PISA 2025, AI and O-Level English: Is AI Helping Students Think, or Helping Them Avoid Thinking?

Singapore topped the reading table in PISA 2025 with a score of 535. Some 86% of students attained at least Level 2 proficiency, against an OECD average of 69%, and 22% were top performers at Level 5 or above, where the OECD average is 6%.

That is worth celebrating.

But the OECD released these results on 8 September 2026, and buried in the same findings is a second set of numbers that deserves just as much attention:

Source: OECD, PISA 2025 Results (Volume I): Singapore country note.

Singapore students are not merely encountering artificial intelligence. They are already using it extensively in the work through which they are supposed to develop their ability to read, think and write.

For many parents, this raises an increasingly common question:

“My child’s homework looks better with ChatGPT, but is my child actually getting better at English?”

That question matters for students preparing for O-Level English, SEC G3 English K300, IP Language Arts and A-Level General Paper, because these examinations do not merely reward polished sentences. They require students to understand, infer, evaluate, organise, argue and communicate independently.

Which raises a larger question: is AI helping our children think better, or helping them avoid the difficult thinking through which they become better?

A simple test for students using AI for English

If you use ChatGPT for O-Level English, SEC English, IP Language Arts or A-Level GP, try these three questions.

That is a simple way to distinguish using AI for learning from using AI to avoid learning. The rest of this article is about why the distinction matters more than it used to.

PISA 2025 does not say “AI is bad”

Across OECD countries, students who reported not using AI to draft written assignments tended to outperform students who did, in science — the focal domain of this cycle.

But that is an association. It does not prove that ChatGPT causes poorer academic performance.

PISA reveals something rather more interesting. Among frequent AI users, students who used AI to support learning and had been taught how to assess the quality of AI-generated information tended to perform better than frequent users who had received no such guidance.

Singapore may be doing something right here. 81% of Singapore students report learning in class how to assess AI-generated information, against an OECD average of 63%.

The Straits Times led its coverage on that weekly-use figure, and it is the right one to lead on. So perhaps the debate should no longer be “should students use ChatGPT?” With only 4% of Singapore students never using it, against 14% across the OECD, that question has already been settled by the students themselves. The more useful question is what they should use it for.

AI can support learning when students remain intellectually engaged. It becomes much less useful educationally when it replaces the reading, thinking and writing students need to practise for themselves.

The most worrying finding may not be about AI at all

It is about reading.

Average reading scores across OECD countries fell by 28 points between 2015 and 2025 — roughly one and a half years of learning. One in five 15-year-olds is now a low performer in science, mathematics and reading, up from 16% in 2022.

The proportion of “hasty readers” — students who rush through a text and answer incorrectly — has nearly doubled since 2018.

That is not a vocabulary problem or a grammar problem. It is a problem of attention, patience and the willingness to stay with a difficult text long enough to understand it. Which is precisely the work that summarising a passage with AI removes. Recall that 41% of Singapore students use AI to summarise texts they were supposed to read.

The skills under pressure are the ones that become more valuable, not less, in a world saturated with AI-generated language:

AI can now generate beautifully fluent prose. But beautifully fluent prose can still contain a terrible argument. It can be logically weak, factually incorrect, poorly evidenced, superficially evaluated, built on an unsupported assumption, or completely irrelevant to the question asked.

A student who cannot recognise those defects has not become more capable simply because the words on the screen look impressive. Fluency is not judgement.

Why homework improves while O-Level English does not

A student submits a composition with sophisticated vocabulary. The sentence structures are elegant. The ideas sound mature. The GP paragraph contains an impressive international example. The homework looks noticeably better.

Then the student sits a timed school examination and cannot reproduce anything close to it.

The AI-assisted product improved. The student’s underlying capability may not have.

This distinction matters enormously. English improvement cannot be measured by asking how good the submitted work looks. We also have to ask whether the student can independently reproduce the reasoning and writing that produced it.

That is a very different test — and it is the one an examination applies.

What this means for O-Level and SEC G3 English

Look at what these examinations actually require. Students have to write independently, comprehend texts, summarise accurately, infer meaning, distinguish relevant from irrelevant information, understand purpose, tone and context, develop ideas coherently, adapt communication to audience and purpose, and express themselves accurately under examination conditions.

Under the current English Language 1184 examination, Paper 1 Writing and Paper 2 Comprehension each account for 35% of the result.

