Objectives and context

AI avatar tutors teach through spoken, embodied Socratic dialogue rather than a text-based interface, introducing a new source of academic support into the learner-teacher relationship. This case study investigates how using Inkling, an AI avatar tutor, influenced trust, authority and interaction within that relationship, and whether it changed where pupils direct their questions.

Methods / Approach

The design is a case study which combined 18 matched pupil surveys with dashboard data on 88 sessions and 22 hours of use between January and March 2026. We then ran two focus groups split by amount of use, to examine pupils’ motivations and attitudes and whether they felt the relationship with their teacher changed. The teacher was interviewed to explore whether she noticed any changes in her interactions with pupils, and whether the usage data informed her teaching practice.

Outcomes & Conclusions

Pupils reported a high degree of trust in the avatar, which was partly attributed to knowing their teacher set a weekly prompt to steer the conversation. All pupils felt comfortable asking the avatar questions they might not raise in class, and ease of access was reported as a reason for taking questions to it instead of their teacher. Pupils did not therefore think it weakened the relationship with their teacher, treating it as welcome additional support to reinforce understanding and rehearse arguments aloud. A small number of pupils found it did not challenge their thinking as they would have liked and therefore preferred to seek their teacher's support instead. AI avatar tutors need not undermine the learner-teacher relationship and may be a beneficial adjunct, particularly where the teacher has visibility of usage, and indeed acts vicariously through prompting the AI tutor.

Advances in conversational AI are transforming computer-based speaking assessment from traditional semi-direct, monologic tasks into interactive dialogues with AI interlocutors (e.g., Chuang & Yan, 2025; Xu et al., 2026). While this shift offers greater opportunities to assess communicative interaction, it also raises a fundamental question: Can rating scales developed for monologic speaking tasks adequately assess speaking abilities demonstrated in dialogues with AI? As AI changes the nature of assessment tasks, it may require a reconsideration of what human judgment values as evidence of speaking proficiency (Xu et al., 2026).

This presentation reports a mixed-methods study that adapted the rating scale of a high-stakes English language test for an AI-mediated dialogic speaking task. A computer-based monologic speaking task was redesigned as an interactive dialogue facilitated by an LLM-powered spoken dialogue system, represented by a humanoid avatar that expanded the conversation based on candidates' responses. The study involved 120 L2 English learners from 23 countries (CEFR A2–C2), whose performances were rated by five trained examiners. To better capture the interactiveness in the new dialogic task, the existing rating scale was adapted by integrating Fluency and Discourse Management into a new criterion, Interactive Discourse Management, with descriptors for turn-taking, clarification requests, communication breakdown management, co-construction of meaning, and interactional fluency added.

The adapted scale was evaluated through examiner think-aloud protocols (n=25) and post-marking interviews (n=5), alongside psychometric analyses of rating quality and reliability. The findings suggest that the adapted scale redirected examiners’ attention towards interaction management, responsiveness, conversational continuity, and the ability to sustain dialogue with the AI interlocutor. This means that AI-mediated assessment goes beyond simply digitising existing speaking tasks; it reshapes the construct being assessed. Thus, rating scales must adapt to the new construct to ensure that human judgment continues to recognise meaningful communicative ability in AI-mediated contexts.

Objectives and context: AI in education is usually discussed through tutoring, assessment, or learning analytics, but educational experience also depends on the physical conditions in which learning occurs. Higher education classrooms support changing pedagogical activities, including lectures, seminars, group work, examinations, and computer-based learning, yet indoor environmental quality is commonly managed through fixed comfort or ventilation thresholds. This paper asks how AI enabled Digital Twins can connect building operations with educational processes to support student comfort, concentration, and learning activity.

Methods: The paper draws on a systematic literature review of Digital Twin and cyber-physical applications in educational buildings, alongside a proposed activity-aware Digital Twin framework for higher education classrooms. The review examines how existing studies represent indoor environmental quality, learners, occupancy, and educational activity. The proposed framework links a Building Asset Twin, representing environmental systems and sensor data, with a Human Learning Activity Twin, representing pedagogical activity type, occupancy patterns, and learner comfort or concentration requirements.

Outcomes & Conclusions: The review indicates that current Digital Twins in educational buildings are largely asset centric. They monitor and optimize temperature, ventilation, CO2, lighting, or energy performance, but rarely model what learning activity is taking place or how environmental conditions relate to attention, concentration, or student experience. The paper therefore proposes a “Twin of Twins” approach in which classroom environmental control is interpreted through pedagogical context. This reframes Digital Twins as responsible AI enabled educational infrastructure: not simply tools for facilities efficiency, but systems that help institutions align physical learning environments with human centred educational goals. The contribution is relevant to responsible AI implementation because it highlights governance, transparency, stakeholder roles, and evidence-based decision making in the design of learning aware educational environments.

AI writing tools are learning to sound human, and people who read and prompt them are beginning to write like AI. The two registers are converging, and detection is losing. In blind tests, detectors miss most AI-assisted work while falsely accusing real students, non-native writers most of all. We argue this is not an integrity problem to police but a validity problem to design around. Every essay we mark is an inference from words to the mind behind them. That inference no longer holds. We call the break the Great Decoupling.

Studies already show student writing drifting toward AI style. Our simulation asks what happens next. In one loop, an AI rewrites a passage until detectors stop flagging it. In the other, a writer keeps reading AI text and absorbs its style. Eventually, an independent detector cannot tell the machine's text from the writer's. The writing stays fluent, but it no longer shows who wrote it.

We respond with design science. The diagnosis becomes six design requirements answering the standard objections, from fluency bias and gaming to privacy and oversight. Socratic Viva is the artifact that meets them, a working system probing understanding as an examiner would, through five minutes of adaptive counter-questioning. On a simulated student cohort, essay grading cannot tell genuine understanding from shallow knowledge dressed in AI-polished prose; the viva separates them, because polish cannot answer a follow-up question. Where AI grading is usually a black box, ours argues its marks in the open. After the viva the student downloads everything: video, transcript, learning outcomes and rubric, and every mark quoting the time-stamped words that earned it, signed off by the teacher. A recorded viva demonstrates the system.

Writing still builds thinking; we keep it. But when text stops proving thought, assessment should ask the thinker.

The rapid adoption of generative artificial intelligence (AI) has challenged long-standing assumptions about assessment, academic integrity, and the purpose of higher education. While institutional responses have largely focused on detecting AI-generated content, such approaches address only a symptom of a more fundamental challenge: how can educators meaningfully evaluate learning when AI can generate, explain, analyse, and communicate knowledge on behalf of learners (LeBlanc, 2026)? This challenge is particularly significant in online higher education, where traditional notions of authorship and assessment authenticity are already contested.

This paper argues that the central question is no longer whether students use AI, but how institutions can design learning environments and assessment systems that provide credible evidence of human learning and capability development (Gilmore & Nguyen, 2025). Drawing on authentic assessment (Biggs & Tang, 2011), self-regulated learning (Zimmerman, 2002), sociocultural learning theory (Vygotsky, 1978), and emerging scholarship on AI and educational purpose, the paper proposes the Human Capability Evidence Framework (HCEF).

The HCEF reconceptualises assessment as the evaluation of learners’ developing understanding, judgment, agency, ethical reasoning, curiosity, creativity, and capacity for transfer, rather than the evaluation of assessment artefacts alone. Evidence of learning is derived from both products and processes, including reflection, decision-making, iteration, collaboration, and the application of knowledge in authentic contexts. Online learning environments and learning analytics can support this approach by making learning processes visible through portfolios, reflective narratives, interaction traces, oral explanations, and authentic performance-based tasks.

Rather than positioning AI as a threat to academic integrity, the framework views AI as a catalyst for rethinking assessment and the broader purposes of higher education. The paper concludes that authentic learning should be evidenced through cumulative, layered, and educationally meaningful assessment practices (Smith et al., 2026) that demonstrate learners’ development of distinctively human capabilities and their capacity to contribute meaningfully both with and beyond AI.

After three decades in educational leadership — including my work in Cambridge schools and the ISB Leadership with AI programme — I've come to believe the central question for education is no longer what AI can do, but what it means to remain human alongside it. The task ahead is not simply adopting AI, but deciding which forms of intelligence — emotional, ethical, reflective, relational — schools must deliberately protect and cultivate as machines take over more cognitive work.

Most AI-enabled education platforms are already good at collecting data, tracking performance, and flagging patterns. Where they fall short is in judgment: they hand educators analytics but no structured way to interpret human experience, leaving a gap between data and decision-making.

I propose an AI–EI Leadership Framework that extends the pipeline from Data → Sentiment to Data → Sentiment → Reflection → Empathy → Action. Using ethically sourced inputs — learner reflections, teacher feedback, engagement data, and text-based sentiment analysis — the system generates reflective prompts rather than verdicts, asking better questions rather than dictating answers. Piloted conceptually through my current app development in sentiment analysis, the framework positions AI as an augmentation of judgment, not a replacement for it.

This has real implications for equity. In under-resourced settings — where trained counsellors, pastoral staff, or even reliable internet are scarce — a lightweight, privacy-conscious sentiment tool can act as an early-warning layer, surfacing patterns in learner well-being or school climate that would otherwise go unnoticed, opening space for human intervention where none currently exists.

As AI increasingly replicates cognitive tasks, education must safeguard what makes us human: empathy, ethical reasoning, curiosity, metacognition, wisdom, and relationship. Used well, AI shouldn't just make education more efficient — it should make it more deeply human, and more widely reachable.

Current EdTech paradigms fail because they treat inclusion as a content-routing problem rather than a cognitive one. Universal Design for Learning (UDL) has been digitally misimplemented as "Static UDL" – burying accessibility tools in menus and assuming students possess the metacognitive "access literacy" to deploy them. Our empirical study concerns over 107,000 exercises (targeting reading comprehension sub-skills in French as the language of literacy) completed by 340 primary students (aged 8-11) in an ecological context. It reveals a profound (in)Equity Paradox: neurodivergent and vulnerable learners use these tools half as much as their peers, driven by behavioral inertia rather than need, resulting in hundreds of invisible missed learning opportunities.

To bridge this gap, we must shift from outcome-based adaptive learning to Process-Aware Scaffolding. This presentation introduces a novel, theory-driven diagnostic architecture that makes hidden cognitive processes observable and actionable. Drawing on fine-grained telemetry from morphosyntactic reasoning tasks, we demonstrate how to computationally disentangle a student’s true underlying competency (Trait) from their momentary ability to retrieve and apply that knowledge (State). By tracking real-time action sequences, our engine infers the student's reasoning strategies and cognitive states in milliseconds, delivering explainable, targeted, fading scaffold messages that respect the student's reflection process.

Finally, we show how this architecture is scalable to amplify, rather than replace, the teacher. Through transparent equity dashboards, educators retain full decision-making power and the capacity to modify scaffolding components, ensuring human-in-the-loop pedagogical judgment while avoiding the ethical black-box of LLMs. The classroom situation becomes a “didactic pyramid” where teachers and human-centered AI-tools collaborate to foster learning for all.

