Opinion: AI and the Invisible Labour of Teaching Writing (Part 2/2)

Feature image credit: Vitaly Gariev on Unsplash

By Ly-ann Tan Low

Ly-ann Tan Low is a homeschooling parent, former MOE teacher, Literature educator, and accredited Institute for Excellence in Writing instructor. She holds First Class Honours in Post-Colonial English Literature and Art History and a Postgraduate Diploma in Education in English Language and Literature, and has spent over twenty years teaching students to read closely, think carefully, and write with increasing independence, across Singapore, Japan, and Hong Kong. She is also trained in AI for learning design. She runs LitLens Studio, a homeschool writing programme in Singapore.

Disclosure: The author runs LitLens Studio, which offers writing instruction to families. This piece’s argument overlaps with the service her studio provides. The author and editor agreed that disclosure was important so readers can weigh the argument with that context in mind.

In Part 1 of this essay, we looked at how policies, guardrails, and the human-to-human feedback loop help keep AI in check within the learning environment. In Part 2, I turn my attention inward to the cognitive cost of what happens in a child’s mind when an AI assessment algorithm takes over thinking, and to the human instinct the algorithm overrides.

Instinct Is Irreplaceable

At this point, as homeschool parents who have opted out of the system, you may be asking, why should MOE’s AI practices dictate how we teach at home? The answer is that MOE’s research and design choices echo the same concern raised by the MIT Media Lab findings: when the tool does too much of the thinking, the child may produce the answer without building the underlying mental capacity. They need the cognitive struggle that builds memory, ownership, and judgment. The research findings do not change whether the child is homeschooled or attends regular brick-and-mortar school.

Polgár, who was also homeschooled, explained that while AI engines can prepare players for chess games, they cannot offer structured human interaction: the kind that builds intuition through personal experience and interaction at the board. She described an incident in which she used chess robots to prepare for a tournament and, during a match, when faced with a fork, chose to follow what her AI practice had told her would lead to a win. This went against her intuition and experience, which told her it would be better to end that match on a draw. Instead, she hedged against her intuition, trusted the computer, and ignored the possibility of splitting the point. Polgár lost that game (an option that the AI never seemed to present in its all-accommodating state) and “was very much pissed off with the computer and I lost my trust (in AI).” It was then she realized that no AI can ever displace the human quality of intuition that develops from repeated struggles with other human beings. There is no shortcut to intuition. It is accumulated time at the board. “It was very painful at some moments,” she said. “Because… At some point, I said, okay, it does make sense. And then later I saw that everybody is using it,” and thought she should too.

Like Polgár, MOS Lau named this in Parliament as “instinct.” “No algorithm can replace the instinct of an educator who senses what a student particularly needs,” regardless of what AI might have suggested. What she was describing is not a feature AI temporarily lacks. Care at scale, which AI can provide, differs from relational presence, which only a human being can.

For those of us who have the good fortune of homeschooling our own children, we have been training that instinct for years. Homeschool parents who treat teaching as ongoing learning, not something they already know, have been refining it. We know when to push, when to pause, when to explain, and when to let the struggle do its work. 

The question is not whether AI has a place in writing instruction. AI can be genuinely useful for research, brainstorming, generating examples, and executing clear instructions where there is no need to weigh meaning against a real human audience. The trade-off is that AI makes practice cheap: a student could draft and revise as many times as she wished, more attempts than a teacher could ever mark by hand, let alone mark fast enough for that volume to mean anything. A student who drafts ten paragraphs and gets quick feedback on structure or grammar brings more to a real teacher than a child who drafts two. As long as the surface pass is never mistaken for the teaching itself, or allowed to become the only review a piece of writing ever gets.

Used to feed a relationship, AI feedback earns its place. Used to replace one, it is the same mechanism wearing a better excuse. 

The real question is whether we can use AI to serve our goals for our children, rather than letting the tool define them for us. ChatGPT does not have a relationship with your child. Claude does not know what they are ready for next. Gemini cannot tell how much help is too much, and DeepSeek cannot filter if your child is discouraged, defensive, tired, or finally ready to hear feedback. Even adaptive platforms built to track a child’s progress over months, such as Khanmigo or MATHia, trained to model mastery and pacing, cannot sense when a child has shut down, is running on empty, or needs a challenge.

