AI in Schools: Three Things to Get Right

Making AI Work for Learning, Not Just for Efficiency

AI in Schools: The Current Landscape

As

Ethan Mollick

argues in Co-Intelligence, AI is best understood as a “General Purpose Technology”, akin to steam power, electricity, or the internet, except it is unfolding at a much faster pace and at the level of cognition rather than mechanics. Earlier general purpose technologies like electricity or the internet took decades to diffuse, while LLMs have scaled orders of magnitude faster, hitting 100 million users almost immediately, and improving at a pace that dwarfs any other major technology.

And AI adoption in schools is moving fast. In the US, 28 U.S. states have issued official guidance on AI use in K–12 classrooms, urging districts to tackle privacy, bias, and proper use head-on (this is a decent summary of how states are responding).

In the UK, the DfE, alongside Ofsted and Ofqual, has issued official guidance on how generative AI can be used in schools. It frames AI use around safeguarding, data protection, and academic integrity. Inspectors are instructed to judge how schools use AI to support learning, not the tools themselves. However in the light of Ofsted’s controversial report card, this is somewhat concerning. Are they in a position to judge schools' AI use without any evidence base yet for what "good" AI integration actually looks like?

Encouragingly, some early adopters are building AI oversight committees, bringing together teachers, administrators, parents, and students to audit algorithms, monitor usage, and address stakeholder concerns (NEA Sample Policy). Transparency is a core theme: policies require that parents and students know which AI tools are in use, and that only GDPR- or FERPA-compliant vendors are approved.

However with many school AI policies overall, what you tend to see are compliance-heavy documents which focus mainly on safeguarding, data protection and academic integrity. Of course this should be our first priority clearly, but from looking at a lot of these over the summer it seems that many schools have no real sense of what to do with AI in three crucial areas: curriculum, instruction and assessment. So here are three things for school leaders to consider from a science of learning perspective.

1. Ensure Curriculum Planning Has Disciplinary Expertise

I get the sense that ahead of the new school year, some schools are working with AI-generated curricula which are effectively plastic plants: they may look convincing on the surface but lack the roots, depth, and vitality of real curricula developed by expert teachers.

For AI to properly inform effective curriculum and instruction, it needs to have three things: firstly a model of what the student knows, secondly a “map” of the domain or subject to be known and lastly, some kind of responsive instructional strategy to bridge the gap between the two. But knowledge is messy: skills overlap, combine, and vary in granularity. The systems also struggle to distinguish between procedural knowledge (knowing how to execute steps) and declarative knowledge (knowing facts and concepts), often collapsing both into simplistic skill estimates.

In conversations I’ve had with developers of large-scale adaptive systems, I’ve seen first hand the complexity hidden beneath apparently simple decisions. Should "addition of fractions" be one knowledge component or several? Data might statistically support highly specific categories like "addition of fractions with unlike denominators where one is a multiple of the other," but such granularity becomes instructionally meaningless and simply unmanageable for teachers.

The key point here is that an effective curriculum must be judged not just on usability or novelty but on how well their underlying models reflect the actual structure of disciplinary knowledge. This is why AI tutoring systems often feel robotic; they're missing the rich conceptual connections that expert teachers intuitively understand. An English teacher knows that a student struggling with a certain conceptual idea might actually just be unaware of what a certain word means and so the instructional input needs to be at the vocabulary level not the conceptual level, but AI systems often can't make those diagnostic leaps.

Policy implications:

Schools need subject specific curriculum teams, not just IT departments, leading AI procurement decisions. The key question isn't "Does this AI system work?" but "Does it understand our subject the way our best teachers do?"

This requires subject specialists to evaluate whether AI tools align with disciplinary thinking patterns, in the the way historians construct arguments differs fundamentally from how scientists test hypotheses.

The problem is that LLMs (to some extent) treat all knowledge as equivalent information to be processed and so some curricula could end up like the plastic plants which decorate but never grow. So schools should map their curriculum's conceptual hierarchy before introducing adaptive AI tools, identifying where genuine prerequisites exist versus where the system might create artificial dependencies.

For example, a science department might map out that 'balancing chemical equations' requires prior understanding of 'conservation of mass' and 'atomic structure,' while recognising that 'periodic table trends' can be learned alongside rather than before equation balancing. Schools should create subject-specific evaluation rubrics that test whether AI systems actually recognise these disciplinary distinctions.

2. Make Adaptive Learning Actually Adaptive

One trend we’re likely to see with rapidly advancing technology is AI being used more frequently for homework and in-class assignments. (Whether that is a good or bad thing is another question, I’m just saying it’s going to happen.) And to be fair, one area where I think AI *could* be a game-changer, after workload and admin which is currently the low hanging fruit, is adaptive learning.

This I think, it the holy grail for AI advocates and traditionally, Edtech adaptive systems have promised truly adaptive and personalised learning but in a lot of these systems so called “adaptivity” often boils down to mere difficulty adjustment rather than meaningful responsiveness. Many of these personalised systems have tracked right/wrong answers but then failed to capture why a student is struggling and more importantly, what to do about it.

However, AI undoubtedly holds huge potential for adaptive learning if its designed with what we know about how learning happens. But we are right to be wary. Let’s face it, a lot of Edtech was basically digitised versions of worksheets which didn’t work in the first place but the dynamic and probabilistic, pattern-matching nature of AI means it can surface subtle trends across huge datasets.

