AI for dialogic education: A research agenda

I was reading a post by Paul Kirschner recently and it made me uneasy. His claim was that too much educational research is not evidence-based but eminence-based: instead of testing ideas rigorously, we defer to persuasive theories and well-known names.

Some of the thinkers he criticised are people I respect. But unfortunately, I had to admit that he is not entirely wrong.

Education has often been vulnerable to ideas that feel right without being either conceptually grounded or properly tested. That raises an uncomfortable question: when people respond positively to concepts that I put forward like “dialogic education” or “expanding dialogic space,” are they responding because these ideas are good and useful, or simply because they sound nice and fit with their worldview?

Kirschner contrasts education unfavourably with medicine, where controlled trials, replication, and systematic review have imposed a certain discipline. There is something to learn from that. Some educational questions really are like medical ones. If we want to know whether retrieval practice improves recall, we can test it directly and accumulate robust evidence.

But taking that as the model for all educational research implies a misunderstanding of the discipline.

Education is not only about what works best to achieve given goals. It is also about what those goals should be or, more fundamentally, what we should teach our children.

Disciplines are defined by their central questions. In medicine, the guiding question is about understanding, preserving, and restoring health: how to keep people well and make sick people better. Education asks what should we teach, and only then, as a secondary question, how should we teach it?

Where there is wide agreement about educational goals — the value of teaching young children basic literacy and numeracy, for example — then research of the medical kind is valuable. But if education were reduced to only that kind of research, it would become both conservative and narrow. Research on effective teaching inevitably focuses on how best to achieve given outcomes, rather than on questioning those outcomes themselves.

A more appropriate comparison might be architecture. Some aspects are technical and measurable. Architects need to know if their bricks or iron girders will support the weight, or if the plumbing will work. But these easily measurable and replicable factors do not tell us whether a building is any good. That depends on what kind of life it enables, what relationships it supports, and how it shapes those who inhabit it. Educational research works in the same way. It is a design science: not only asking how things are, as the natural sciences do, or how to fix what is broken, as medicine does, but how things ought to be. And in education, that means asking what is worth cultivating — asking what kind of future we would like for our grandchildren.

When I trained as a teacher in Bristol in 1990, I became aware of something odd. Reading the official documents and talking with others on the course, it was not clear that anyone really knew what education was for. At one level, there were answers. A government report at the time, led by a banker rather than a Professor of Education, suggested that education existed to support the British economy — to help us compete internationally. Alongside that were familiar aims: socialisation, preparing people for roles, public health messages, and so on.

But none of these quite captured what seemed most essential.

As I began to look into this more deeply, and later during my PhD, it struck me that the answer might not be complicated. There is a widespread intuition about what education is for, even if it is rarely made explicit.

I saw this most clearly working with small groups of students in classrooms in Milton Keynes. A student would be struggling with a problem, locked into a particular way of seeing it. Then something would shift. There would be a pause. They would stop assuming they understood. Then they would begin to hold more than one perspective at once. They would question the frame that had constrained them. And then — often visibly — understanding would dawn.

This moment is familiar to teachers: the light going on.

What is happening here is not simply the acquisition of a correct answer. It is an opening of a space of reflection. That space is then widened — as multiple perspectives are held in relation — and deepened through questioning and integration.

This is what I mean by education as expanding dialogic space. It can sound abstract, but in practice it is simple and observable: it is what happens when understanding grows through the encounter of perspectives.

This process does not only occur at the level of individual learners. It also scales. As we develop knowledge, we bring more perspectives into relation and deepen the frameworks through which we understand the world. Despite setbacks, there has been a broad expansion of shared knowledge over time: not simply as an accumulation of facts but as an increasing capacity to relate perspectives in order to understand our situation and solve problems.

After many years of enquiry, my working hypothesis is this: education is for expanding dialogic space. What should be taught is dialogic intelligence — the capacity to think with others, including machines, by opening, widening, and deepening dialogue. How this should be taught is through structured activities that induct learners into effective participation in dialogic space.

Advocating for the importance of dialogue and expanding dialogic space in education is relevant precisely because it is counterfactual. Contemporary education systems are not organised explicitly around this aim. While moments of opening, widening, and deepening do occur in classrooms, they are not what the system is primarily designed to produce. The dominant structures of schooling — shaped by print literacy, standardised curricula, and assessment regimes — tend to privilege convergence, recall, and individual performance. The whole system seems to be built around high-stakes exams which do not have any obvious intrinsic educational function but are useful for certification, ranking, and the allocation of scarce resources.