No amount of AI tracking sophistication adds up to our humanity.

Human intuition and emotional awareness of our students’ readiness, frustration tolerance, and our relationship with them matter, especially when we remember that our goal is not just a better paragraph, but a stronger writer. 

Same Board, Different Endgame

“But what about the composition piece for the PSLE?” you say.

Preparing students for national exams and becoming a strong writer are not competing goals. Neither are they the same. Building deeper foundations in structural thinking leads to stronger writing than one built on model essays and AI corrections. Both can deliver PSLE-grade compositions.

However, genuine writing growth builds intuition for structure, vocabulary, and sentence variety for effect. Still, it takes longer, and we have to be prepared for the inevitable messiness of being human. It looks less straightforward than mimicking a piece and more unpolished, with students’ unskilled ways of manipulating writing styles because they are doing the work themselves. Which means we might have to sacrifice the quick score because the process is the goal. That is both a comforting thought and an uncomfortable space to occupy.

Lastly, we already know that machines mark the optical answer sheet portion of the national exams. In the time it took for me to gather evidence from the ground on AI’s impact, synthesize research from MOE’s position, work through a feedback framework, wrestle with ongoing parental concerns whilst articulating intuitions about writing and human connection, and develop this thesis, AI models had already leapfrogged each other a few times. The horizon where machines will scan, transcribe, and mark compositions is already here. You may have already put your finger down on the “AI Slop” that sounds so polished, with vocabulary so extensive, hits all the points, and yet weighs none of them by their importance. It is a dystopian future when everybody starts producing the same thing. Even what a person genuinely creates starts sounding like AI because she has spent years practicing for AI-writing evaluation.

Now consider a student who devoted that time to learning how to structure an argument, sift for key ideas, revise for a target audience, engage that audience for feedback, and, in doing so, move from practice to craftsmanship. That student will distinguish themselves in discursive discussions, oral defenses, and difficult debates. That is the endgame. Those will be the truly significant performance events that stand out but cannot be delegated to an AI. As the horizon for fully AI-driven assessments approaches, now is the time to build skills in our students that cannot be replaced.

Polgár has her own rule for her chess students. Learning a good move is ten percent. The other ninety percent, she says, is understanding why it is a good move.

AI provides the 10%. People, patience, and practice make up the other 90%.

As for my daughter’s composition, I could not make head nor tail of that useless piece of information. In my limited capacity, I couldn’t do one darn thing with it, so it stayed buried under the other bajillion queries I have made of AI. I do know, though, that she takes pride in retelling the compositions she has written and in fitting in an idiom like the last piece of a jigsaw puzzle. At one PSLE Chinese practice session, I asked her Chinese teacher whether she could share her thoughts and whether my daughter should try to memorize “stuff” — “stuff” being my highly technical term for the things I did not even know she needed. She responded that all those things would take care of themselves if my daughter kept reading, kept writing, and kept learning how to say what she meant with increasing clarity. So, I’m giving her that 90%, putting aside my impatience for a blitz answer to a classical problem, and offering her an audience, my uncomfortable patience, and a long runway of practice.

Filtering AI Feedback

As an educator, one of the frameworks I rely on is Grant Wiggins’ research on feedback: to what end, how much, and how reliable.¹ Feedback is only feedback if it matters to a specific goal. In my experience using ChatGPT to evaluate my daughter’s Chinese composition, I still had no bearings on where her writing stood against a curriculum, where she should be on that learning pathway based on her background, experiences, capabilities, and interests. I did not really know how to tweak my approach to “get” her. I’m sure I’m not alone in this goal: it’s not just PSLE; it’s, “Where should she be on this journey?” And AI feedback is not ‘human’ enough to understand her humanness or to tell me how her work in relation to where she is as a student and a person.

Then, there is the matter of how much feedback. Too much is counterproductive. Reimagine this with me: receiving a composition paper with red marks highlighting every error on the page. You might start correcting them from the first paragraph onward, but do you really know your most pertinent weakness, and what might elevate your writing if you improved that? That is not unlike the comprehensive feedback ChatGPT gave me on my daughter’s composition, which told me everything but nothing at the same time. It actually matters that there is a cognitive limit (no more than three elements) and a sequence that prioritizes which concerns and in what order. It should also be consistently aligned against an unchanging standard, not a goalpost that keeps moving between AI conversations, because the editor-machine in not only unknowable, but it is constantly changing. To get consistent, curriculum-aligned feedback, we need to know which model to use, how to structure the prompt, what constraints to set, and how to verify that the output is aligned with both the syllabus and your child’s developmental level. Juggling all those variables creates a moving target and leaves you with feedback you cannot tell is measured against the same standard used last week.