Artificial intelligence is increasingly used in higher education for information retrieval, tutoring, assessment, and decision support. Less attention has been given to its potential to support reflection, personal development, and the design of meaningful lives and careers in uncertain contexts. This conceptual paper examines AI as a reflective partner within Entrepreneurial Life Design (ELD). Rather than functioning as an expert system or recommendation engine, AI is conceptualised as a dialog partner that supports reflection, perspective-taking, assumption testing, and exploration of possible futures. With its integration of ELD, career construction theory, narrative identity, positive psychology, and human–AI collaboration research, the paper introduces the concept of Hybrid Sensemaking Systems. The concept describes interactions in which human values, experiences, aspirations, and judgement are combined with AI-supported questioning, reframing, pattern recognition, and scenario generation. In this role, AI can support self-exploration, future-oriented thinking, opportunity recognition, and reflective decision-making, while responsibility for interpretation and action remains with the individual. Building on this concept, the paper proposes a framework for integrating AI into educational settings in ways that support reflection while preserving learner autonomy and the benefits of social cognition. Design principles for responsible implementation are discussed, including transparency, agency, critical engagement with AI-generated outputs, and the continuing role of human facilitation. The paper contributes to ongoing discussions about the role of AI in higher education by positioning AI not only as a tool for learning and performance, but also as a resource for reflective and developmental processes within Entrepreneurial Life Design.

The rapid adoption of generative artificial intelligence is forcing universities to reconsider not only how students learn, but also what assessment is intended to measure. Much of the debate has centred on preventing AI-assisted cheating, while comparatively little attention has been given to developing assessment methods that evaluate the capabilities becoming more valuable as AI makes information generation increasingly accessible. We argue that designathons represent a promising assessment modality because they evaluate collaborative judgement, adaptive reasoning and collective problem-solving rather than information production alone.

Drawing on observations from the London School of Hygiene & Tropical Medicine’s inaugural Sexual Health Designathon, we conceptualise the designathon as a team-based, open-book assessment in which participants respond to a novel challenge under significant time constraints. These conditions limit opportunities for extensive research or over-reliance on external tools, instead rewarding participants' ability to synthesise existing knowledge, integrate diverse perspectives, negotiate competing priorities, communicate effectively and defend decisions in real time.

Unlike traditional essays or examinations, designathons assess the process of problem-solving as well as its outcome. Participants collaborate with relative strangers, respond to evolving ideas and justify their proposed interventions to subject-matter experts acting as assessors. As a result, behaviours such as leadership, creativity, teamwork and strategic judgement become directly observable - competencies that are increasingly important in professional practice but remain difficult to evaluate through conventional assessments.

Rather than attempting to exclude AI from learning environments, designathons assess the capabilities that are necessary in AI-enabled environments: human judgment, collaboration, critical evaluation and the ability to mobilise collective intelligence. As generative AI increasingly automates information production, assessment should shift from evaluating what students can produce independently to examining how effectively they work with others, navigate uncertainty and translate knowledge into action. In an AI-enabled world, universities should reward not simply what students know, but what they are able to accomplish together.

Generative Artificial intelligence (GenAI) continues to disrupt education systems at an unprecedented rate, outpacing formal guidance and creating uncertainty. The AiEd Certified Framework (Alcock et al., 2026), developed collaboratively by academics and educators, aims to provide schools and colleges with a structured approach to GenAI implementation. In this pragmatic, practitioner-researcher study we used mixed methods to investigate whether representatives of twelve institutions who had completed Framework-based certification perceived improved confidence in AI use across their institution. All provided a written explanation of their Likert rating; these were analysed using inductive Reflexive Thematic Analysis (RTA).

Representatives of all institutions were positive about the impact of certification against the Framework on institution-wide AI confidence, built from multiple factors. The Framework served as a catalyst for strategic direction, transitioning institutions from fragmented experimentation toward formalised policies and dedicated AI leadership. Structured reflection allowed schools to evaluate current practices and identify development gaps with greater clarity. Institution-wide collaboration was essential in fostering a shared understanding of safeguarding risks and pedagogical opportunities. Practical efficiency gains were found to empower staff, allowing them to redirect time towards human-centric education. AI integration was reframed as an ongoing journey rather than a static goal, providing a foundation for institutional change. Overall, results indicate that certification against a structured Framework can be a powerful motivator for building collective confidence, and for embedding AI in ways that are human-centred, trustworthy and educationally meaningful. These preliminary findings from early adopters indicate the Framework's potential; further research is warranted and planned as more institutions certify.

Alcock, S., Goodall, C., & Russell, A. (2026). Developing an institutional AI implementation framework for English schools: a practice-informed model for systemic adoption. In EDULEARN26 Proceedings (in press). IATED. https://doi.org/10.21125/edulearn.2026 (DOI pending).

Artificial intelligence is reshaping education at speed. Touted as transformative, AI systems are being integrated into assessment, personalisation, and analytics with remarkable urgency, and remarkable unevenness. Drawing on Schumpeter's concept of creative destruction, this paper argues that AI in education does not simply innovate: it dismantles existing pedagogical, ethical, and social safeguards while failing to replace them with equitable alternatives. In their place, it often reproduces, and accelerates, the very inequities it claims to disrupt.

We introduce the concept of creative distraction to name this dynamic: the way institutional enthusiasm for AI innovation diverts attention from structural injustices affecting racialised learners, women and girls, and disabled students. Bias in educational AI is not incidental. It is encoded in training data shaped by historical discrimination, amplified by male-dominated development teams, and deployed in systems that were never designed with marginalised communities in mind. Automated essay scoring penalises linguistic diversity. Proctoring tools misidentify racialized students. STEM recommendation algorithms steer girls away from high-value pathways before they have chosen anything at all.

Drawing on critical race theory, feminist technoscience, and disability studies, this paper maps these mechanisms and their consequences, from widened achievement gaps to the psychological toll of algorithmic surveillance. It then asks what justice-centred innovation would actually require: participatory design, mandatory bias audits, transparent decision-making, and a willingness to slow down.

AI's promise is real. So is its potential for harm. The task is not to reject transformation, but to insist that it serves everyone, not just those already well served by the systems being destroyed.

Objectives and Context

Across educational settings, three words increasingly spearhead the opening charge in disputes over marks, feedback, policy and pedagogy: 'But ChatGPT said…'. Uttered by learners, parents, colleagues and institutions, the phrase can position generative AI as an apparent counter-authority against which education professionals must defend their judgement.

Professional judgement must remain open to challenge. Those making educational decisions may be mistaken or unjust, and expertise cannot place them beyond scrutiny. The concern arises when the terms of challenge delegitimise their standing as professional knowers. That standing is earned through disciplinary formation, experience, situated knowledge, duties and responsibility for consequences. Generative AI offers a shortcut to its appearance.

Approach

Drawing on Risse’s (2021) account of epistemic actorhood and Glenster’s (2025) development of epistemic rights as protections for human agency under technological conditions, this conceptual paper analyses disputes involving assessment, misconduct, parental complaints and institutional decision-making. It examines how AI-generated outputs function as apparent second opinions despite lacking professional formation, contextual understanding and responsibility when wrong.

Outcomes and Conclusions

It argues that this asymmetry cannot be understood adequately as a problem of accuracy, trust or professional autonomy alone. Education professionals remain answerable, while the systems invoked against them—and the private actors shaping what those systems can say—may remain unanswerable. These professionals sustain shared standards of evidence and inquiry across the education sector.

It therefore develops the case for epistemic rights for professional knowers. These would not confer immunity from criticism or restore unquestionable authority. They would protect reciprocal contestability: professionals’ ability to inspect and challenge AI-generated counterclaims, retain agency over decisions for which they remain responsible, and participate in governing systems that mediate their work. The right at stake is not a right to be believed, but to retain the epistemic standing and agency that make accountability meaningful.

As AI becomes embedded in education, discussions increasingly focus on what learners should know about AI, while giving little attention to what children themselves already believe about thinking, intelligence, and the human role in learning. This study asks a more fundamental question: what do children perceive as uniquely human in relationships, in reasoning, and in their own work and how do these beliefs shape their view of AI's place in education?

A multilingual, paper-based survey was administered to 120 students in Grades 5–8 across four government schools in Maharashtra, India, prior to any formal AI instruction. This paper draws on three underexamined items: an open-ended question asking what students would ask a human teacher but never an AI one; an open-ended question asking what AI cannot do however smart it becomes; and a forced-choice scenario in which students submit either their own imperfect writing or a polished AI-generated paragraph. These are read alongside a scenario item on anticipated cognitive dependency. Responses were analysed thematically, with quantitative items analysed descriptively.

Children consistently distinguished intelligence from humanity. They reserved certain questions as personal, emotional, or care-related for human teachers alone, and named qualities like empathy and personal understanding as things AI could not replicate regardless of its capability. When forced to choose between authentic-but-imperfect and polished-but-outsourced work, many privileged their own authorship, often citing concern that reliance on AI would weaken their own thinking. Rather than framing AI as threat or solution, children positioned it as a tool whose value depends on preserving human judgment and effort. These findings suggest that children already hold rich, largely unprompted, theories of human distinctiveness that should inform how educators think about assessment, teacher-student relationships, and cognitive agency in an AI-mediated classroom rather than assuming such boundaries must be taught from the outside in.

Debates about generative AI in student writing often frame AI use as shortcuts, substitution, or offloading. A recent exploratory study complicates that account by examining how neurodivergent and neurotypical writers use AI during academic composing. The project began with a 2025 autoethnographic analysis of a ChatGPT writing corpus, which examined AI as a scaffold for a neurodivergent writer. That work generated the central question for the 2026 study: if AI can support thought organization, task initiation, and revision for one neurodivergent writer, do similar patterns appear in scaled ND writing processes? To answer that question, this study shifts to student-centered empirical evidence through survey data and Think-Aloud Protocols (TAPs). It asks how AI-supported writing may operate as cognitive orchestration: a distributed process through which writers manage evaluation, revision, and authorship judgment.

The project combines survey data with TAPs from an N=100 study at Kennesaw State University. Participants were recruited from undergraduate populations through approved outreach. The survey dataset includes 77 responses to date; 25 respondents, 32.5% of the full dataset, self-identified as neurodivergent, with ADHD/attention differences as the largest subgroup. TAPs provide concurrent verbalizations and screen-recorded AI interactions during an argumentative writing task.

Preliminary analysis suggests that these students often use AI during planning and revision rather than drafting alone. Early patterns indicate that prompting functions as an external scaffold for thought organization, helping writers articulate ideas and revision needs before and during composing. These interactions suggest that prompting can make planning and self-regulation practices more observable. This study develops a research pathway for examining AI-assisted writing as cognitive orchestration and scaffolding, especially among neurodivergent writers who may use prompting to externalize thought organization, manage tasks, and make rhetorical decisions.

What is education? As a professor, I have posed this question to countless current and future educators. Still, I have yet to find a consensus definition for it. Education can be understood as imparting or acquiring knowledge, as any formative activity affecting the mind or ability of an individual, or even as the result of a good upbringing.

Language is the foundation of effective communication. However, even with a common language, misunderstandings occur ubiquitously. The assumption that individuals mean the same by using the same words is widely overestimated. In reality, the meanings we assign to words are mostly based on our own beliefs, rather than on facts.