From 2027, students move into the Singapore-Cambridge Secondary Education Certificate. G3 English carries the new K300 subject code, which SEAB cross-references to the current 1184 syllabus, and SEAB states that there is no change in overall examination standards under the SEC. We have explained that transition in detail in our guide to what parents need to know about SEC G3 English K300.

None of those skills disappears because generative AI tools exist. If anything, they matter more.

A student in an examination hall cannot outsource the difficult moment of deciding what a passage really means, which information matters, what inference can reasonably be drawn, what an example actually proves, how to develop an argument, and what to leave out.

Those decisions are the student’s own. The danger is not that students use AI. The danger is that students use AI to repeatedly perform the exact cognitive work they themselves need to become good at.

For A-Level General Paper, the issue is sharper

General Paper is not an examination of how much information a student can accumulate. A student can now ask for five arguments about inequality, ten examples about climate change, or arguments for and against censorship, and receive them in seconds.

But GP 8881 requires something more demanding. Candidates must comprehend, infer, analyse, evaluate, summarise and synthesise. Paper 1 requires a coherent response to one of eight essay questions. Paper 2 requires close work with written arguments, and ultimately the application of ideas beyond the passages themselves.

The decisive question is not whether the student knows an example. It is whether the student understands what the example proves. Can they see its limitations? Identify an assumption? Distinguish correlation from causation? Produce a meaningful counterargument? Decide whether evidence genuinely supports the claim? Qualify an overstatement? Explain why one interpretation is stronger than another?

That is judgement. And judgement cannot be outsourced indefinitely without weakening through disuse.

The Application Question makes this clearest. Students often fear the AQ because it cannot be mastered by memorising examples. The student must understand an author’s argument, determine whether and how it applies to a wider context, select relevant evidence, and explain the relationship between evidence and claim.

AI can certainly supply Singapore examples. But possessing an example is not the same as understanding why it is relevant, what it proves, and where its limitations lie. That final layer is where strong GP performance is won or lost.

How students should actually use AI for English and GP

The useful shift is to treat AI as an intellectual opponent rather than an intellectual substitute.

And instead of consuming an AI answer passively, mark it. Where is the unsupported claim? Where is the weak example? Where does correlation become causation? Where does sophisticated vocabulary disguise imprecise thought? What would an examiner challenge here?

That is AI literacy. It is also English education — and the PISA finding that guided students outperform unguided ones suggests it works. We have written more on this in If AI can write an essay, why does your child still need to learn to write?

Five things parents can do

Parents do not need to ban AI. But it helps to change what you look for.

What English education is becoming

If students increasingly encounter machine-generated language in everyday learning, English education acquires a new responsibility. Students need to learn not only how to produce clear language, but how to interrogate it: to test claims, identify assumptions, evaluate evidence, recognise bias, distinguish confidence from accuracy, compare interpretations, and recognise when fluent language conceals weak reasoning.

In that sense English may become one of the most important subjects for AI literacy in Singapore. Comprehension, inference, argument, synthesis and evaluation are no longer simply examination skills. They are how a person navigates an environment in which plausible language can be generated almost infinitely.

There is an irony in generative AI: it may make traditional English skills more valuable, not less. When machines produce unlimited words, human value moves upstream — towards choosing, judging, questioning, interpreting, evaluating, connecting, and deciding what deserves to be said.

For O-Level and SEC G3 English K300, students still need strong comprehension, writing and independent expression. For IP Language Arts, sophisticated reading and textual judgement remain central. For A-Level General Paper, critical thinking and evaluation become more important still.

The student of the future should not simply be able to ask AI for the answer. They should be able to read that answer and say:

“That sounds convincing. But it isn’t.”

That may turn out to be one of the most important forms of literacy we can teach.

How a parent can tell whether a student really has these skills

The more useful question is not “does my child use ChatGPT?” but “can my child independently comprehend, reason, evaluate and write at the level the examination requires?”

That can be diagnosed.

At The Raffles English Studio, our work in Secondary English, O-Level and SEC G3 English K300, IP Language Arts and A-Level General Paper focuses not on the final written product but on the capabilities underneath it: comprehension, inference, argument, evidence, evaluation, structure, precision and independent judgement.

The aim is not students who can get AI to write impressive English. It is students who can recognise when an impressive answer is actually weak — and then produce something better themselves.

Frequently asked questions about PISA 2025, AI and English