The child who hands up less than they can do

There is a particular silence that comes down over a kitchen table when a child has been accused of something they did not do. I have sat in it. Most parents of a neurodivergent child have sat in it at some point, and what makes it so hard to shift is that the child usually cannot produce the explanation in the shape the explaining is being demanded in.

A piece of research published on Tuesday has been in my head all week. The Center for Digital Thriving, a research centre at Project Zero at the Harvard Graduate School of Education, surveyed more than a thousand American teachers and principals and interviewed students about artificial intelligence. What it describes is what the researchers call two-way suspicion. Teachers do not trust what they are reading. Students do not trust how their work will be read.

Two sentences from it. Beck Tench, the principal investigator, said: “Teachers can't see their students anymore. They have a hard time understanding what their students know.” And Lydia Bach, a twelfth-grade student, said: “I think that AI in general, students are taught to feel ashamed of their use of it.”

But the finding that stopped me was neither of those. It was that students reported deliberately scaling back their best work. Writing worse than they can write. Handing up something safely mediocre, because a strong piece now carries a risk that a weak one does not.

Who pays for suspicion

I want to be careful here, because this is not a teacher problem and I will not let it be framed as one.

Picture the teacher. Thirty students. A correcting pile that was already too big before any of this started. No detection tool she trusts, because the detection tools are not trustworthy and she knows it. No departmental guidance that tells her what to do on a Tuesday afternoon with one essay in front of her that reads oddly. She has been handed an impossible job and told to use her judgement, while the conditions that make judgement possible — knowing the child, seeing them work, having time — are the exact things that have been eroded. If she is suspicious, it is because she has been put in a position where suspicion is the only tool left on the desk.

Now put a particular child into that room.

A child whose written voice has never matched their spoken voice. I have taught autistic students who write with a fluency and precision they cannot reach out loud, and I have taught students for whom the opposite is true. A child who has spent a decade being correctly taught to use scaffolds: sentence starters, planning frames, speech-to-text, a reader, a structured writing app, every bit of assistive technology we fought to get them. Their work has always looked, to a teacher who does not know them, slightly too polished for the child sitting in front of her. That gap used to be evidence that the supports were working. This year it reads as evidence of something else.

And if that child is challenged, they cannot do the performance of innocence. They will not make the eye contact. They may go completely flat, or completely over-explain, or freeze, or laugh at the wrong moment. Every one of those responses is read, by a system under pressure, as guilt.

So the child learns the lesson the Harvard students described. Hand up less. Write below yourself. Make it look like what they expect from you.

That is the opposite of everything we have spent twenty years trying to build.

The thing we are actually measuring

This is where I think our own system is badly exposed, and it is worth saying plainly. The Irish model was never designed for the range of children now sitting inside it. What we have is a softened cousin of an older, more elitist, unbending model — it has been made kinder at the edges, it has had supports bolted onto it, but its spine is still a sorting machine. A terminal written exam, a narrow band of acceptable ways to show what you know, and an assumption that the standard output of a standard child is the thing being measured.

A system built to sort does a specific thing when it gets nervous. It tightens. It falls back on the format it trusts. That is why the first instinct everywhere this year has been to reach backwards — more handwriting, more supervised conditions, more oral defence, more of the room with a pencil and a piece of paper. Nobody setting those policies is asking who cannot write by hand for three hours, or who cannot speak to defend their own work under pressure. Access is not argued down in those conversations. It simply never comes up.

There was a second study this week that makes the same shape visible from another angle. A randomised trial of an AI study assistant across 2,379 undergraduates found it lowered final grades and cut course engagement — and that the worst effects fell on first-generation students. The new thing landed hardest on whoever had the least scaffolding around them to begin with. It always does. That is not a fact about AI. It is a fact about what happens when you add something to an environment nobody redesigned.

What I would actually do

Three things, and none of them require a budget.

Separate the policy from the accusation. A school AI policy that only describes detection and sanction is not a policy, it is a threat. Write down, before you need it, what happens to a student who is wrongly suspected — who hears about it, how it is recorded, how it is undone, and who checks on the child afterwards. If that paragraph does not exist, the cost of being wrong is being carried entirely by children.

Ask about process, not product. If you want to know whether a student understands their own work, the question is not “did you write this”. It is “show me how you got here” — the plan, the notes, the drafts, the dead end they abandoned. That question is answerable by a child who cannot defend themselves verbally under accusation, and it is a better assessment of learning anyway.

And design the alternative in at the start. Every time an assessment gets redesigned in a hurry, the alternative arrangement arrives six weeks later through the appeals process, by which point a family has had to fight for it. Put it in the first draft. It costs nothing at the drafting stage and it costs a family enormously at the appeals stage.

There is one more cost I want to name, because it rarely gets counted. When a household goes into a dispute with a school, it takes the whole household. The sibling whose match nobody got to. The parent who used the annual leave on meetings rather than on a holiday. The evening that was supposed to be ordinary. None of that appears in any policy document, and all of it is real.

We are at risk of teaching a generation of children that the safest thing to do with their ability is hide it. Change the environment, not the child. Because inclusion by presence is not inclusion — and neither is trust extended only to the students who already look like what we expect.

Elaine Walsh is the founder of Autism & Me (autismandme.ie), a Waterford-based neuroinclusive education consultancy. She writes here on Universal Design, inclusion and Irish education, from the twin vantage points of the classroom and the kitchen table.

Sources

Center for Digital Thriving, Project Zero, Harvard Graduate School of Education, reported by ABC News, 30 September 2026: https://abcnews.com/GMA/News/new-report-ai-impacting-trust-students-teachers/story?id=136874802

“The Effects of Course-Integrated AI Tutoring on Student Performance and Engagement: A Randomized University Trial”, EdWorkingPaper No. 26-1598, Annenberg Institute, Brown University, October 2026 (working paper, not yet peer reviewed): https://edworkingpapers.com/ai26-1598