As universities race to adopt AI tools, a new paper in Frontiers of Digital Education warns that the technology is quietly reshaping the conditions under which teachers make professional judgements—and that without deliberate pedagogical design, it risks making teaching “technically active but normatively thin.”
The conceptual paper, by Ying Ji and Neil Mercer of the University of Cambridge, introduces the concept of the judgement space: the pedagogical condition in which professional judgement is not only exercised but articulated, examined, and sustained through interaction. In AI-supported higher education, they argue, this space becomes a three-party condition, as AI-generated evaluations enter the conversation alongside teacher and student.
The question is not whether AI can evaluate student work—it increasingly can, the authors write. “The question is whether teaching is designed so that AI’s evaluative outputs become material for professional deliberation rather than a substitute for it.”
Drawing on educational philosopher Gert Biesta’s account of education’s three purposes—qualification (knowledge and skills), socialisation (becoming part of disciplinary communities), and subjectification (developing as an autonomous person)—the authors identify a persistent epistemic asymmetry. AI readily captures qualification-domain evidence such as correct answers, error patterns, and performance metrics, whose quality criteria can be defined in advance. Evidence of socialisation and subjectification—whether a student’s hesitation signals genuine struggle, whether a challenge to a peer reflects real engagement—depends on normative, contextual judgements that only a teacher who knows the student, the discipline, and the purposes can make.
This asymmetry, they argue, is not a technological limitation to be solved by better AI, but reflects the nature of those domains, where what counts as quality is itself a normative question. The result: An evidence base systematically weighted toward what is easiest to measure, narrowing the judgement space even without removing the teacher’s authority.
The authors cite evidence consistent with this concern, including a randomised experiment in which learners using generative AI improved essay scores without corresponding gains in knowledge acquisition or transfer—a pattern the study’s authors termed “metacognitive laziness.”
To counteract this, Ji and Mercer advance the dialogic inquiry triangle, a framework developed in school-based practitioner research, extended here to AI-supported higher education. It structures interaction across three dimensions: inquiry INTO dialogue (building analytical capacity to interrogate AI outputs), THROUGH dialogue (collaborative deliberation through “exploratory talk”), and FOR dialogue (orienting toward independent judgement). From this, the authors derive six structural principles for teacher–student–AI interaction—for example, that AI provides data for professional judgement rather than definitive evaluation, and that teaching should preserve occasions for productive struggle rather than automatically resolving difficulty—along with a Professional Judgement Audit for faculty development.
The authors are explicit that the framework is conceptual. The judgement space construct, the structural principles, and the audit are theoretically derived and remain untested in higher education; the triangle has so far been examined only in school contexts. They set out a three-stage research agenda to empirically investigate the judgement space, test which principles matter most across disciplines, and validate the audit as a reflective tool. They also acknowledge that dialogic practices carry cultural and linguistic specificity, and that highly standardised institutions may offer teachers little discretion to create such conditions.
Faculty development in AI-supported environments must extend beyond equipping individual teachers with AI competencies, the authors conclude. “It must address the pedagogical conditions under which professional judgement can function—conditions that can either be enriched or impoverished by how AI is positioned.”
DOI: 10.1007/s44366-026-0107-x