teachable:

• model verification and source evaluation in class;
• provide structured prompts and examples of strong and weak AI use;
• teach metacognition, including how students know that they understand something;
• provide supervised access to institutionally governed tools; and
• train teachers so student guidance does not depend on individual staff confidence.

MOE’s current AI-literacy direction includes understanding AI’s benefits, limitations and risks, questioning outputs and identifying bias or false and misleading information.

Singapore’s direction through 2029

The available policy signals point toward deeper integration, not retreat. Singapore is likely to continue embedding AI literacy across learning stages, expanding approved AI-enabled tools and using SLS to support personalised practice and formative feedback.

Universities are also likely to continue favouring disclosure, accountability and course-level rules over detector-led enforcement. Assessment will place greater value on process, application, dialogue and judgment—especially where the learning objective is independent reasoning.

The central test will be whether AI expands students’ capacity to think or quietly replaces the practice that thinking requires. Singapore’s strongest model is therefore not “AI everywhere” or “AI nowhere.” It is structured use with protected time for unaided effort.

Students should graduate able to work effectively with AI, recognise when it is wrong, decide when not to use it and produce sound reasoning when it is unavailable. That is the difference between learning with AI and outsourcing learning to it.