When discussing education, some emphasize knowledge; others, the teacher; and others, the student. This idea is intricately linked to the purpose of education –which will also invariably fluctuate based on individual values and beliefs. To properly align the meaning of education and its purpose with the challenges posed by Artificial Intelligence (AI), we must first engage in self-reflection and introspection. This will allow us to shift our understanding of education away from a process of superficial formality and toward the nurturing of the forms of human intelligence relevant to an era of increasingly capable AI.

In addition to readjusting our understanding of education and its purpose, I propose three themes to be further explored in this rapidly evolving environment: first, the need to support students in the development of their critical thinking skills; second, our obligation to provide students with tools to properly question all forms and sources of knowledge; and third, our responsibility to instill in them the necessary integrity to be responsible for their own learning, minimizing the need for external regulation.

Generative artificial intelligence is frequently discussed as a tool that automates cognitive tasks, raising concerns about its impact on student agency, cognition, and critical thinking. This paper explores an alternative approach in which AI is designed not to provide answers, but to create structured environments that require learners to exercise judgment and agency under conditions of uncertainty. Drawing on game-based learning, educational scaffolding, and Johan Galtung’s conflict triangle, a customized conversational assistant was configured as an autonomous role-playing game director for geopolitical crisis simulations.

The intervention was implemented with 101 undergraduate students who participated in sixturn simulations requiring them to manage complex conflicts through constrained policy choices, probabilistic outcomes, and competing ethical priorities. A mixed-methods evaluation combined Likert-scale survey data with qualitative reflections. Findings indicate that students perceived the simulation as highly effective in exposing them to moral ambiguity, systemic constraints, and the unintended consequences of decision-making, thereby promoting forms of higher-order cognition and reflective agency. Qualitative responses suggest a shift from binary understandings of success and failure toward more nuanced conceptions centered on stabilization, harm reduction, and the management of structural violence. Participants increasingly viewed conflict management not as the pursuit of optimal solutions, but as the navigation of uncertainty, competing priorities, and unavoidable trade-offs.

The study argues that AI can function as a mechanism for productive cognitive friction, compelling students to confront uncertainty rather than bypass it. Rather than diminishing learner agency, the simulation required participants to continuously make decisions, adapt strategies, and reflect on the relationship between values, institutions, and outcomes. These findings contribute to emerging discussions regarding cognition, agency, and epistemic thought in AI-mediated learning environments, suggesting that carefully structured AI systems may foster forms of epistemic humility, reflective judgment, and human intelligence that are central to higher-order learning.

As generative AI becomes embedded in recruitment, career education faces a question of human capability as much as access: how can students learn to use AI without losing the voice, judgement and self-authorship that make career development meaningful and distinguishable to employers? This paper reports early findings from an embedded case study with non-traditional undergraduate and postgraduate business students in an Irish university.

The study designed, implemented and evaluated a universally designed co-curricular AI for Career Management module. The module was built on a multi-level behavioural competency framework and PROVE process (Purpose, Redact, Ownership, Verify, Explain), both developed by the researcher to support responsible AI use. Students completed pre- and post-competency self-assessment, career tasks, ethical scenario judgements, career artefacts and reflections. The design was informed by focus groups and interviews with faculty and programme staff in the case-study business school (n=9), international business school career educators (n=30), graduating students and alumni (n=8), and ten employer interviews across sectors. All students in the in-scope cohorts were invited to participate, with 20 completing the module, submitting artefacts, and taking part in interviews.

Initial analysis suggests that students were either using AI uncritically or avoiding meaningful engagement because they lacked a responsible method for applying it to career development. Interviews and reflections indicate that the module made this gap visible, increased students’ confidence, and helped them articulate how they would apply PROVE in future career tasks. Stakeholders emphasised ethical use not only to protect student voice and individuality but also to support honesty, integrity, confidentiality and responsible handling of data. Engagement patterns frame time as an equity issue: part-time students reported persistent time pressure, while undergraduate participation clustered outside teaching term for a non-credit activity. The paper invites reflection on career education’s role in developing human judgement for AI use.

Generative AI is increasingly entering assessment design in higher education, but much available support emphasises either policy compliance or rapid production of assessment artefacts. This paper presents ADA, an Adaptive AI Agent for Assessment Design, as an exploratory design-based study of how AI-supported faculty development might preserve the judgement-rich work through which academics connect learning purposes, disciplinary standards, student experience, validity, fairness, workload and institutional rules. ADA was developed around a Collaborative Assessment Design Decision-Making Map that structures interaction across four interconnected decision spaces: purposes and constraints; design alignment and stakeholder reasoning; validity, fairness and integrity; and artefact building and review. The study compares two versions of ADA across five simulated assessment-design scenarios: ADA A, which embeds the Map, and ADA B, a comparable general-purpose assessment-design facilitator using the same model and knowledge base. Sentence-level analysis examined decision-space engagement, transitions across spaces and dialogue acts. ADA A showed greater engagement with design alignment and validity-related reasoning, more frequent movement across decision spaces, and a facilitation pattern marked by more recommendations, validation, summarisation and procedural direction. ADA B relied more heavily on probing and elaboration. These findings do not demonstrate improved assessment quality or direct effects on faculty practice. They provide provisional evidence that map-structured GenAI facilitation can make assessment-design reasoning more navigable, inspectable and reviewable. The paper argues that future faculty development should focus not only on helping academics use AI tools, but on designing sociotechnical conditions that support academic judgement with AI.

Objectives and context

While AI tools are being rapidly introduced into classrooms, there is little evidence demonstrating a causal link between tutoring based on large language models (LLMs) and learning outcomes in primary education, especially within non-English-speaking contexts. This paper presents the development and evaluation of a teacher-mediated AI tutor for Computational Thinking (CT), deployed within Uruguay's national CT and AI programme at Ceibal, which serves approximately 30,000 sixth-grade students in public schools. The work moves across three registers: constructive design, baseline measurement, and experimental evaluation.

Methods and approach

The tutor was built iteratively by Ceibal’s educational team, based on research and direct work in the field. An earlier review categorised generative AI applications into four instructional functions and documented gender gaps within Ceibal’s digital ecosystem (Levy Yeyati et al., 2025).In 2025, a preliminary phase to the 2026 pilot project was carried out, during which the chat was iterated, classroom observations were conducted, interviews with teacher pairs were held, and more than 2,000 natural interactions between students and AI were recorded—revealing that, without explicit instruction, participation is uneven: approximately one-third of the sessions were technical in nature, another third were social or off-topic, and a smaller group engaged in productive and sustained debugging. Teachers, unprompted, identified learning to ask good questions as the central student gain. Baseline data from 2026 sharpen this picture. Asked how they would prompt an AI tutor to solve a CT scenario, 33% of students (n=340) could not produce a relevant question, 47% asked something vague, and only 21% framed the problem with genuine precision — despite two-thirds recognising the chat as a potentially useful learning tool. In the 2026 trial, 126 volunteer teachers enrolled and were randomly assigned at the classroom level: 58 to the treatment group and 68 to the control group. Classes in the treatment group followed a 10-week Scratch programming sequence with structured access to a teacher-mediated AI tutor; classes in the control group followed the same sequence without such access. CT learning is measured using an adaptive assessment calibrated according to the TRI method at the beginning of the study and an aligned post-test instrument. Additional results include motivation, interest, self-efficacy, and the quality of instructions; subgroup analyses by gender, socioeconomic background, and prior performance have been previously reported.

Outcomes and conclusions

Pre-pilot observation and baseline data converge on the same finding: most students arrive knowing they need help, but not knowing how to ask for it. This gap — between recognising the chat as useful and being able to use it productively — points to prompt quality as a meaningful outcome in its own right, not merely a proxy for engagement. Whether a teacher-mediated AI tutor can develop that capacity is the trial's central question: does structured instructional use of an AI chat tutor improve CT learning, and does it shift the way students interact with AI — from delegation toward genuine inquiry? Post-test data collection runs through July 2026; preliminary trial results — ITT estimates on CT learning, non-cognitive outcomes, and prompt quality improvement — will be presented at the summit, offering early causal evidence on whether structured teacher mediation can turn that recognition into real capability.

Debate about artificial intelligence in education has centred largely on adoption, integrity, and productivity. This presentation reframes the question through the lens of existential risk. Alongside its benefits, advanced AI is increasingly recognised by researchers as a potential catastrophic and even existential threat, whether through loss of human control, the dangerous concentration of power, or its capacity to amplify other hazards such as engineered pandemics. If education exists in part to prepare each generation for the world it will inherit, a serious question follows: what human capabilities must education cultivate so that young people can understand, deliberate on, and help to mitigate and navigate the risks that frontier AI poses?

Drawing on Sen's and Nussbaum's Capabilities Approach, the study treats education as the expansion of the real freedoms people have to think, judge, and act. This presentation draws on the first phase of doctoral research, now underway: structured interviews with a cohort of approximately forty experts in AI and existential risk, in which participants rate the severity of distinct AI risks and the relative importance of candidate capabilities for navigating them.

Interviews are currently underway, with analysis to be completed ahead of the summit. The presentation will therefore report the study's first emerging findings: which capabilities this expert cohort judges most important for navigating AI risk, how far they converge, and where they disagree - including on how much of this responsibility formal schooling can realistically carry.

This work speaks directly to the summit's questions about educational purpose, frontier technologies, and the protection of independent thought. It offers an evidence-informed basis for one of the gathering's central concerns: what forms of human intelligence education should most seek to protect, cultivate, and assess.

Advances in large language models have significantly expanded the generative capacity of educational systems. AI can now produce explanations, hypotheses, and professionally fluent texts at scale, often outperforming humans on measures of pattern-based reasoning. Much of the current debate has focused on adoption, productivity, and academic integrity. However, a more fundamental ethical question remains underdeveloped: where does responsibility reside when educational thinking is increasingly co-produced with machines?

This paper argues that AI has not solved intelligence, but selectively amplified competence while leaving judgment, evaluation, and ethical accountability unresolved. In practice, this creates a structural imbalance: fluency accelerates faster than defensibility, and agreement often replaces challenge. The resulting risk is not error alone, but the production of work that appears credible while obscuring the conditions under which it was generated.

Drawing on examples from SEND systems, professional documentation, and AI-assisted reasoning, the paper develops a model of distributed ethical responsibility. Here, the asymmetry at the heart of education does not disappear but is reconfigured: responsibility is shared across practitioners, system designers, institutions, and regulatory frameworks. Yet this distribution risks becoming dilution unless the locus of accountability is made explicit.

The paper argues that the human role must be reasserted not as primary generator of knowledge, but as ethical anchor: the one who is answerable for decisions, who stands in relation to the learner, and who must interpret, challenge, and, where necessary, refuse machine-generated outputs.

A practical implication follows: AI use in education should be structured to separate generation from evaluation. By designing explicit “regulatory phases” that audit groundedness, evidence, and constraint, we can reintroduce ethical friction into systems optimised for fluency.

The central claim is that AI systems do not remove the ethical burden of education rather, they redistribute it. The critical question is whether, in doing so, they clarify responsibility or render it harder to see.