The emergence of generative AI makes this misalignment between real education and our education system more visible. AI does not simply add a new tool to existing practices. It challenges the assumptions of a system built around the transmission and testing of fixed knowledge. If machines can now perform many of the tasks that print-based education has traditionally prioritised, then the question of what education is for can no longer be deferred.

This returns us to Kirschner’s concern. If educational research is to avoid becoming merely eminence-based — driven by persuasive ideas rather than cumulative evidence — then claims about aims must be grounded. The hypothesis proposed here is one perspective in the ongoing debate about what education is for. Research does not simply settle educational questions by declaring propositions true or false. It feeds a disciplined cultural debate about what is worth knowing, what is worth becoming, and what forms of life education should sustain. This is how knowledge develops in any serious discipline: not by the mere accumulation of facts, but through argument, interpretation, evidence, and challenge. In that process, identities shift, consciousness expands, and shared culture develops. Research matters because it keeps this debate answerable to experience rather than to preference, fashion, or authority.

A body of recent empirical work on AI in education is beginning to reveal when and how AI expands dialogic space — and when it contracts it. Across four studies, the lesson is fairly consistent: AI can help to open and support dialogic space when it is designed as a moderator, role-holder, reflector, or provocateur, not merely as an answer machine.

Feng — GPT-4 in collaborative problem-solving

Six small groups used a GPT-4 chatbot for collaborative problem-solving. The key warning: AI can become the centre of interaction. Students may talk more to the chatbot than to each other. That expands access to ideas, but can shrink peer-to-peer dialogic space if the framing is wrong. Design matters as much as the technology.

Doherty et al. — LLM jigsaw-partner agents

LLM agents used in jigsaw-partner roles reduced social loafing and produced more respectful collaboration language. But the researchers stress an important caveat: LLMs can also reinforce poor social dynamics unless carefully designed. The quality of the dialogue depends on the quality of the design, not on AI capability alone.

Breideband et al— CoBi, dialogue as object of inquiry

Students set community norms for their discussion; CoBi listened for examples of those norms; the class then reflected on the AI’s output together. This made dialogue itself an object of inquiry, rather than using AI simply to complete a task. An interesting experiment in using AI to help build community and collaboration.

Nguyen et al. — role-based environmental science chatbots

Multiple chatbot personas — fisherman, student, influencer, researcher, engineer — gave students access to situated perspectives on climate change and local ecosystems. This is a clear design for widening dialogic space by multiplying voices. Students were not receiving a single authoritative account but navigating a range of positioned ones.

The pattern across these studies points toward a design principle: AI can widen the space of voices and hold open the space of inquiry. This requires intentional design. The default behaviour of most AI systems — efficient, helpful, convergent — works directly against it.

The problem is not a lack of evidence. It is a lack of measures that would allow this evidence to accumulate coherently — measures of dialogic space itself, not just of proxies like test scores or talk-time. If “expanding dialogic space” is to anchor a research programme, it needs to be observable. As a starting point, the following provisional definitions are proposed:

Questioning: Do contributions open new directions, surface assumptions, or reframe the problem?

Widening: Is a broader range of perspectives brought into play and genuinely engaged with?

Deepening: Are underlying assumptions examined, leading to more integrated understanding?

Integration: Do new responses emerge that incorporate previously incompatible positions?

Judgement: Are decisions made that draw on multiple perspectives and navigate uncertainty?

Sustaining: Is engagement maintained in the presence of disagreement without collapse?

The challenge is to turn these verbal pointers into indicators that can be observed and refined without reducing them to tick-boxes that capture the surface of dialogue while missing its substance.

Design-based research is appropriate here because the aim is not simply to test predefined variables, but to iteratively develop and refine both concepts and practices. Small-scale studies can explore how different AI-supported designs affect dialogic processes. Over time, through comparison and refinement, more stable constructs and design principles can emerge.

Central to this is the development of a research community. Teachers, researchers, and designers need to share cases, compare findings, and jointly refine both theory and method. The community is not secondary to the research. It is the primary site in which knowledge is constructed.

Only once constructs are sufficiently clear does it make sense to move towards larger-scale validation.

The broader aim is a reorientation of education around the expansion of dialogic space — understood as the development of collective wisdom: the capacity to engage with complexity, difference, and uncertainty in ways that support human flourishing through shared processes of judgement and decision.

This is an open research programme. It will only develop through shared critique, comparison, and collaboration.