Artificial intelligence tutoring apps can deliver real, measurable improvements in GCSE maths results — but only for pupils who own a personal device and have reliable home internet, according to a major new study from The Alan Turing Institute. The findings, drawn from a year-long trial tracking more than 4,000 secondary school students across England, suggest that AI learning tools risk entrenching the very inequalities they are often marketed as solving. The research arrives at a politically sensitive moment, with the Department for Education (DfE) actively weighing whether to subsidise AI learning platforms across the state sector.
What the study found
Conducted across 38 secondary schools between September 2024 and July 2025, the Turing study compared pupils who used a popular AI maths tutoring app against a control group receiving standard teaching. On average, regular users of the app improved by roughly two-thirds of a GCSE grade — a substantial gain in a subject where small percentage differences can determine sixth-form and apprenticeship eligibility.
But the headline figure conceals a stark split. When researchers disaggregated the data by home circumstances, almost all of the benefit was concentrated among pupils with a personal laptop, tablet or smartphone and a stable broadband connection. For students reliant on shared family devices or intermittent mobile data, the measurable improvement shrank to near zero.
“The technology works — that’s the uncomfortable part,” said Dr Priya Nandakumar, the study’s lead author and a senior researcher in education technology at the Turing. “It genuinely helps children learn maths. But it helps the children who were already best placed to succeed, and it does almost nothing for those who weren’t. That isn’t a neutral outcome — it actively widens the gap.”
Why the divide is so sharp
The researchers point to the nature of effective AI tutoring itself. The apps rely on frequent, low-stakes practice — short sessions of 15 to 20 minutes, often in the evenings or at weekends, when the software’s adaptive algorithms can reinforce weak areas. Pupils without a private device simply cannot build that habit.
The study identified several compounding barriers for disadvantaged pupils:
- Device access: Many shared a single household device with siblings or parents, limiting evening study time.
- Connectivity: Pupils on pay-as-you-go mobile data rationed usage, avoiding bandwidth-heavy interactive content.
- Quiet space: Crowded or noisy home environments made sustained, focused sessions difficult.
- Digital confidence: Households with less technology experience offered weaker support when pupils hit technical snags.
Notably, the gap persisted even when schools provided in-class access to the apps. The researchers found that classroom-only usage produced only marginal gains regardless of background — suggesting the real value, and the real inequity, lies in out-of-hours independent study.
A dilemma for the Department for Education
The timing is awkward for ministers. The DfE has signalled growing enthusiasm for AI in classrooms, having funded earlier pilots to reduce teacher workload, and is now understood to be considering a subsidy scheme that would give state schools discounted access to approved AI tutoring platforms.
Education analysts warn that a poorly designed subsidy could backfire. Subsidising software without addressing the hardware and connectivity gap, they argue, would effectively channel public money into a tool that benefits wealthier families most.
“If you fund the app but not the laptop, you’re not levelling up — you’re paying to widen the attainment gap,” said Marcus Whitfield, an independent education policy analyst. “The lesson from this research isn’t ‘don’t use AI’. It’s that AI is only as equitable as the infrastructure around it. Any subsidy has to come bundled with devices and connectivity, or it’s worse than doing nothing.”
Some campaigners have called for any AI rollout to be paired with a revived national device-lending scheme, echoing the emergency laptop distribution seen during the pandemic. Others argue funding would be better spent on additional human tutoring through the National Tutoring Programme, which research has consistently shown narrows gaps rather than widening them.
Cautious optimism from researchers
The Turing team stopped short of opposing AI tutoring outright, stressing that the underlying tools are pedagogically sound. Their recommendation is conditional: deployment should be accompanied by guaranteed device access, subsidised home connectivity for eligible families, and protected supervised study time within the school day for those without home provision.
“We’re not technophobes — we want every child to have what these apps offer,” Dr Nandakumar added. “But equality of access has to come first. Otherwise we’re building a two-tier system and calling it innovation.”
What this means
For schools, parents and policymakers, the study reframes the AI tutoring debate from “does it work?” to “who does it work for?” The evidence suggests these tools can genuinely lift attainment — but left to the market, they will reward the already-advantaged and deepen the digital divide. As the DfE finalises its position on subsidies, the Turing’s findings draw a clear line: investing in AI software without simultaneously closing the device and connectivity gap risks spending public money to make educational inequality worse, not better. The technology may be ready for the classroom; the infrastructure around it is not.
Photo by Kampus Production on Pexels