Many students from disadvantaged backgrounds do not receive the same support as their peers when preparing for medical school interviews. Applicants from well-resourced schools access coached mock interviews through family and school networks, while applicants eligible for free school meals (FSM) often prepare alone. This study evaluates AVA, an online AI platform on which students practise answering interview questions and receive feedback, and asks whether AI-based preparation narrows, leaves unchanged, or widens gaps in admissions outcomes.

The study is a multi-institution randomised controlled trial, co-led across partner institutions including Anglia Ruskin Medical School. All students shortlisted for interview at ten UK medical schools are invited to participate. Consenting applicants, approximately 5,000 across cohorts, are individually randomised at the point of interview invitation to AVA or an active video-course comparator. Five medical schools hold day-one data-sharing agreements returning objective, binary offer outcomes linked to UCAS ID, addressing the attrition and survivorship bias that have undermined self-report evaluations in this field. A parallel Sutton Trust longitudinal cohort of 500 widening participation students uses AVA across the full application cycle, completing repeated self-efficacy surveys.

The analysis is pre-specified around equity. The primary question is whether treatment effects differ by FSM eligibility; secondary axes include first-generation status, school type, POLAR4/IMD, ethnicity, and gender. We report no effectiveness results here: full-cycle outcomes are still being collected, and these questions remain open by design.

The pilot cycle demonstrates feasibility. Recruitment at invitation, centralised UCAS-ID randomisation, and institutional return of objective outcomes all functioned at scale. The evidence so far is very promising for the viability of objectively measured, equity-focused evaluation of AI in university admissions.

Human capability has historically developed through repeated inquiry, problem solving, evidence evaluation, and reflection. These activities were not simply mechanisms for completing work; they were the processes through which judgment, expertise, and professional capability emerged. Generative artificial intelligence fundamentally changes the environment within which those capabilities develop.

Building upon Simon and Newell's process-oriented theory of human problem solving, this paper argues that AI represents more than a new technological tool. It changes the conditions under which cognitive effort is allocated. As increasingly capable AI systems assume tasks such as researching, drafting, synthesizing information, and generating ideas, many cognitive activities that were once necessities become choices. The central question therefore shifts from what AI can do to how humans exercise agency when cognitive effort becomes increasingly optional.

The paper introduces the conceptual framework Cognitive Necessity → Cognitive Choice, proposing that the defining consequence of generative AI is not simply increased productivity but a fundamental transformation in the processes through which human capability develops. Rather than treating AI adoption as an end state, the framework positions AI as a change in the environment that reshapes inquiry, judgment, reflection, evidence evaluation, and responsible delegation. Human agency becomes the critical mechanism through which individuals decide when to think independently, when to verify information, when to challenge AI-generated outputs, and when to delegate cognitive work. These choices ultimately shape the development of human capability.

Although the empirical work emerging from this framework is situated within higher education, the conceptual contribution extends more broadly to organizational learning, workforce readiness, leadership development, and public policy. The paper argues that AI-mediated environments require new ways of understanding not only AI systems, but also the conditions under which human capability and agency continue to develop.

By reframing AI as a change in the environment rather than simply a change in technology, this work positions human capability and agency as foundational questions for human-centered AI. Understanding how people develop judgment, inquiry, and epistemic responsibility under changing technological conditions may become one of the defining scientific and societal challenges of the AI era.

As machines increasingly perform, support, and simulate aspects of human thought, education has a timely opportunity to clarify what it most values and why. This contribution offers Critical AI Literacy (CAIL) (Hauck, Moore & Wright, 2025) as a framework that approaches that question on educational rather than technological terms and through a social justice and inclusion lens. It treats human intelligence as relational, situated, and culturally grounded.

CAIL extends AI literacy beyond competent use towards critical engagement: understanding how generative systems are built, whose knowledge they encode, and what educational purposes they serve. A defining dimension invites learners and educators to interrogate "AI" itself as a concept, recognising that naming shapes expectation and agency. CAIL also positions informed choice – including the decision to work without AI – as a characteristic of literacy and judgment, restoring human discernment to the centre of learning.

The framework draws its ethical foundations from relational traditions such as southern African ubuntu ("I am because of you"), which locate intelligence, wellbeing and worth in human connection and mutual responsibility. The value education should protect is the human capacity to reason, relate, and act with care – capacities that flourish through dialogue and shared inquiry across cultures.

These commitments take form in a four-day, university-wide Promptathon that engaged 157 participants and 16 facilitators in interdisciplinary teams across four rounds, each mapped to a CAIL theme: surfacing embedded bias, communicating accessibly, interrogating AI's social and environmental costs, and designing AI-resistant and inclusive assessment. Analysis of submissions, annotations, and a post-event survey found bias awareness dominant, attention shifting from crafting prompts to evaluating outputs, and collaboration driving the richest learning. This conference session offers the case as an approach to curriculum, assessment and educator development, inviting a shared conversation about an education that keeps human purpose and relationship at its heart.

This talk presents findings from a literature review of symbols and AI as supports for inclusive communication, and a forecasting exercise drawing on interviews, focus groups and workshops with teachers and SENDCos.

There is a renewed focus on Oracy in the UK and on dialogue and human interaction and collaboration more broadly. This necessitates approaches for including learners with communication differences and preferences, including those with auditory impairments, children who stammer, children not fluent in the language of instruction and children with autism.

Novel AI tools might support the inclusion of some learners, but it is not yet always clear how they might interact with traditional supports, such as symbols and visual icons. Additionally, they raise a range of questions related to access, efficacy, data security privacy and consent.

I present the imaginaries of teachers and SENDCos, I argue for the importance of including teachers and learners in foresight work on the desirability and risks of emerging technologies with the potential for educational application before they become ubiquitous.

In this presentation we describe research that investigates how examiner behaviour in digital marking can be inferred from data on response processes, rather than relying on self-reported marking “strategies” (Suto and Greatorex, 2008). We capture marker behaviour 'between the clicks', with eye-tracking and pupillometry data from 25 participants completing 30 marking trials. We modelled attention and judgement with semantic context behaviours with GNN (Graph Neural Network) and T-based architecture. Model outputs were synthesised with Retrospective Think Aloud (RTA) data under a mixed method methodology, to infer behaviours behind the Deep Learning gaze-based states.

Our extension of data, analysis, and consideration of intelligence beyond human perception reveals micro-behaviours shaped by interface design and marking content. We contrast that analysis with self-report evidence from more conventional retrospective think aloud (RTA) protocols. We will present our findings, and reflect on the methodological and ethical potentials and challenges that the approach presents. Those include - the potentials to work with AI to improve human on-screen experience, and the challenges of using data and inferences that – while they are ‘explainable’, are beyond human perception.

We discuss the research-oriented challenges of outsourcing epistemological judgement to AI models and their treatment of sensory data in the context of contrasting marker judgements, prior experience, and human diversity. We draw conclusions on the ways that improved Human–AI Interaction for assessment might reduce marker cognitive load, to support fairer, more reliable assessment. The findings contribute to a wider research initiative on Human-AI interaction under the EPSRC Prosperity programme, involving a collaboration between AQA and Kings College London.

Objectives

A graduate student building a donor segmentation model produced a well-performing classifier. The AI-generated code used zip code as a key predictor, encoding residential segregation into the fundraising strategy. The technical output was flawless. The judgment required to identify why it was inadequate was not technical, and it was not something the AI could supply. Such cases are not exceptional; they are the defining pedagogical condition of graduate analytics education. When AI output is accepted uncritically, what is compromised is not accuracy alone; it is the learner’s epistemic agency: the capacity to form independent, warranted judgments. That is an epistemic rights problem as much as a pedagogical one.

Approach

The paper uses practitioner inquiry (Cochran-Smith & Lytle) and case study analysis (Yin), drawing on two anonymized cases from the Master of Science in Analytics at Northeastern University. Both share a structural feature: AI produced output that was well-formed and confidently wrong in ways requiring human judgment, judgment that was ethical and contextual rather than technical. Working professionals with real organizational stakes extend the analysis beyond the residential undergraduate setting that dominates this literature.

Conclusions

Two findings follow. First, technical and humanistic formation are not separable in professional analytics education: teaching students to evaluate AI output is teaching epistemology; asking whose interests a model serves is teaching political philosophy; explaining a technical finding to a non-technical stakeholder is teaching rhetoric. These capacities are not a curricular choice but a structural feature of what competence requires. Second, AI has not created this need; it has made it legible. Bureau of Labor Statistics data show programmer employment fell 27.5 percent between 2023 and 2025, removing institutional cover for treating credential attainment as a proxy for genuine formation. Assessments worth designing require students to show reasoning against outputs they did not generate.

Objectives and Context

Educational AI is increasingly used to support assessment, student interventions, admissions, and other consequential decisions. Governance approaches often assume that keeping a human "in the loop" preserves accountability and professional judgment. This presentation challenges that assumption and introduces the concept of Human Oversight Theatre: situations in which formal human oversight remains in place while independent judgment gradually diminishes through reliance on AI recommendations.

Methods / Approach

Drawing on practical work in AI governance, accountability design, and organisational implementation, the presentation examines how Human Oversight Theatre can emerge in educational settings. It explores three mechanisms that contribute to this phenomenon: automation bias, procedural review practices that become routine rather than reflective, and the gradual erosion of professional decision-making capabilities through repeated dependence on AI-generated outputs.

Outcomes and Conclusions

The presentation argues that the most significant risk may not be the absence of human oversight, but the illusion of it. As AI systems increasingly shape recommendations, evaluations, and decisions, educators may retain formal responsibility while exercising progressively less independent judgment. This raises important questions about professional agency, intellectual autonomy, and the role of human judgment in education. The session proposes that educational institutions should evaluate oversight not by the presence of a human reviewer, but by the extent to which meaningful human judgment remains active within the decision-making process. Human Oversight Theatre offers a new lens through which to examine the relationship between AI, accountability, and the preservation of human agency in education.

In the context of private English as a Second Language (ESL) education within the Hong Kong K-12 tutorial support sector, a persistent tension exists between maximizing instructional personalization and maintaining operational viability. While traditional 1:1 tutoring offers superior pedagogical quality, it presents severe constraints regarding financial scalability. Grounded in Vygotsky’s scaffolding theory, this study investigates a hybrid, university-style elective model designed to deliver affordable, high-fidelity personalized learning. The objective is to evaluate how integrating AI-driven diagnostic tools with human-centric professional guidance can optimize student learning pathways while preserving the essential human connection necessary for language acquisition.

The study deployed a five-stage blended learning ecosystem. First, students completed an automated AI diagnostic assessment to baseline their proficiency. Second, learners were clustered into homogenous ability bands. Third, a modular, single-focus curriculum (targeting discrete competencies in reading, writing, listening, or speaking) was delivered via synchronous instruction, supplemented by asynchronous digital repositories. Fourth, a dynamic "co-scaffolding" mechanism was introduced: students autonomously selected their learning pathways based on AI metrics, actively mediated by personalized teacher consultations. Finally, periodic AI-driven re-assessments triggered adaptive group realignments. Within this framework, AI managed standardized diagnostics, liberating human educators to focus on empathetic mentorship and strategic pathway counseling.

