AI in Schools: How the Digital Divide Could Widen the Inclusion Gap.
AI in Education Part 2 | Reading time: approx. 10 minutes.
In Part 1, we explored the genuine and substantial promise of artificial intelligence as a tool for closing the inclusion gap.
Personalised learning. Assistive technology. AI tutoring. Greater access to support for disadvantaged pupils.
It was, deliberately, an optimistic read.
This one is a little harder.
Because the same things that make AI potentially transformative for inclusion also make it capable of accelerating the inequalities we are trying to address.
And the evidence, from UK research, parliamentary testimony and recent school experiences, suggests that without serious structural intervention, we may be more likely to reinforce existing inequalities rather than remove them.
That should concern all of us working in education.
The uncomfortable headline: private schools are already ahead
In July 2025, the Sutton Trust published Artificial Advantage?, a landmark piece of research examining the emerging AI divide in English education.
Its findings were stark: private schools are adopting AI tools far faster than state schools, and the gap is widening.
Independent schools educate around 7% of pupils in England, yet it is well evidenced that they disproportionately shape access to elite universities and professions. The Sutton Trust found that these schools are investing significantly more in AI-powered learning tools, personalised adaptive platforms and staff AI training than their state-school counterparts.
Their pupils are, in effect, beginning to receive an AI-augmented education that pupils in under-resourced schools may simply not have access to.
The Sutton Trust's CEO commented that without intervention, AI risked becoming
"the latest in a long line of educational innovations that entrench advantage rather than challenge it."
That is not an exaggeration. It is pattern recognition.
EdTech Innovation Hub, reporting on the same research, highlighted the structural conditions driving this divide: funding gaps, infrastructure deficits, teacher capacity and leadership bandwidth all play a part.
They are, notably, many of the same conditions that have driven previous technology-related inequalities in education.
Although the technology changes, the dynamic does not.
The digital divide starts before anyone opens an AI tool
Beneath almost every conversation about AI in education sits an awkward reality: not every school has the infrastructure needed to use it reliably.
The Department for Education's Generative AI in Education guidance, updated in January 2025, acknowledges significant variation in schools' digital readiness, but does not mandate the investment needed to address it.
Schools serving higher proportions of disadvantaged pupils, children with SEND and pupils for whom English is an additional language are often also the schools dealing with older devices, slower broadband and fewer ICT support staff.
That matters.
Because we can talk all we like about the transformative potential of artificial intelligence, but a transformative tool is of limited use if the infrastructure to use it properly is not there.
Written evidence submitted to the House of Commons Education Committee's inquiry into AI in Education in 2024-25 raised repeated concerns about this.
The National Education Union (NEU) warned that without meaningful investment in infrastructure and workforce development, government guidance on AI would be "aspirational for many, achievable only for few."
The University and College Union (UCU) made a similar point in its written evidence (AIE0169), warning that without equitable access to the tools themselves, AI in education risks reproducing "a two-tier system in which advantaged institutions use technology to extend their lead."
This is not a fringe concern.
It is mainstream, evidenced and, at present, largely unaddressed by any ring-fenced statutory funding.
And it is not difficult to see how that happens in practice. Schools with greater financial capacity can pay for whole-school licences, premium AI subscriptions for staff, training and the infrastructure needed to make meaningful use of them. They can begin embedding AI into teaching, planning, administration and professional practice while other schools are still working within much tighter constraints around devices, software, training and staff capacity.
The result is not simply that some pupils have access to better technology.
It is that some schools are able to build an entirely different level of AI capability into the way their school operates.
If AI really is going to become part of the future of education and the workplace which we are preparing students for, then access to that future cannot depend on the ability of a school to pay for it.
Algorithmic bias: what happens when the tool gets it wrong?
Access is only one part of the problem.
Even when AI tools are available, there is a deeper question we need to ask:
What happens when the tool itself is biased?
England already has a very recent example of what can happen when algorithms are used to make decisions about children's education.
In August 2020, the government's algorithm for allocating grades following the cancellation of A-level examinations systematically downgraded many pupils from state schools and historically disadvantaged areas, while inflating results for pupils in smaller independent school classes.
