Artificial intelligence is often discussed in education through the language of transformation. Yet the most useful question is usually much more practical. Where can AI solve a genuine educational problem without weakening professional judgement or creating new dependencies? One UK government initiative to support the development of AI tools for education offers a useful case through which to think about that question. Its focus on formative assessment, feedback and teacher workload addresses problems that schools have faced for years. At the same time, the way such programmes are funded, governed and technically designed will determine whether their promise becomes sustainable practice. (Department for Education et al., 2025).

Why formative assessment is a sensible place to start

Formative assessment is one of the strongest areas in which AI can support education because it is closely connected to learning while also creating substantial workload. Unlike summative assessment, which usually records achievement at the end of a unit or course, formative assessment is intended to help learners improve while learning is still taking place. It can reveal misunderstandings, show where further explanation is needed and give students opportunities to act on feedback before a final judgement is made. (Education Endowment Foundation, 2026).

The difficulty is scale. Meaningful formative assessment requires repeated opportunities for students to practise and receive useful feedback. For teachers, that can mean a large amount of marking, checking and analysis. When workloads are already high, the ideal of frequent individualised feedback can become difficult to sustain.

AI can help with the repetitive parts of this process. It can assist with initial analysis, identify common patterns in responses, support the drafting of feedback and help teachers see where a class is struggling. Used carefully, this does not remove the teacher from the process. It can instead reduce some of the routine work so that professional time is directed towards interpretation, explanation and the students who need more focused support.

The value of educationally grounded data

Another important feature of the initiative is the emphasis on an educational content store. General-purpose generative AI systems are trained to operate across an enormous range of topics and contexts. That flexibility is useful, but education often requires something more controlled. A tool that supports a particular curriculum, age group or assessment task needs relevant source material, clear boundaries and an understanding of the context in which its output will be used. (Department for Education, 2026).

A well-managed repository of educational content can therefore be more than a technical resource. It can help developers build systems around curriculum-relevant material rather than asking schools to rely on generic models with little connection to local teaching aims. It can also make evaluation more meaningful because the tool can be judged against an agreed body of educational content.

The principle matters beyond this particular programme. Educational AI should not begin with the question of what a model can generate. It should begin with what teachers and learners need, what knowledge the system is expected to use, and how its output will be checked.

Funding and scope still matter

The ambition of any AI programme has to match the resources available to it. The initiative discussed here allocated £1 million across sixteen developers. Supporting several organisations has advantages. It avoids placing the whole programme in the hands of a single provider and allows different approaches to be tested. However, spreading a relatively modest budget across many development teams also raises a practical question about depth. (Department for Education et al., 2025).

Educational technology needs more than a working prototype. Tools intended for schools require testing, safeguarding, accessibility work, data protection, technical support and careful evaluation with educators and learners. A system may appear impressive in a demonstration while still being unsuitable for routine classroom use. If funding is too thinly distributed, there is a risk that programmes produce a collection of promising pilots without providing enough support for rigorous development and long-term use.

Scope is equally important. Trying to address many subjects and assessment types at once can dilute effort. Generative AI is particularly suited to some language-rich tasks, including feedback on writing and open responses. Other areas may already have mature computer-assisted assessment systems or may require very different forms of reasoning and verification. A smaller number of clearly defined educational problems can therefore be a stronger starting point than an attempt to cover everything.

Independent oversight should be built in

The technical quality of an AI system is only one part of its educational value. Commercial developers understandably bring expertise in software, product design and AI. But decisions about educational validity, learner development and appropriate feedback also require independent expertise.

Independent review should therefore be part of the design process rather than something added after deployment. Universities can contribute AI researchers, education specialists, psychologists, accessibility experts and experienced teachers who are able to examine both the technology and the educational assumptions behind it. Their role should not be to block innovation. It should be to test whether the system is doing what it claims to do, whether the evidence supports its use and whether foreseeable risks have been addressed.

This is especially important for formative assessment. Feedback can influence how students understand their ability and what they do next. An inaccurate or poorly framed response is not simply a technical error. It can affect learning. Human oversight and independent evaluation are therefore essential parts of responsible design.

Avoiding unnecessary technological dependence

A further question concerns long-term dependence on major AI providers. Commercial foundation models can offer powerful capabilities, but institutions should consider what happens when prices, terms of service, data arrangements or technical features change. A system that is affordable during a pilot may become expensive at scale. A workflow designed around one external platform may also be difficult to move later.

Open-source models and centrally managed educational infrastructure are not automatically the right answer in every case, but they should be part of the options considered. Universities and public organisations can play an important role in evaluating models, hosting systems where appropriate and sharing technical expertise. The aim should be to preserve flexibility and public value rather than allowing educational practice to become locked into a small number of commercial services.

From promising pilots to useful educational infrastructure

The initiative has several sensible foundations. It focuses on formative assessment rather than technology for its own sake. It recognises teacher workload as a real constraint. It supports more than one developer and places value on education-specific content. These are all constructive choices.

The harder work is what comes next. Educational AI needs sufficient funding for testing and support, a manageable scope, independent evaluation and a clear plan for technological sustainability. It also needs an explicit understanding of where human judgement remains essential.

The wider lesson is that good educational AI will not be defined by the sophistication of the model alone. Its value will depend on whether it solves a real problem for teachers and learners, whether its outputs can be trusted and checked, and whether institutions retain enough control to use it responsibly over time. If those conditions are treated as core design requirements, AI can support formative assessment in ways that are genuinely useful rather than simply novel.

Sources and further reading

The sources below are those used in, or directly relevant to, the original article and the claims retained in this blog adaptation.

Department for Education, Department for Science, Innovation and Technology, Phillipson, B. and Kyle, P. (2025). AI teacher tools set to break down barriers to opportunity. GOV.UK, 13 January 2025. https://www.gov.uk/government/news/ai-teacher-tools-set-to-break-down-barriers-to-opportunity

Department for Education (2026). AI Content Store. Public beta resource for educational content designed for use in AI applications. https://aicontentstore.education.gov.uk/

Education Endowment Foundation (2026). Feedback. Teaching and Learning Toolkit. https://educationendowmentfoundation.org.uk/education-evidence/teaching-learning-toolkit/feedback