Empirical tracking of this hybrid model demonstrated a 20% increase in students' academic English performance, accompanied by a 50% reduction in operational delivery costs and a 70% increase in student learning motivation. Furthermore, qualitative evaluation revealed an exceptional parental satisfaction rate exceeding 90% regarding structural flexibility and individualized attention. These empirical insights suggest that the automation of routine diagnostic assessments does not erode the human dimension of instruction. Instead, this delegation structurally optimizes the interpersonal connection by elevating the practitioner from a procedural evaluator to an indispensable pedagogical facilitator, thereby harmonizing scalable personalized learning with operational sustainability.

Objectives and Context

Higher education institutions are rapidly developing frameworks, policies, and guidance to support the integration of generative artificial intelligence into teaching and assessment. However, the success of these initiatives depends not only on institutional strategy but also on the capability of academic staff to interpret, implement, and operationalise new approaches within disciplinary contexts.

This paper reports on an institutional research project accompanying the implementation of Birmingham City University's GenAI Integrated Assessment Framework. The framework was developed to support assessment redesign through principles of authenticity, transparency, visible learning, professional relevance, equity, and responsible AI use. The study investigates how staff understand these principles and examines the capabilities required to translate institutional aspirations into educational practice.

Methods / Approach

The project adopts a qualitative case study design. Data are collected through cross disciplinary staff focus groups and interviews exploring perceptions of AI, assessment redesign, confidence, professional judgement, and implementation challenges. Documentary analysis of institutional policies, assessment guidance, and framework materials provides additional insight into how responsible AI principles are communicated and enacted. Thematic analysis is used to identify patterns relating to capability, governance, decision making, and assessment practice.

Outcomes and Conclusions

Preliminary findings suggest that responsible AI implementation requires capabilities extending beyond technical AI literacy. Staff identify assessment literacy, pedagogical judgement, ethical reasoning, disciplinary expertise, and confidence in redesigning learning as critical factors influencing implementation. The study highlights tensions between institutional ambition and local practice and demonstrates how capability gaps can influence consistency, equity, and assessment quality.

As AI becomes embedded in education, there is a need to ensure that opportunities to develop AI literacy are accessible to all learners. However, little is known about how home-educated neurodivergent learners engage with AI or the factors shaping equitable participation.

This talk presents findings from Talking Itchen, an eight-session outreach programme centred on the River Itchen, one of England's internationally important chalk streams, where nineteen home-educated neurodivergent learners aged 7 – 15 explored how artificial intelligence and electronics can support environmental monitoring and community action.

The programme introduced AI through hands-on activities centred on the River Itchen, using environmental conservation as a context for learning about AI, electronics and public engagement. While designed for young people, the programme also engaged parents, whose perspectives evolved alongside their children's and helped shape attitudes towards AI. Insights were drawn from observations, participant discussions and conversations with parents.

Both learners and parents initially expressed scepticism towards AI, often citing concerns about misinformation, environmental sustainability and broader societal impacts. However, these perceptions shifted through participation, suggesting that authentic, purpose-driven experiences can foster more positive and informed engagement with AI. The study also highlights the role of parents in shaping children's attitudes towards AI and identifies barriers to equitable engagement, including limited guidance for home-educating families and unequal access to resources.

Although drawn from a small-scale outreach programme, the findings suggest that AI education requires more than equitable access to technology. Building trust, connecting AI to learners' interests and communities, and providing opportunities for meaningful application may be equally important. The findings offer practical insights for educators, outreach practitioners and policymakers seeking to design more inclusive approaches to AI education, and informed published evidence submitted to the UK Parliament Education Committee's inquiry into the use of artificial intelligence and educational technology in education.

Objectives and context

AI-based marking can improve the speed and scalability of educational assessment, but widespread adoption requires more than high average accuracy. Performance metrics summarise behaviour across large numbers of responses, but operational deployment requires evidence about the reliability of individual marking decisions. Reliable confidence estimates are therefore essential for determining when an AI-generated mark can be accepted and when a response should be routed for human marking.

This paper presents a study of confidence calibration for automated short-answer grading within Cambridge University Press & Assessment's Digital Mocks Service, which uses questions from previous high-stakes examinations.

Methods / Approach

For 107 short-answer items, BERT-based classifiers were trained using approximately 1,600 examiner-marked responses from the corresponding high-stakes examination. Each model was trained to predict the mark awarded by an examiner. The models were subsequently applied to 85,340 Digital Mocks responses, alongside full human examiner marking.

We investigated whether model confidence scores could be transformed into reliable estimates of the probability that an AI mark would agree with a human mark for a Digital Mocks response. Three calibration approaches (Platt scaling, isotonic regression and temperature scaling) were evaluated at three levels of granularity. Calibration models were fitted using held-out examination data and evaluated on Digital Mocks predictions.

Outcomes & Conclusions

Raw model confidence scores from the BERT classifiers were systematically overconfident. Among the approaches evaluated, item-level temperature scaling best reduced calibration error and produced confidence estimates that more accurately reflected the probability that an AI mark matched an examiner mark.

These findings suggest that accuracy alone is insufficient for operational AI-first marking. Confidence calibration provides a practical mechanism for distinguishing between responses that can be marked automatically and those requiring human review, enabling more efficient deployment while maintaining appropriate human oversight.

Despite rapid AI adoption in education, rigorous evaluation of AI-powered educational (AIED) systems remains critically underdeveloped, particularly in second language (L2) education, one of the most common yet least evaluated AI applications. We introduce L2-Bench, an open-source benchmark of 1,000+ task-response pairs to aid the pedagogy-led evaluation of LLM capabilities relating to language learning and assessment. Crucially, L2-Bench measures model performativity on the application of learning experience design principles rather than mere knowledge of those principles or broad learning outcomes. Our contributions include: (1) a validated taxonomy of 12 competencies and 31 subcompetencies validated by 200+ expert practitioners (task authenticity = 4.42/5.00, criteria adequacy = 4.18/5.00); (2) a rubric-based evaluation methodology that we believe can, if adapted, generalize to similar (open-ended, qualitative) disciplines; (3) an evaluation dataset that produces reliable signal about model strengths, weaknesses, and contextual robustness across diverse L2 education scenarios. We find that, among large models, Claude Opus 4.7 performs best overall (85.5%), though is marginally outperformed on several constituent tasks. We also find that performance drops notably on harder tasks ∈ (69.9%–73.4%). L2-Bench provides education stakeholders better methods to make more informed decisions about real-world AIED adoption, use, and governance, while advancing the maturing science of AI evaluations for education.

Objectives and context

When generative AI can explain, assess, and persuade, the question “what is essentially human in learning?” no longer has an obvious answer. This paper rejects the defensive reflex of fencing off individual cognition from the machine. It asks what learning is once AI is inside it and argues that judgment, discernment, and ethical reasoning are not possessions of separate minds but properties of the relations in which those minds are formed. Learning, on this account, is not something a person does; it is something that happens between.

Methods / Approach

The argument rests on three qualitative studies, selected for their minimal overlap. The first examined relations among 29 students, two faculty members, and a Socratic AI tutor delivered via WhatsApp in a low-bandwidth Kenyan teacher-education course. The second study examined what emerged between fifty doctoral students and a deliberately scaffolded AI “critical friend,” with the work’s analytic core walled off. The third followed doctoral researchers from underrepresented backgrounds, asking whether AI widens or narrows equitable participation. In each, the unit of analysis is the relation, not the individual.

Outcomes & Conclusions

The same pattern surfaced in all three. Learning was located in no one: it did not reside in the student nor pass from the machine, but occurred as the relation itself — in the contesting, resisting, and reorganizing through which thinking became possible. Because the pattern holds across a Kenyan classroom and a doctoral researcher’s guarded analytic space alike, convergence rather than any single case carries it. I name this learning ontological emergence: a relational event that reconstitutes what its participants are becoming. Education’s task shifts accordingly, from protecting the bounded individual to cultivating the dispositions, attunement, humility, and reciprocity that enable anyone to participate well in relations they do not own.

This presentation argues that AI should not be adopted primarily to strengthen the existing education system: more content delivery, more assessment, more surveillance, and more pressure on students. Instead, AI should be guided by a learning-centred framework that restores education to its original purpose: drawing out students’ intelligence, agency, curiosity, judgement, and creativity.

To support this argument, the presentation builds on a case study from schools in Shenzhen, China. These cases illustrate how AI can support human-centred problem-based learning. Through the case study, the presentation uses Liz Kolb’s Triple E Framework (Engage, Enhance, Extend) to evaluate to what extent AI can be used to genuinely support learning goals rather than simply adding another layer of digital tools to existing instruction. The framework shifts attention away from mainly teaching students how to use AI tools technically and toward a more fundamental question about augmenting human intelligence: does AI help students ask better questions, think more deeply, collaborate more meaningfully, and express understanding in richer and more creative ways?

Importantly, the presentation further argues that, while the Triple E Framework provides a valuable starting point, it should be revised and updated for the AI era. Previous digital tools could engage, enhance, and extend learning, but AI makes possible a fourth E: Elevate. This fourth dimension captures what earlier technologies could not fully achieve: the capacity of AI to expand students’ intellectual agency, creative power, ethical judgement, and autonomy as learners. To elevate is to raise both the education system and human capacities beyond the existing limits, boundaries, and scope, rather than simply making that system more efficient. In this regard, AI should not replace learning or automate schooling. It should help return education to the hands of students, enabling them to become more active, reflective, and capable agents of their own learning.

Objectives and context

Automated systems for assessing second language (L2) writing proficiency have been studied for decades. With the advent of transformer-based architectures, these systems have achieved impressive performance in holistic assessment (i.e., assigning a single overall score to a learner's writing), in some cases even surpassing human-human agreement. More recently, research has shifted its focus towards analytic assessment (i.e., assigning separate scores for different dimensions of writing proficiency) and the related topic of feedback. This task is considerably more challenging due to the intrinsic correlations among analytic dimensions and the limited availability of annotated data for training automated systems. Large language models (LLMs) offer a promising alternative, for a practice test to help learners improve, as they can leverage their emerging capabilities without requiring task-specific fine-tuning.

In this work, we investigate different strategies for LLM-based feedback for L2 learners, using essays from a popular test of L2 English as a benchmark.

Methods

As well as directly prompting LLMs, we explore several alternative approaches, including natural language-based methods and techniques inspired by item response theory and cognitive diagnostic modelling. All these approaches rely on highly interpretable natural language descriptors, expressed in the form of can-do statements, that are strongly grounded in pedagogy. We evaluate our methods at both the inter-learner and intra-learner levels. The latter, which remains largely underexplored in previous research, is particularly important for providing meaningful diagnostic feedback.

Outcomes & Conclusions

We will show results about the potential of these approaches for producing feedback and analytic scores. Beyond predictive performance, we discuss the implications, limitations, and potential risks of deploying LLMs for language learning and related purposes, with particular attention to data privacy, transparency, and fairness.

This proposal examines how responsible AI implementation can extend, rather than replace, human expertise in educational assessment. Drawing on three years of work by Cambridge University Press & Assessment with an external provider, it evaluates AI-assisted auto-marking within the Cambridge Digital Mocks Service as a practical case of human-centred innovation.