The algorithm had learned from historical patterns.
And historical patterns in English education are shot through with inequality.
The backlash was swift. The decision was reversed.
But the lesson remains important: when the data reflects an unequal system, an algorithm can reproduce that inequality at scale.
Research published in the International Journal of Artificial Intelligence in Education (Springer, 2021), and replicated in subsequent studies, has documented algorithmic bias across a range of educational AI applications.
This could include assessment tools that perform less accurately for students with disabilities, language models that produce lower-quality feedback for pupils writing in non-standard English dialects, and attention-detection systems that can interpret the behaviour of autistic pupils as disengagement.
A 2024 study published in Computers and Education: Artificial Intelligence found that AI-based student progress monitoring tools showed differential accuracy across pupil demographics, with less reliable outputs for pupils with SEND and those from ethnic minority backgrounds.
The practical implications are uncomfortable.
Imagine a school leader who genuinely wants to use AI to improve provision.
Imagine them using AI-generated progress data, where an AI system has analysed, inferred, predicted or summarised information about a pupil's progress, to identify pupils who may need intervention, without the right training, prompting or agentic working practices to understand its limitations.
Imagine that the data is less accurate for pupils with SEND.
Imagine nobody realises.
A decision about intervention, grouping or an EHCP review could then be based on information that looks objective, but isn't.
And that is one of the most difficult things about algorithmic bias.
It can look like objectivity.
The school leader may be acting with the best of intentions. The system may present its output with confidence. The data may sit neatly in a dashboard or report.
But if the underlying system is producing less reliable information for some pupils, that apparent objectivity can make the bias harder, not easier, to spot.
The problem can be hidden behind the appearance of neutrality.
Farrer & Co's 2024 legal analysis, AI, SEND and Reasonable Adjustments, is clear that schools must ensure any AI tool used in provision for disabled pupils does not introduce or amplify disadvantage.
But that requires staff to be able to recognise the risks in the first place.
The Equality Act 2010 is technology-neutral. The fact that a discriminatory outcome has been produced by an algorithm rather than a person does not make it any less discriminatory. The difficulty is that algorithmic discrimination can be far harder to recognise. It does not necessarily look like the discrimination staff have been trained to spot. It can appear as neutral data, an objective recommendation or a seemingly evidence-based decision.
AI in schools creates another question: whose data?
There is another issue that is harder to see, but potentially just as significant.
What happens to the data?
AI systems learn from data. In education, that data can include some of the most sensitive information we hold about children.
A learning profile.
An SEND diagnosis.
An emotional and behavioural history.
A set of family circumstances.
The Data (Use and Access) Bill, currently moving through Parliament, has prompted significant concerns from children's rights advocates about how children's data may be used to unintentionally train AI systems and what safeguards will apply.
The Digital Futures for Children Centre's written evidence to the Education Committee (AIE0181) emphasised that children must not become "data sources for AI development without meaningful consent or understanding".
It also highlighted the particular vulnerability of pupils with SEND and children in care, whose sensitive information may be used in ways they are unable to understand, control or challenge.
For families already navigating complex systems, including EHCPs, tribunals and multi-agency reviews, the idea of their child's most intimate educational information being processed by commercial AI systems raises some very serious questions.
Who has access to it?
How is it being used?
How long is it kept?
Can a family challenge a decision made using it?
And perhaps most importantly, who gets to decide whether that use is appropriate in the first place?
AI cannot replace an inclusive workforce
There is a persistent temptation in conversations about AI in education to treat the technology as a solution in itself.
Buy the platform.
Roll it out.
Train people to click buttons.
Problem solved.
It isn't.
In SEND, that approach could be actively dangerous.
The SEND Code of Practice 2015 places the SENCo at the centre of coordinating SEND provision and working with teachers to ensure that the graduated approach is applied with rigour.
Using AI effectively in SEND therefore requires more than basic access to a platform.
It requires educators who understand what the technology can do, what it cannot do and when its output should be questioned.
It requires teachers and SENCos who can interrogate AI-generated information rather than simply accept it.
And it requires professionals who can place that information within a much bigger picture of the child.
Because a child's education cannot be reduced to a dataset.
