What Makes a Teacher Replaceable

A paper went up yesterday in Frontiers in Artificial Intelligence asking a question most of us have had put to us at a staff meeting by now, usually half as a joke. Will AI take teachers' jobs? Georgios P. Georgiou of the University of Nicosia put it to 300 people who work in education and in AI, and then did something more useful than collect opinions: he tested which institutional conditions actually predict the answer.

The reassuring part first. Overall, expectations of teacher job loss over the next decade came back low to moderate. Nobody in the sample is packing up the classroom.

It is the next layer that I have not been able to put down. The strongest predictors of expecting teachers to be replaced were, in the paper's own words, "automated feedback, large-scale AI adoption, and standardized content delivery". Expectations clustered around the direct automation of instruction, rather than around weak governance or policy gaps.

Read that list of three again and notice what it actually describes. Automated feedback. Scale. Standardised delivery. That is not a prediction about robots. That is a description of a school that has been quietly reorganised around sameness — one route through the content, one acceptable form of response, one kind of feedback, delivered to everybody at once.

People expect teachers to become replaceable precisely where teaching has already been flattened into delivery. Which means the study, without ever intending to, has made the clearest argument for universal design I have read this year. If flattening is what makes a teacher look replaceable, then variability is what makes one irreplaceable. Multiple means of representation, expression and engagement is not a kindness bolted onto a standardised system. On this evidence it is also the part of the job that nothing automated is anywhere near doing.

I went looking, because I always do, for three words: accessibility, disability, learner variability. None of them appear in the paper. I do not say that as a criticism of the author, who asked a narrow question and answered it carefully. I say it as a description of the field. The research on AI and teaching is being written as though learner variability were a footnote rather than the condition of every classroom any of us have ever stood in.

On the same day, a second piece landed that I think belongs beside it. Steven Yoder reported for The Hechinger Report that young workers are steering away from the careers they think AI will take and towards the ones they believe it cannot. Pew found in August that 73 per cent of adults under 30 now expect AI to mean fewer jobs, up from 61 per cent two years earlier. Stanford's Digital Economy Lab found that 22 to 25-year-olds in AI-exposed jobs saw a 19 per cent employment decline between late 2022 and the middle of this year. The destinations are the trades, healthcare, and — I noticed this with some feeling — teaching.

Here is the question the coverage does not ask, and the reason I am writing about two American stories on an Irish blog. The routes those young people are moving into are, with real affection for all of them, among the least adjusted routes we have. Apprenticeships, nursing, classroom teaching. They run on interviews, placements, rosters, vetting, long unbroken shifts in loud rooms, and an enormous amount of social instruction that nobody ever writes down. Academic routes have spent twenty years learning to build in reasonable accommodation — unevenly, grudgingly at times, but visibly. Vocational and clinical routes have had far less of that conversation.

So if a cohort pivots into hands-on work at scale, something entirely predictable happens. The neurodivergent young people in that cohort meet a selection process that was designed for somebody else, and they get read as unsuitable rather than unaccommodated. Nobody records that as a failure. It appears in the data, if it appears at all, as a person who did not complete.

There is a lever here, and it is lying on the ground in Ireland unused. The Department of Education and Youth's own Guidance on Artificial Intelligence in Schools, published with Oide in October 2025, already says in plain terms that AI can assist in implementing Universal Design for Learning by offering multiple means of representation, expression and engagement. It is advisory rather than statutory, and it describes itself as a living document. But the sentence is written down, in a national document, with the Department's name on it. Anyone building a school AI policy this term can quote it.

That is the whole of my practical advice, and it comes in two halves. In the school: write the AI policy as a design question rather than a cheating question, because a rule written to catch people will always be enforced hardest against the pupils who most needed the tool. And beyond the school gate: if we are about to send more school leavers down apprenticeship and care routes, somebody should be asking now what those routes assume. Is the application an unprepared interview? Is there a written alternative? Does the placement supervisor get told anything useful before day one?

Inclusion by presence is not inclusion. A training place that somebody cannot survive is not access, and a classroom that offers one route through is not neutral just because everybody is sitting in it. Change the environment, not the child — and this time, change it before the numbers arrive rather than after.

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.