Across successive phases, AI-generated marks were produced in parallel with examiner marking, using models built primarily from mark scheme information. The work increased in scale and complexity, covering a wider range of item types and subjects while retaining examiner judgement as the reference point for quality, trust and accountability.

The most recent phase focused on higher-tariff responses and created new possibilities for faster feedback, particularly in contexts where examiner capacity may be limited.

Responsible deployment was governed through layered quality assurance. Every response was double-marked by both an examiner and the auto-marker, enabling response-level comparison. Agreement was evaluated using multiple statistics, including quadratic weighted kappa, with values above 0.7 observed in the latest phase. A tolerance rule of +/- one mark allowed approximately 37% of responses to receive the AI-generated mark only when the quality threshold was met.

Operationally, the workflow was successful, with no significant technical barriers to deployment and positive feedback from participating examination centres. Importantly, the model positioned AI as a potential support for examiners’ expertise: automating lower-risk decisions, preserving human oversight, and focusing examiners’ attention where it adds greatest educational value.

Ongoing work focuses on deeper evaluation of the validity of AI-generated mark justification statements and formative feedback for learners. The presentation will argue that the responsible implementation of AI in assessment depends on transparency, continuous validation and clear human accountability. Used in this way, AI can open the way for scalable, timely and trustworthy digital assessment, enabling richer learning opportunities without compromising standards.

Generative artificial intelligence has created profound challenges for assessment, academic integrity and literacies in higher education. Systems whose fundamental operation relies on probabilistic next-token prediction are increasingly capable of producing outputs that resemble critical analysis, evaluation, synthesis, and other complex written communication. As universities seek to redesign assessment in response, debates frequently turn on fundamentally different assumptions about what generative AI is and is not.

These assumptions are frequently expressed through metaphor. Educators, policymakers, and researchers routinely describe AI as an “autocomplete on steroids”, a “calculator”, a “ghost”, a “copilot”, or a “co-intelligence”. While such metaphors help make a complex technology accessible, they also shape educational judgements concerning assessment validity, academic integrity, and the continuing significance of human intellectual capabilities. Different metaphors imply different responses: if AI is an advanced tool, assessment design may focus on effective use; if AI approximates human cognition, more fundamental questions arise about what remains educationally significant.

However, all current metaphors fail to adequately explain how apparently simple probabilistic mechanisms can generate increasingly sophisticated outputs, sometimes surpassing humans. In this presentation we identify three established categories - diminishers, utilities, and partners - and analyse the insights and limitations of each, before introducing a new category, transformative metaphors, which focuses on the relationship between underlying mechanisms and emergent capabilities. Examples including "clay and sculpture", "erratic factories", and "knives for everyone" illuminate processes of transformation, opacity, capability, and risk without relying on direct comparisons to human intelligence.

Assessment reform requires more than practical guidance on AI use; it requires conceptual clarity about the nature of the technology itself. By making explicit and enhancing the metaphors that underpin educational decision-making, the proposed framework offers a foundation for more coherent discussions of assessment design, likewise academic integrity, AI literacy, and educational purpose in an age of generative AI.

Objectives and context

Neil Postman argued the telegraph, photograph, and television fractured public discourse into a “peek-a-boo world,” where information appears, then vanishes, leaving a culture entertained but unable to act on what it knows. Conversational AI creates an analogous fracture, turned inward, in the record of a researcher’s own questions, abandoned paths, and insights. On platforms including Claude and Copilot, search reaches only an automatically generated title, not the content of what was said, and asking the AI to recover a past conversation tends to generate a new exchange instead, leaving the record more muddied than before. ChatGPT fails differently: it can turn a detail from one conversation into a stored fact, shaping later replies without the user knowing why. In each case, the platform does something with the researcher’s own work that cannot be fully seen or controlled.

Methods / Approach

The paper extends Postman’s framework to a more individual loss: knowing oneself by seeing and building on one’s own past thinking. Through documentation of the author’s own use of three widely available platforms, Claude, ChatGPT, and Copilot, it identifies a distinct failure in each. A working archiving system, built by the author, serves as a counter-case, showing the gap is a design choice, not a technical limit.

Outcomes & Conclusions

Microsoft’s Copilot, widely available to higher education, once let individuals export their conversations and removed it; what remains is available only to administrators. The architecture imagines the institution, not the researcher, as owner of the record. Education has a stake in cognitive sovereignty: the principle that learners should own, control, and verify their own AI-assisted thinking. Two measures follow: treating retrievability as a criterion when adopting tools, and teaching learners to preserve their own record and distinguish their thinking from what a system has quietly added.

Objectives and context. This paper examines how personal AI knowledge twins can be responsibly integrated into (legal) education in ways that are human-centred, trustworthy, and educationally meaningful. The study is based on a small-scale pedagogical experiment in a data protection law course which has been taking place during spring semesters at Vilnius University Law faculty since 2024, where two lecturers introduced conversational AI twins trained on their own academic materials, together with other generative AI tools. The objective was not to replace the professor, but to extend access to expert guidance, support student learning, and create a supervised and well-structured environment for developing AI literacy, critical reflection, and AI skills that would prepare the students for a legal career where AI tools are integrated in different stages of daily tasks.

Methods / Approach. The course design combined three elements: (a) classroom teaching and guided AI use, (b) hackathon style collaborative workshops where students interacted with personal AI twins, and (c) “open-to AI” exam tasks requiring transparent documentation of prompts, outputs, and reasoning processes. Students were encouraged to consult AI tools critically rather than passively accept their outputs. The analysis draws on student surveys, course evaluations, and instructors’ reflective practice to assess how the integration of AI tools affected learning, anxiety, engagement, and critical and creative thinking.

Outcomes & Conclusions. The experiment suggests that carefully designed AI integration can support more human-centred education. Some students reported reduced anxiety, deeper engagement, and greater confidence in exploring complex legal concepts. The AI twins enabled classroom time to be redirected toward discussion, interpretation, and critique, while assessment shifted attention from final answers to documented learning processes. The findings also highlight the importance of transparency, teacher guidance, institutional trust, and clear pedagogical purpose. We conclude that responsible AI implementation in education requires more than access to tools: it requires governance, AI literacy, reflective documentation, and pedagogical designs that keep human judgment, creativity, and accountability at the centre of learning.

Jonathan Sansom is Director of Digital Strategy at Hills Road Sixth Form College, Cambridge, a post-16 college with 3000 A level students. This presentation draws on his practitioner work leading digital strategy, GenAI development and curriculum-facing implementation in a live sixth form college setting.

As generative AI becomes normalised in education, institutions face a key question: how can they protect learners’ right to think, learn and judge for themselves when machines can support, simulate or replace intellectual work?

The presentation introduces ‘cognitive sovereignty’ as a practical and ethical frame for GenAI implementation. Cognitive sovereignty refers to students’ right and capacity to remain the primary authors of their thinking, judgement, intention, learning and academic work.

The argument is developed through three practical evidence strands. The first is institutional strategy work with leading MAPLE Group sixth form colleges, exploring how different organisational positions on GenAI affect safeguarding, assessment, infrastructure, staff confidence, student access and educational purpose. The second is student-facing curriculum-specific AI development at Hills Road, creating pedagogical agents connected to A level resources and designed to support reflection, explanation, disciplinary language and progression. The third is sector-facing teacher professional learning, focused on helping teachers make subject-specific judgements about when AI scaffolds learning and when it substitutes for the thinking students need to do themselves.

These strands suggest that the current post-16 GenAI landscape is fragmented, with colleges, teachers and students often following different paths in uncertain conditions. The presentation argues that this fragmentation can become more coherent and productive through collaboration, shared inquiry and honest exchange about what works, and how the risk of ‘cognitive surrender’ might best be tackled. By learning together, we can help our students and teachers use GenAI tools with greater confidence, integrity and educational purpose.

AI offers significant opportunities to personalise learning, yet many current AI systems used in education rely on collection and processing of learner data in cloud environments. This creates challenges relating to privacy, trust, transparency, institutional governance and learner autonomy. This paper presents the design principles underpinning an AI-powered study assistant developed to support higher education learners, in particular disabled and neurodivergent students, while maintaining a zero data retention approach. The work explores how meaningful personalisation can be achieved through adapting AI learning and study support to disabled and neurodivergent students’ psychological learning profiles and the learning context, all the while preserving privacy.

Using a human-centred design methodology, the study assistant was developed through iterative collaboration with students, disability specialists, learning technologists and academic staff. Key features include lecture capture support, transcript interaction, adaptive study assistance and personalised learning tools. Rather than creating permanent cloud-based learner profiles, the approach investigates on-device and session-based personalisation, enabling AI systems to adapt to individual learner needs without the need for data retention. The Learning and Study Skills Inventory (Weinstein, Palmer & Acee, 2016) and pedagogical theory were leveraged to train the study assistant and design decisions were evaluated against principles of transparency, explainability, accessibility, privacy and learner agency.

The work demonstrates that meaningful personalisation using AI does not necessarily require large-scale collection of student data. Early findings suggest that privacy-preserving approaches can maintain learner trust while delivering adaptive support that enhances engagement, comprehension and independent study. Next steps will be to hyper-personalise the study assistant to mitigate for barriers faced by individual learners and to develop multimodal outputs beyond the text and speech interactions currently available. These developments will offer institutions a practical pathway for deploying AI study assistants that are educationally valuable, ethically grounded and aligned with emerging expectations for trustworthy AI in education.

We know students are already using generative AI to explain ideas, summarise material and prepare for assessments. The question is whether this use strengthens or bypasses learning, and more importantly how can we support learners to develop the skills they need for a changing world of work?

At City St George’s, University of London, our response is Duologue. Developed with academic and external partners, it is an evidence-led, institutionally-governed learning initiative that asks whether AI can help learners build foundational capabilities while protecting human judgement and agency? Duologue provides structured, personalised AI-supported learning and contextual literacy pathways in critical thinking, AI literacy and ethics, and computational thinking. The platform offers guided, adaptive learning that goes beyond open-ended chatbot use and supports more consistent, purposeful capability development.

Our work reports an ongoing design-based, mixed methods pilot across student, staff and professional learner contexts. The evaluation combines platform analytics, baseline and end of pilot surveys where available, comparative analysis where feasible, qualitative feedback and tutor reflection. We examine engagement, progression, confidence, trust, accessibility, responsible AI use and operational feasibility.

Our contribution is a practical model for responsible AI implementation in higher education that sees AI capability as a matter of curriculum, governance and learner agency, rather than tool access. Our central claim is that the educational value of AI depends on the learning conditions built around it; that by using AI to support a carefully designed and pedagogically-sound learning experience we can help learners to think better.

The rapid integration of Artificial Intelligence (AI) into higher education presents an unprecedented paradox. Tools designed to augment human cognition risk, if left ungoverned, eroding the very intellectual capacities higher education institutions are mandated to cultivate. This abstract introduces the concept of Cognitive Stewardship. A principled framework through which higher education institutions assume the responsibly for preserving, developing, and protecting the cognitive agency of learner in AI-saturated environments.