Indeed, I’ve been thinking for some time that the near endless variable complexity of trying to line-up core drivers of learning like retrieval practice, spacing and interleaving might be a problem that AI will do far better than humans. Effective systems must balance competing strategies across timescales from immediate decisions (when to provide hints or feedback) to session-level choices (which topics to practice, when to stop) to long-term pathways (spaced repetition timing, prerequisite relationships, maintaining motivation).

The problem is that AI still lacks the diagnostic insight of a teacher who recognises misconceptions in real time from the look in a kid’s eye of their tone of voice, or more usually the specific nature of their misunderstanding.

Even sophisticated AI systems face data challenges that undermine adaptive decision-making. Students don't engage as intended, they guess rapidly, cheat on difficult problems, or abandon tasks midway etc. These behaviours create systematic biases in the data used to train adaptive algorithms. Worse, adaptive systems create feedback loops where their own decisions influence the data they collect. If an algorithm incorrectly estimates that certain problems are easy (due to cheating), it may assign them inappropriately, leading to frustration and more problematic behaviours.

School Policy implications:

As I said, where I think we will see a lot of schools adopt AI in the name of pupil learning will be in carefully monitored in-class use and homework. Policy here should emphasise that adaptivity must go beyond mere pacing and remediation. It should be judged on whether it genuinely supports effortful thinking, rich understanding, addresses misconceptions, and integrates transparently with teacher expertise rather than locking it out. Policies should emphasise that adaptive learning must mean more than merely speeding up strong students and slowing down weaker ones, it should be truly adaptive in the way the best teachers are.

This requires schools to evaluate systems based on instructional effectiveness rather than engagement metrics or completion rates, while requiring transparency about how systems model knowledge, make decisions, and handle conflicts between strategies. Schools must ensure teacher agency in overriding system recommendations based on classroom knowledge and pilot carefully with clear criteria for what constitutes genuine adaptivity versus mere difficulty adjustment. They should avoid optimisation traps where systems maximise easily measurable proxies like time on task or completion rates while undermining actual learning, demanding instead evidence of misconception detection rather than just right/wrong tracking.

3. Keep Humans In The Loop for Assessment

All of the best models I’ve seen stress the importance of human oversight. A really great example of this in terms of assessment is the exciting work by

Daisy Christodoulou

and No More Marking using AI for comparative judgement. This model acknowledges that AI can speed up the process (presenting scripts, sorting data, identifying patterns etc) but the act of judgment still lies with the teacher.

Their work shows why the “human in the loop” matters. In 2023, they hoped AI could instantly grade essays and give personalised feedback, but hallucinations and generic comments soon proved otherwise. The breakthrough was redesigning the workflow: teachers still judge and add short audio notes, while AI transcribes, collates, and scales. This keeps teachers as the arbiters of quality while AI handles the legwork, delivering comparable scores and feedback with far less time than traditional marking.

Of all the use cases of AI in education, this seems to me to be one of the most viable for two reasons, firstly it directly tackles a genuine pain point for teachers in terms of workload, namely marking and secondly it eliminates well known teacher bias and inaccuracy in assessment.

Policy implications:

Schools should state clearly that AI can support but not replace teacher judgment in assessment. In practice this means: using AI to collate and organise student work for possible comparative judgement sessions; allowing AI to generate draft feedback that teachers review and edit; and prohibiting AI from making final grading decisions. Policies should also require training for teachers so they know how to use these tools efficiently while keeping professional judgment at the centre.

Schools implementing AI-supported assessment should start tentatively. Begin with one department piloting comparative judgment sessions for 4-6 weeks, measuring both time savings and teacher satisfaction with feedback quality. Teachers typically need 2-3 training sessions to become comfortable with the workflow, but report significant time savings once established, often reducing marking time by 60-70% for essay-based assessments.

As Dylan Wiliam recently cautioned, AI tools may help with logistics like scoring or collating, but “there are limitations to AI's feedback and assessment capabilities”. What matters is how we design assessments to be forward-looking, formative, and student-centered, not just efficient.

Final Thoughts: A Complex Set Of Trade-Offs

A key problem with implementing the science of learning is that it often highlights that “X improves learning,” but doesn’t tell you how to do it and rarely answers whether it is practical, affordable, or scalable.

Most school leaders I work with are really grappling with a complex set of trade-offs in terms of teacher time, the effectiveness of a particular strategy, and the opportunity costs of what gets left out. The question now for school leaders is now doubly complex because suddenly we all have the capability of an AI-powered exoskeleton giving us a set of tools that can dramatically extend what’s possible, but also introduce new risks, costs, and dependencies.

Every AI tool forces trade-offs between precision and usability, complexity and cost, individualisation and fairness. A school policy must surface these trade-offs openly, deciding for example, whether the marginal gains of a highly complex adaptive system are worth the additional costs in teacher time, training, or infrastructure.

Ultimately, the challenge for schools is to treat AI as a general purpose technology with unpredictable spillovers as Ethan Mollick observes. Like steam, electricity, or the internet, its effects will extend well beyond the first use cases we see in classrooms today. That means school policies should be living documents: flexible, transparent, and continually revised in light of evidence. The real test of leadership will not be whether a school adopts AI, but whether it does so in a way that strengthens disciplinary knowledge, amplifies teacher expertise, and preserves the human dimensions of learning that no algorithm can replace.