AI in Education: What Actually Changed This School Year

Every year produces a wave of predictions about AI and education, and most of them describe a classroom that does not exist yet. The honest version of this school year is less dramatic: the tools improved, the policy caught up unevenly, and the change showed up in staff work before it showed up in lessons.

What follows is what changed, what did not, and what is worth watching — without invented statistics.

The change happened in planning, not in teaching

The visible shift this year was in teacher workload rather than pedagogy. Planning, question writing, and feedback drafting moved into tools that produce a document, and that is where the hours went.

Three reasons:

• Planning has no audience. A weak draft costs nothing until you teach it, so teachers were willing to experiment there first.
• The artifact is editable. A plan is a page you can fix. A live lesson is not.
• No student data is involved. The first thing teachers checked was whether they had to paste student work anywhere. Planning does not require it.

Classroom use — students using AI directly — moved more slowly, and mostly in schools that had already written a policy.

Policy arrived, unevenly

The pattern across districts this year was policy following practice rather than leading it. Teachers were already using planning tools; the acceptable-use document appeared afterwards, usually distinguishing between teacher-side drafting and student-facing use.

Three positions emerged:

1. Permitted with disclosure. Teacher-side tools allowed; student use allowed with citation and teacher approval.
2. Permitted but unmanaged. No formal rule, individual judgement, wide variation between departments in the same building.
3. Restricted. Blocked on school networks, sometimes with teacher-side tools caught in the same filter.

The practical consequence: whether a teacher could use a planning tool this year depended more on their network administrator than on their pedagogy.

Detection tools lost credibility

AI detectors were the most discussed and least trusted technology of the year. They produced false accusations against careful writers and missed lightly edited machine output, and several institutions quietly stopped relying on them.

What replaced them, where anything did, was process evidence: drafts with revision history, in-class writing, oral defence of written work. Slower and less scalable, but defensible when a student's integrity is at stake.

For teachers, the takeaway is narrow and useful: do not build an assessment around a detector. Build it around work you can see being made.

What did not change

Class sizes, timetables, and the number of duties. The structural constraints on teaching time are unchanged, which is why time saved on planning so quickly gets absorbed.

The core of the job. A model can draft a plan; it cannot read a room. The parts of teaching that teachers describe as the reason they stay — the explanation that finally lands, the conversation after class — are untouched so far.

Assessment design. Most assessments still measure what was easy to measure before, and generating questions faster has made that more visible rather than less.

Where AI in education still struggles

Three areas where the gap between demo and classroom stayed wide:

Feedback that changes behaviour. Drafting comments is solved-ish; making sure a fourteen-year-old reads and acts on them is not, and no model fixes that.

Differentiation at scale. Tools can produce three versions of a worksheet. Coordinating those versions across a class, a support teacher, and a parent conversation is still manual work.

Anything requiring institutional memory. Knowledge of last year's cohort, the department's unwritten standards, the fact that this unit always runs long. Models have no access to it.

What to watch next year

Three developments worth tracking, in order of likely impact on a working teacher:

1. Integration with gradebooks and learning platforms. The value is in removing the copy-paste step, not in better prose.
2. Cheap, school-hosted models. Data policy gets easier when student work never leaves the building; cost and maintenance become the new problems.
3. Assessment redesign. More schools are moving weight toward in-class, process-visible work. That is a bigger change than any tool.

The thing that will not happen next year: a model that knows your class. The information it would need — who is improving, who is pretending, who needs a different example — is not in any training set, and it is not going to be.

Frequently asked questions

Is AI in education mostly hype?

The classroom-transformation claims are. The workload claims are not: planning and feedback drafting genuinely take fewer minutes than they did three years ago, and that is where the adoption is.

Should schools block AI tools?

Blocking teacher-side planning tools costs time and is hard to enforce. Blocking student-side use during unsupervised assessment is a narrower, more defensible rule, and easier to explain to families.

What should a teacher do this term?

Pick the one artifact that consumes an evening, use a tool that returns a printable document, and check the output. Document what you changed. That record is more useful than any district policy, and it is what you will be asked for if the question ever comes up.

Bottom line

AI in education this year changed the staff room more than the classroom. Planning and drafting got faster, policy arrived late and unevenly, and detection tools lost credibility. The teaching itself — the judgement, the room-reading, the relationships — is still the part no one has automated, which is a useful thing to remember when the next prediction arrives.