As large language models, generative AI platforms, and automated tutoring systems become embedded in academic workflows, traditional assessment paradigms face a fundamental legitimacy crisis. Conventional examinations, written assignments, and standardised evaluations, designed to measure individual intellectual effort, are increasingly vulnerable to AI-mediated substitution, raising questions about what higher education is measuring, and for whom. This paper argues assessment is not merely a technical problem of detection or policy enforcement, but a deeper epistemic challenge. Institutions must reconceptualise what they value, how they measure it, and why authentic human cognition remains the irreducible core of educational achievement.

Drawing on emerging scholarship in AI ethics, educational psychology, and learning sciences, this paper propose a Cognitive Stewardship model built on three pillars: inattentional cognitive engagement, AI-transparent pedagogy, and process-centred assessment. The model reorients assessment away from product-based outputs that are easily delegated to AI, and toward observable cognitive processes, metacognitive development, and demonstrated reasoning under conditions of transparency.

The implications are institution wide. Faculty, instructional and curriculum designers, and academic policymakers must collaborate to redesign assessment architectures that are simultaneously AI-resilient and pedagogically generative.

Cognitive Stewardship, this paper contends, is not a defensive posture against technological disruption, it is an affirmative commitment to the enduring purpose of higher education, the formation of independent critically capable human minds.

Keywords: Cognitive stewardship, artificial intelligence in higher education, assessment paradigms, cognitive agency, generative AI, process-centred assessment, AI-transparent pedagogy, and metacognitive development.

Objectives and context

Cognitive offloading has become the dominant lens through which higher education interprets generative AI's impact on student learning, yet this framing is limited: it focuses attention on what AI does to cognition - outsourcing thinking, rather than on how AI can be deliberately designed into learning to transform approaches to critical thinking. This paper challenges the framing of 'offloading' and asks how educators can intentionally design applied AI learning experiences that cultivate independent judgement and epistemic agency, reframing the central question from preventing reliance on AI to scaffolding learners towards critical, professional, and ethically responsible use.

Methods / Approach

Drawing on Vygotsky's Zone of Proximal Development, Bruner's theory of scaffolding, Biggs' constructive alignment, and the ABC Learning Design Framework, the paper develops cognitive scaffolding as an alternative conceptual and pedagogical model to offloading. It further draws on Flavell's work on metacognition to argue that critical thinking in AI-mediated environments is increasingly metacognitive, requiring learners to evaluate machine-generated output, recognise bias and hallucination, and regulate their interaction with intelligent systems as well as their own thinking. The approach is illustrated through discipline-level assessment examples showing scaffolding applied in practice, and also linked to a wider international policy context, including UNESCO's human-centred guidance on generative AI and the OECD's work on future competencies.

Outcomes & Conclusions

The paper concludes that the principal educational challenge is not protecting learners from cognitive offloading but designing learning environments that progressively cultivate epistemic agency and employability skills through carefully scaffolded human-AI collaboration. This repositions AI not as a substitute for thinking but as a mediated learning resource whose role changes as expertise develops, with clear implications for curriculum and assessment design that supports cognition, agency, and critical thinking in AI-enabled education. AI provides an opportunity to rethink how we teach critical thinking - we should embrace it.

Objectives and Context

Current AI tools are increasingly effective at generating complex code. If misused, they risk becoming harmful shortcuts for inexperienced programmers, particularly students with limited technical backgrounds or enrolled in courses where programming is a secondary skill. This work reframes programming education around student agency over AI agents, shifting the focus from writing boilerplate code to problem formulation, solution orchestration, and critical evaluation, while still developing foundational coding skills.

Methods

We designed and evaluated teaching materials based on core data science tasks, including database programming and machine learning. The approach centres on an interactive workflow in which students specify problem requirements and unit tests, then orchestrate AI agents to generate code. Students iteratively refine specifications, guide generation, and critically assess outputs. We collected both process- and product-level metrics aligned with learning outcomes, capturing students’ ability to specify, evaluate, and improve solutions.

Outcomes

We developed replicable modules for quantitative courses where programming and AI agency are explicit learning objectives. Students engage in structured iteration by revising specifications, regenerating solutions, and validating outputs. Preliminary evidence indicates improved engagement with problem structure and increased critical evaluation of generated code. Instructors also develop practical fluency in agent-based programming workflows.

Conclusions

By teaching students to design, guide, and evaluate AI agents, we mirror how mature technology organisations responsibly deploy AI in production. This approach equips students with transferable skills needed to operate effectively in AI-assisted programming environments and contributes to the development of competent, critically engaged programmers.

As AI systems are increasingly initiating, participating in and completing tasks, making decisions and engaging in dialogue with humans, questions about agency, dependency, and what counts as human understanding and expertise are ever more important. These hybrid arrangements can open new possibilities for shared sense making, while they also surface tensions around opacity, bias, and the redistribution of cognitive labour and competences. It is within this evolving space of human-AI collaboration that we have designed a large-scale learning innovation, undertaken by 1,300 university students from across the world and disciplinary backgrounds over the last 3 years while studying at UCL. The learning initiative engages students directly with digital and AI technologies and seeks to create a practice-based alertness to issues of accuracy, bias, ethics, and responsible application. Through an intensive and applied team ‘sprint’ project, students collaborate to undertake a research-underpinned project for a chosen organisation. Using a range of digital and AI tools, teams design an interactive website to present their project. Supported by in-class sessions on digital and AI strategy, ethics and ‘dark sides’, digital consumer behaviour, and AI-led innovation, students are required to collaborate with AI tools, whilst simultaneously reflecting on that interaction. They learn to question outputs, better recognise limitations, identify bias and reflect on the ethical implications of the technologies they use. We used a multi-method approach to study how students collaborate with AI, including individual written reflections, weekly questionnaires and observations of teamwork. We introduce an AI Accelerator & Detractor Heat Map that specifically illustrates the intensity of human-AI collaboration, human pushback, and efforts to reclaim agency. We believe this work will be of interest and relevance to education scholars as well as education practitioners.

Objectives

As higher education seeks to integrate AI in ways that are human-centred, trustworthy, and meaningful, there is growing recognition that learners should shape such initiatives. This presentation describes co-creation of an AI Ethics, Inclusion and Society MOOC at the University of Glasgow. The project responds to the need for responsible AI implementation by embedding student perspectives and supporting critical engagement with issues including bias, power, environmental impact, and social responsibility. The work explores how participatory approaches can contribute to institutional AI literacy strategies that are inclusive and responsive to diverse experiences.

Approach

The course has received internal and external funding; from the Learning Innovation Support Unit, Learning & Teaching Development Fund and the EPSRC “AI Underpinning” Fund. The resulting asynchronous MOOC is hosted on FutureLearn, free at point of access, and available beyond academia. The MOOC is also integrated into credit-bearing programmes within Glasgow Chemistry and Engineering.

Academic staff and student partners collaborated throughout curriculum design, content creation, evaluation, and dissemination. One student co-creator, who is now pursuing a master’s in technology law reflected: “I can't express how much this internship has meant to me”.

Outcomes & Conclusions

Findings demonstrate that student partnership can support the responsible implementation of AI education by producing learning experiences that are contextually relevant, critically informed, and trusted by learners. Since its launch in May 2025, the MOOC has engaged more than 1,600 learners internationally across higher education, industry, and the public sector. Evaluation data show that 92% of learners gained new knowledge and skills, while 96% reported that the course met or exceeded expectations. One learner described the experience as “Incredibly thought provoking; I’m asking questions surrounding AI I have never considered before.” The project provides a transferable model for institutions seeking to embed responsible AI literacy through participatory, inclusive, and sustainable approaches.

Objectives and context

As AI becomes increasingly embedded in education, teachers are making daily decisions about when, how, and whether AI should be used in their professional practice. While much of the current discourse focuses on AI adoption and implementation, less attention has been paid to how teachers make sense of these decisions. Understanding these judgements matters not only for supporting responsible AI use, but also because they offer insight into the professional values, purposes and practices that teachers seek to preserve when AI becomes part of teaching. This study asks: How do teachers make sense of decisions about AI use and non-use within their professional practice, and what do they identify as shaping those judgements?

Methods

This interpretive qualitative study used reflective CPD workshops to create space for teachers to articulate and discuss AI-related decisions with their peers. Data was collected through audio recordings, written reflections from workshop activities, and researcher field notes. Eight workshops involving 88 teachers were carried out. Analysis followed a two-pass Reflexive Thematic Analysis: an initial inductive analysis followed by a theoretically informed interpretation using Priestley et al.'s (2015) ecological model of teacher agency.

Outcomes and conclusions

Preliminary analysis suggests that teachers' decisions about AI are shaped not only by technical considerations but also by professional values, knowledge of individual learners, contextual conditions, and beliefs about the role of teaching. Rather than representing simple decisions about technology adoption, AI use and non-use emerge as acts of professional judgement through which teachers negotiate their professional responsibilities and educational priorities. These findings position teachers as active decision-makers rather than passive adopters of technology and suggest that teachers' professional judgements provide a valuable lens for understanding the educational purposes and professional values that AI should support, rather than replace.

The increasing use of generative AI tools in higher education has changed the student experience in different ways, from providing support with learning difficult topics, to acting as a colleague to discuss topics with.

As educators, we began a research project aimed at understanding how doctoral students use generative AI when generative AI tools first came into common use in the winter of 2022. We found doctoral students assigned four different roles to generative AI tools: a subject-expert tutor, a colleague, a family member, and a personal tutor for wellbeing and lifestyle advice. This outcome provided us with evidence we used to design workshops to prompt critical thinking when using generative AI, taking into account the four roles doctoral students identified.

In the last two years (2024-2026), we have delivered these workshops to enhance the critical thinking of doctoral students when using generative AI tools. Our intention is to give skills which will support them in the doctoral projects, whatever role they give to generative AI, and to prepare them for their future career in academia and beyond.

From the first iteration of our workshops, we have observed the evolution of doctoral students’ perceptions, uses, and understanding of generative AI tools. In our presentation, we share our exploration of how generative AI is changing the research process with the aim of inspiring other educators from the perspective of doctoral students. We discuss how students approach the four roles, including advantages and disadvantages; how our workshops mitigate some of these disadvantages; and how doctoral students move from “passive users” to “critical users” of generative AI.

The integration of Generative Artificial Intelligence in education transforms how learners relate to knowledge, construct beliefs, and exercise agency within knowledge processes. This paper examines these epistemic transformations through the lens of Epistemic Agency (EA). Epistemic Agency is the capacity and responsibility of learners to actively participate in the creation, evaluation, and advancement of knowledge. Current educational responses, particularly Artificial Intelligence (AI) literacy frameworks, emphasise skills and competencies to meet this transformation. We demonstrate that these responses remain limited to the user paradigm, decontextualising both the technologies and the learners who engage with them, and fall short of addressing the complex epistemic relations that emerge when AI systems function as an epistemic infrastructure. We propose cultivating an agentic stance as a pedagogical orientation that goes beyond traditional AI literacy and critical-thinking approaches. Drawing on critical pedagogy, eco-pedagogy, and media studies, the agentic stance is investigative, contextualised, and relational, shaping learners' relationships with knowledge structures. Rather than focusing on what learners can do with AI, the agentic stance emphasises how learners position themselves within AI-integrated epistemic ecologies. This approach requires learners to view AI systems as objects of investigation, make the embedded power relations visible, and recognise their participation in complex human-machine interdependencies.