Assuming that we all concur that certain types of AI reliance do not serve our children, and that MOE’s cautious approach to AI is noble, i.e., to provide equal access to curriculum-aligned resources, and we even agree that it augments teachers’ professional practice, the other concern raised was whether AI could really deliver something it promises: individualised attention at scale. Every sentence gets scanned, every paragraph perused, and every student gets a response. Really? Having agreed that any AI feedback needs an adult to parse, we still have to ask whether the teacher has the capacity to unpack it individually, for a class of 30 or a class of 1. Parents of homeschool children may not share the same worry as parents of mainstream students that the teacher will be the choke point for feedback, but we can all agree that attention at scale is not the same as being seen. We all want our child to be guided by someone who has walked with this student and seen her progression, what she wrote last month, and whether or not this piece of writing is in her voice.

Here are questions I keep in mind when giving and filtering feedback to students. Use them when using AI to evaluate your child’s work:

Image credit: Cash Macanaya on Unsplash

Conclusion

When AI feedback is mistaken for teaching, we gain short-term efficiency in the final product but risk losing the long-term gains of genuine growth. We know intuitively that the theory of “desirable difficulties” shows that some forms of struggle that slow performance in the short term can strengthen long-term learning and transfer. When AI eases that struggle, we bypass the muscle practice needed to cultivate intuition and discernment. Instead, we stop exercising our own judgment and begin repeating the machine’s learned judgment, generated from the averages across all the data it has learned from. But there are advantages to AI: it makes practice cheaper, allowing students to iterate through mistakes quickly. The nuance lies in balancing both: an adult with a relationship with and deep knowledge of the child can discern when that child’s struggle is productive. In this article, I presented five practical questions to help us decide when and how to filter for feedback. The key is that AI feedback must be filtered to serve our purpose, and delivered through the lens of relationship and intent.

Who I Am Actually Writing For

A few closing thoughts that this piece owes to you, the reader. While I write this for the larger homeschool audience, I am addressing very specific people in mind. Not because they need to hear this, but because I want my writing to impact them. Specific interactions with people have compelled me to write.

These are individuals whom ChatGPT and Claude cannot fathom. Nor could they tell why I press on certain points harder because of what these people said. The most pertinent feedback this piece could receive is whether or not my reader is motivated to act or reconsider their position. In short, AI is not one of them.

* * * * *

Editor’s Note: In editing this piece, I used Claude (Anthropic) as a thinking aid to pressure-test my reads of the author’s arguments and flag factual claims that needed checking against primary sources. All editorial judgments and final decisions were my own. The author’s argument and structure are entirely hers. Phrasing was edited in places for clarity, as is standard editorial practice. The author’s own note on AI use follows below.

Author’s Note on AI Use: AI tools were used in preparing this piece as research and thinking aids, including organizing source material, testing the argument, and refining clarity. The claims, judgments, final wording, and responsibility for this article remain mine. This reflects the distinction the piece makes: AI may support thinking, but it cannot replace human judgment.

Footnotes

1. Grant Wiggins, “Seven Keys to Effective Feedback,” Educational Leadership, vol. 70, no. 1, Sept. 2012, pp. 10–16. The five feedback questions later in this piece also derive from Wiggins’ seven criteria, specifically goal-referenced, user-friendly, consistent, ongoing, and actionable feedback. No further footnotes are assigned to those passages. ↩︎

References

Fitzpatrick, Dan. “The Queen of Chess on What Schools Are Getting Wrong About AI.” Forbes, 7 May 2026, www.forbes.com/sites/danfitzpatrick/2026/05/07/the-queen-of-chess-on-what-schools-are-getting-wrong-about-ai/.

Wiggins, Grant. “Seven Keys to Effective Feedback.” Educational Leadership, vol. 70, no. 1, Sept. 2012, pp. 10–16.