The NEU's written evidence to the Education Committee highlighted that many teachers currently lack even basic AI literacy, while professional development provision remains patchy and often commercially driven.
Without meaningful investment in educator training, particularly for those working with the most vulnerable pupils, AI adoption in school system contexts could produce something that looks sophisticated but is poorly implemented; a dangerous combination.
Who gets to define "inclusion"?
Perhaps the biggest question is also the one that receives the least attention.
Who decides what inclusion looks like?
Most AI products used in education are developed by commercial companies. They are often designed around measurable performance indicators.
Accuracy.
Attainment.
Engagement.
Efficiency.
These things matter.
But they are not the whole picture of inclusion.
A system designed to identify "low engagement", for example, might usefully flag a pupil who needs support.
Or it might interpret a neurodiverse child's way of learning as a problem to be fixed.
Those are very different things.
The SEND Code of Practice is explicit that the views of children and young people should be central to decisions about their provision.
The Equality Act 2010 recognises dignity and autonomy as fundamental.
Yet many AI tools used in education have been developed without meaningful co-production with pupils, their families or the SENCOs and practitioners who understand their needs.
The Royal Society's written evidence to the Education Committee (AIE0215) called for greater transparency in AI tool development and procurement, alongside independent ethical review of educational AI before it is deployed in schools, particularly where vulnerable children are involved.
That feels like a sensible starting point.
Because inclusion cannot simply mean making technology available to everyone.
The technology itself needs to be inclusive.
So where does that leave us?
None of this means that AI should not be used in education.
Quite the opposite.
The potential is too significant to ignore.
But it does mean that deploying AI without addressing the structural conditions can widen rather than close the gaps we already see.
It is a choice for schools, and that choice has consequences.
The inclusion gap is not new.
The attainment gap between disadvantaged pupils and their peers is not new.
The gap between pupils with SEND and those without is not new.
Neither is the gap between children who arrive at school already behind and those who arrive with every advantage.
These inequalities are persistent and well documented. AI offers an exciting opportunity to change that, but technology will not close them on its own.
More concerning still, technology deployed carelessly, inequitably or without accountability could make these gaps considerably harder to close.
That is why the question cannot simply be:
How can schools use AI?
We need to be asking:
Who benefits from it?
Who might be disadvantaged by it?
Where are there knowledge gaps?
Whose data is being used?
Who is making the ultimate decisions?
And perhaps most importantly:
Are we using AI to make education more inclusive, or are we simply finding an easier, more sophisticated way to reproduce an unequal system?
The answer will depend on what happens next.
Part 3 of this series looks at what getting it right might actually require.
Key references and further reading
Sutton Trust (2025). Artificial Advantage? AI in Education. Read here | Full PDF
Sutton Trust (2025). State Schools Falling Behind in New AI Digital Divide. Read here
National Education Union (NEU) (2024-25). Written evidence to the House of Commons Education Committee AI in Education Inquiry (AIE0220). Read here
University and College Union (UCU) (2024-25). Written evidence to the House of Commons Education Committee AI in Education Inquiry (AIE0169). Read here
Royal Society (2024-25). Written evidence to the House of Commons Education Committee AI in Education Inquiry (AIE0215). Read here
Digital Futures for Children Centre (2024-25). Written evidence to the House of Commons Education Committee AI in Education Inquiry (AIE0181). Read here
Farrer & Co (2024). AI, SEND and Reasonable Adjustments: What Schools Need to Know. Read here
Baker, R.S. & Hawn, A. (2021). Algorithmic Bias in Education. International Journal of Artificial Intelligence in Education, 32, 1052–1092. Read here
Computers and Education: Artificial Intelligence (2024). Investigating Algorithmic Bias in Student Progress Monitoring. ScienceDirect. Read here
Department for Education (2025). Generative Artificial Intelligence in Education. Gov.UK. Read here
Department for Education / Department of Health (2015). SEND Code of Practice: 0 to 25 Years. Gov.UK.
Equality Act 2010. Chapter 15. legislation.gov.uk.
Axios (2020). How an AI Grading System Ignited a National Controversy in the UK. Read here