The dominant question in many evaluations of AI in education is simple: did performance improve? Yet a score increase measured at a single moment, in a single context, with AI support present cannot answer a more consequential question: is the learner more capable when the tool is removed than they were before?

This paper argues that AI-in-education research faces a “snapshot problem”: a tendency to evaluate short-term task performance rather than durable changes in human capability. The problem is especially acute in studies of generative AI, where the boundary between learner performance and tool performance is often unclear. If students produce better answers with AI assistance, this may indicate improved learning, improved scaffolding, increased dependency, or some combination of all three.

The paper identifies three measurement gaps in current AI-mediated learning research. First, evaluations often privilege task completion over the cognitive and metacognitive processes that make learning transferable. Second, benchmarks and outcome measures are frequently repurposed across human, AI, and human-AI contexts without sufficient validity checks. Third, relatively few studies follow the same learners across time, contexts, and changing levels of tool availability, limiting what can be inferred about whether AI builds, masks, or erodes human capability.

Drawing on recent large-scale AI-in-education trials, benchmark validity debates, and the author’s work on metacognitive discernment, this paper proposes five requirements for evaluating AI’s impact on human learning: transfer tests, metacognitive calibration measures, dependency probes, information-literacy tracking, and discernment rate, a proposed metric for assessing how often learners question, correct, reject, or override AI-generated outputs over time.

The paper concludes by outlining an improved longitudinal study design for AI-mediated learning that treats learner agency as a primary outcome. Rather than asking only whether AI helps students perform better in the moment, education research must ask whether learners become more capable, discerning, and intellectually independent across time.

Generative AI has not created a crisis in academic writing. It has exposed one. Higher education has long treated polished prose as reliable evidence of a student’s thinking and understanding. Now that AI can generate, transform, and refine academic text, that assumption demands scrutiny. The question is not how to regulate AI, but how to go back to fundamentals and reconceptualise what authorship means.

This presentation introduces Digitally Enabled Translanguaging for Academic Purposes (DETAP), a framework grounded in translanguaging theory, writing-process research, academic literacies, and post-plagiarism scholarship. DETAP argues that academic writing has always been a mediated process of meaning-making, and that contemporary writing is now largely digitally mediated. Generative AI has simply made this reality impossible to ignore. It reframes writing as the accountable governance of linguistic, cognitive, and technological resources, not the isolated production of standard academic English.

The presentation challenges three assumptions that underpin conventional approaches to assessment and integrity: that academic English is a neutral medium rather than a socially authorised prestige register; that polished prose evidences thinking; and that authorship requires unaided textual production. In their place, DETAP distinguishes ideation from expression and redefines authorship as the combination of intellectual ownership and evaluative control. Writers remain authors when they retain intellectual ownership of the ideas and exercise evaluative control over how those ideas are expressed, even where AI assists with linguistic expression.

Rather than lowering standards, this reconceptualisation relocates them. The educational challenge becomes developing learners’ capacity to supervise, evaluate, and take responsibility for AI-mediated writing, while enabling more equitable participation for multilingual, neurodivergent, and other writers for whom conventional academic English has posed an unnecessary barrier. DETAP therefore offers a principled account of human authorship that distinguishes thinking from expression and provides a foundation for rethinking teaching, assessment, academic integrity, and what counts as evidence of human intelligence in the age of AI.

Generative AI is increasingly available to secondary school pupils, including in mathematics, but its educational significance extends beyond whether it produces correct answers. This paper asks how the teacher’s role changes when pupils work with AI systems that can explain, prompt, assess and persuade. It argues that the central challenge is no longer evaluating AI outputs, but protecting pupils’ epistemic agency: their capacity to question, justify and retain ownership of mathematical reasoning. Central to this is the quality of the student–AI interaction and the learner stance adopted towards AI as a source of mathematical authority.

Drawing on practitioner research in a Year 11 mathematics classroom, including think-aloud protocols, student interviews and analysis of student–AI interactions, the paper examines how pupils engage with generative AI during mathematics revision. It also draws on subsequent BERA-funded work with pupils with SEND, highlighting how interactional viability, the extent to which an AI interaction remains accessible and productive, shapes whether AI supports or undermines independent mathematical thinking.

The findings suggest that the same AI system creates different pedagogical problems depending on learners’ stance towards algorithmic authority. While some pupils use AI transactionally as an answer machine, others engage critically and dialogically, with important consequences for reasoning, judgement and autonomy. The research further suggests that correctness is an insufficient measure of educational value: mathematically accurate explanations may still diminish productive struggle, overwhelm learners with symbolic or linguistic complexity, or prematurely close down sense-making.

The paper argues that AI does not diminish the need for teachers but redefines their expertise. In AI-mediated mathematics classrooms, teachers must learn to interpret not only pupils’ misconceptions but also the student–AI interactions that shape them. Their distinctive role becomes the protection and cultivation of epistemic agency, ensuring that AI supports rather than displaces independent mathematical thinking.

Objectives and context

Generative artificial intelligence has introduced a significant challenge for higher education assessment by weakening the relationship between performance and cognition. Traditionally, Assessment relies on the assumption that student work provides interpretable evidence of understanding, reasoning, and disciplinary knowledge. While much of the recent discussion has focused on academic integrity, assessment security, and AI-resistant task design, this paper argues that the more fundamental issue concerns construct validity. AI systems can generate academically credible outputs across a wide range of assessment tasks, and there is an inferential relationship between performance and a progressive decline in the stability of the constructs being assessed.

Methods / Approach

This paper adopts a design-based practitioner inquiry approach, drawing on four years of iterative assessment redesign within fully online postgraduate programmes in health economics, sustainability, and public policy at the James Lind Institute in Geneva, delivered to learners across more than 25 countries. It develops the concept of epistemic traceability to examine the extent to which reasoning processes remain observable and attributable within assessment. The paper analyses assessment approaches including scenario-based reasoning tasks, structured oral defences, critical evaluation of AI-generated outputs, and reflective disclosure of AI use.

Outcomes and conclusions

This paper argues that generative AI reveals a previously underexamined assumption within assessment practice: that performance can be treated as evidence of individual cognition. It proposes that assessment should increasingly focus on the traceability and defensibility of reasoning rather than on completed artefacts alone. This paper concludes that the central challenge for higher education is not simply adapting assessment to AI, but reconsidering what constitutes valid evidence of learning when cognitive production can be externally augmented.

Most research on GenAI in education measures individual performance: whether it raises test scores, accelerates feedback, or personalises instruction. Far less attention has been paid to the human connections through which learning actually happens, i.e. to what becomes of the learner–teacher relationship, of peer collaboration, and of belonging when GenAI enters shared learning communities. This review addresses that gap. It proceeds from the premise that learning is relational: classroom communities depend on the relationship between learner and teacher as a source of authority, trust, and intellectual direction, and on relationships among peers as a source of mutual support and shared meaning. The objective is to examine what is known about GenAI presence within these relationships and how it reshapes the connections on which engagement and learning depend.

The authors performed a critical review of peer-reviewed literature on GenAI in shared learning contexts, published between 2022 and 2026. Sources were selected for their bearing on the relational dimensions of learning and were analysed thematically across three dimensions of learning community: conceptual understanding, collaboration, and belonging.

Three patterns emerge. First, when GenAI systems assume explanatory and scaffolding roles traditionally held by teachers, authority shifts away from the teacher, and human-to-human interaction gives way to human-GenAI exchange, reconfiguring the trust and role relations that anchor the learner–teacher relationship. Second, despite advances in voice, image, and immersive technologies, learners rarely perceive GenAI as a genuine participant in their learning community in the way they perceive teachers and peers, exposing an unresolved gap between technological sophistication and authentic membership. Third, human-GenAI interaction remains predominantly learner-initiated, lacking the reciprocal initiative (noticing, reaching out, and redirecting) that characterises authentic teacher–student dialogue.

The review concludes that GenAI reshapes the social conditions under which conceptual understanding, collaboration, and belonging are formed. The authors discuss future research directions for ensuring that GenAI's presence in learning communities supports, rather than substitutes for, the human connection on which learning depends.

The literature on human–AI collaboration has primarily examined how AI systems improve decision making, productivity, and the allocation of work between humans and machines, while responsible AI implementation has focused on guardrails that keep systems aligned with institutional goals. What we know far less about is what happens after deployment, when users learn and test system guardrails, especially in education, where AI tutors are meant to support learning without substituting for student effort. To examine this gap, we analyze two years of de-identified interaction data from an AI tutoring platform designed to scaffold reasoning rather than provide answers. We used natural language processing to classify prompts, code answer disclosure, and trace subsequent behavior using student, question, month, and exact-prompt fixed effects. We find that students do not merely collaborate with AI tutors; they strategically manage them. Direct answer requests receive explicit answers about 30% of the time, while answer-only and jailbreak-style prompts succeed roughly 69% of the time. When the tutor gives the answer, continuation falls by about 33 percentage points after fixed effects, suggesting that the tutor has shifted from collaborator to solver. We also document a “retry lottery”: after an initial refusal, identical repeated prompts later obtain the answer in 66% of cases, and 88% of repeated question-prompt cells show both disclosure and withholding. We contribute by theorizing responsible AI implementation as the governance of strategic task capture: the process through which users learn inconsistent guardrails and shift work from themselves to the system, requiring organizations to preserve not only technical safety, but also the intended human–AI division of labor. This is particularly urgent for higher education institutions adopting AI tutors at scale: responsible implementation must assess not only whether these tools are accurate, safe, or engaging, but whether everyday use preserves the learning work that justifies their adoption.

ContextGenerative AI tools have become many students' entry point to the research literature, and they can be a genuine support. Even tools built for scholarship now provide citations and links, so the issue is what happens next: reading a generated summary often means students and faculty never open them. We call this pattern zero-click research. Norwegian national library statistics show a year-on-year fall in journal article downloads even as student citation counts appear to rise, so sources may be cited more than they are read. Others have named a related effect, "zero-click AI" (Ozturkcan, 2026), from a publisher lens; our question is about the learner and the judgment that reading builds.

ApproachBuilding on the Norwegian data, we are conducting a comparative analysis of publicly available library usage statistics (COUNTER-based and equivalent open datasets) from 2024-2025 in Germany, Austria, France, and Finland, to test whether the decline in article downloads is a national anomaly or part of a broader European trend. A download is an imperfect proxy for reading, which we weigh accordingly.

ConclusionWe expect to find that declining engagement with primary sources is not confined to Norway. We argue this reflects how answer interfaces work rather than a passing quality problem that better models will fix: an interface that leads with a finished answer leaves the reader little reason to open the sources behind it, and so little reason to do the reading that learning depends on. The issue is less the students than the learning design we invite them into, so we discuss how institutions might build AI practices and assessment that protect independent thought.

DisclosureAuthor is affiliated with Keenious, a research-discovery tool. We raise this because we think it matters for learning, not to promote a product; the argument and data stand on their own.