AI in Education: What Is Actually Changing, and What Still Depends on Humans

The education sector has spent the past two years absorbing an unprecedented volume of AI-powered product launches. Adaptive learning platforms that adjust content in real time. AI tutors that answer student questions at two in the morning. Automated marking tools that process essay responses in seconds. The pace of product development has outrun the pace of evidence, and school leaders, university administrators, and edtech buyers are now facing a familiar dilemma: how to distinguish genuine improvements in learning outcomes from well-funded hype.

The Adaptive Learning Opportunity

The core premise of adaptive learning is straightforward and well-supported by decades of educational research. Students do not all learn at the same pace or in the same way. A classroom model that delivers identical instruction to thirty students at once is structurally incapable of meeting all of them where they are. AI-driven systems that can adjust the difficulty, sequencing, and format of learning material based on real-time performance data represent a genuine improvement on that model, at least in principle.

The evidence from well-designed implementations is genuinely encouraging. Platforms deployed at scale in structured subjects, mathematics and language learning in particular, show measurable improvements in progress rates when the adaptive system is properly integrated into a teaching workflow rather than used as a standalone replacement for instruction. The key word is integrated. The platforms that work best are those where teachers can see what the system is tracking, adjust its priorities, and use its data to inform their own decisions about which students need direct intervention. The ones that underperform are those deployed as self-contained solutions that operate independently of the classroom relationship.

Automated Assessment and the Time Question

One of the most significant pressure points in education is teacher workload. Studies consistently find that teachers in UK schools spend between 35 and 45 percent of their working time on tasks other than direct instruction, with marking and administrative reporting accounting for a substantial share. AI tools that can produce consistent, fast feedback on structured written responses represent a meaningful opportunity to recover some of that time.

The current generation of AI assessment tools performs well on tasks with relatively constrained answer spaces: short-form comprehension questions, structured argument paragraphs, factual recall. Performance degrades on tasks that require genuine evaluative judgement, including assessment of creative writing, nuanced historical argument, and extended scientific reasoning. The risk is that institutions, attracted by efficiency gains, deploy automated assessment in contexts where the tool is not fit for purpose, and students receive feedback that is superficially detailed but fundamentally inaccurate. Saving teacher time by producing poor feedback is not a net benefit.

AI can tell a student whether their essay is well structured. It cannot tell them whether their argument is worth making.

The Equity Problem That Most EdTech Ignores

AI-powered learning tools perform best in conditions that are not evenly distributed. A well-functioning adaptive platform requires a stable internet connection, a reasonably modern device, and a learning environment where a student can engage with a screen without significant distraction. These conditions are far more likely to be present in households with higher incomes and in schools with more resource. The students who could benefit most from personalised, responsive learning support are often precisely those for whom the conditions for effective AI-assisted learning are least reliably available.

This is not simply an infrastructure problem that better connectivity will solve. Research into how students from different backgrounds interact with AI learning tools suggests that the framing, language, and cultural assumptions built into most AI tutors reflect the perspective of the developers who built them rather than the full range of learners they are deployed with. An AI tutor calibrated to a particular social context will be less effective for students outside that context, even when connection and device conditions are identical. EdTech buyers who are serious about equity need to ask much harder questions about training data, cultural calibration, and accessibility than most current procurement processes require.

What School and University Leaders Should Be Asking

• What evidence exists for this tool's impact on learning outcomes, specifically for students comparable to ours, not just for the most advantaged cohort in a pilot study?
• How does the tool integrate with teacher workflow, and does it give teachers more visibility into student progress or less?
• What happens when the AI gets something wrong, and how quickly can a teacher correct it without undermining student trust in the feedback?
• Does the tool work equitably across the range of devices, connectivity conditions, and accessibility needs present in our student population?
• What data does the tool collect, how is it used, and what are the data governance implications for a student population that includes minors?

The Teacher Relationship Is Not Optional

The most consistent finding across evaluations of AI-powered edtech is that tools deployed without teacher buy-in and proper professional development consistently underperform tools that involve teachers as active users rather than passive bystanders. This finding is not surprising to anyone who has spent time in schools, but it is routinely ignored by technology procurement processes that treat software as something to be installed rather than something to be embedded in practice.

The institutions getting the most out of AI in education are not those with the most sophisticated tools. They are those that have invested in helping teachers understand what the tools can and cannot do, where AI judgement should be trusted and where it should be questioned, and how AI-generated data can inform better professional decisions rather than replace them. That investment is not glamorous. It does not feature in product demos. But it is the difference between AI that genuinely improves learning and AI that produces usage metrics without improving outcomes.