Source-linked AI summary
The Interplay of Learning, Analytics, and Artificial Intelligence in Education: A Vision for Hybrid Intelligence
Mutlu Cukurova
TL;DR
The paper addresses the narrow treatment of AI in education as applied tools and the resulting separation of AI research from broader accounts of human intelligence and learning. It develops three conceptualisations—externalising, internalising, and extending human cognition—and examines educational examples and their limitations. It concludes that hybrid intelligence and AI’s role in understanding learning require broader empirical, educational, and systemic attention.
Problem
AI in education is often conceptualised as applied tools for automating decisions, while human intelligence and AI’s role as a way to understand learning receive narrower treatment.
Method
The paper compares three AI conceptualisations in education: externalising human cognition, internalising AI models, and extending cognition through tightly coupled human-AI systems.
Results
The paper presents hybrid intelligence as combining complementary human and AI abilities, while educational examples show both useful analytics-supported interventions and substantial challenges for agency and social learning.
Takeaways & Limitations
AIED should examine human-AI interaction, use AI to understand learning, and consider how education can remain relevant in an AI-ubiquitous world.
Takeaways & Limitations
AI interventions in complex social learning contexts face challenges involving threatened human agency, prediction, normativity, and alignment between human values and AI goals.
Abstract
from arXiv · showhide
This paper presents a multi-dimensional view of AI's role in learning and education, emphasizing the intricate interplay between AI, analytics, and the learning processes. Here, I challenge the prevalent narrow conceptualisation of AI as tools, as exemplified in generative AI tools, and argue for the importance of alternative conceptualisations of AI for achieving human-AI hybrid intelligence. I highlight the differences between human intelligence and artificial information processing, the importance of hybrid human-AI systems to extend human cognition, and posit that AI can also serve as an instrument for understanding human learning. Early learning sciences and AI in Education research (AIED), which saw AI as an analogy for human intelligence, have diverged from this perspective, prompting a need to rekindle this connection. The paper presents three unique conceptualisations of AI: the externalization of human cognition, the internalization of AI models to influence human mental models, and the extension of human cognition via tightly coupled human-AI hybrid intelligence systems. Examples from current research and practice are examined as instances of the three conceptualisations in education, highlighting the potential value and limitations of each conceptualisation for education, as well as the perils of overemphasis on externalising human cognition. The paper concludes with advocacy for a broader approach to AIED that goes beyond considerations on the design and development of AI, but also includes educating people about AI and innovating educational systems to remain relevant in an AI-ubiquitous world.
1. Human Intelligence and Artificial Information Processing
AI is often treated as information-processing tools that replace or support decisions, but human intelligence also involves contextual, embodied, social, and emotional abilities. This distinction motivates tightly coupled human-AI systems that combine complementary strengths while keeping humans engaged.
- Human intelligence includes learning, reasoning, decision-making, adaptation, and emotional and social interaction with the world.
- AI in education commonly uses big-data analysis and prediction to replace decision-making through a user interface.
- Transformer-based language models process inputs through neural-network layers, predict outputs, calculate loss, and adjust weights through backpropagation.
- Generative AI tools are black boxes with limited real-world contextual understanding, possible incorrect information, and unclear educational success measures.
- Tightly coupled human-AI systems combine human flexibility, reflection, planning, and contextual understanding with AI data processing and analytics.
2. AIED and The Direction of Research towards AI as an Applied Tool
AIED has developed systems that automate pedagogical tasks, while the field’s earlier connection between AI and learning as an analogy for human intelligence has weakened. Evidence indicates tutoring systems can be effective, but their broader direction and adoption remain open questions.
- Early learning sciences and AIED treated AI as an analogy for studying human intelligence, but this perspective became less prominent.
- AI in Education can externalize, internalize, or extend human cognition through tools, human mental models, and tightly coupled hybrid systems.
- Intelligent tutoring systems adapt content, pacing, and feedback to students’ mastery levels and needs.
- 83 adult learners showed no significant learning-gain or learner-experience difference between recorded human lectures and AI-generated synthetic media.
- 0.66 standard deviations was the median test-score improvement from intelligent tutoring systems over conventional levels, rising from the 50th to the 75th percentile.
3. Research evidence on effectiveness versus real-world impact of AI in Education
AIED effectiveness does not by itself determine real-world educational impact. Adoption depends on institutional, pedagogical, infrastructural, social, and psychological conditions, while fully automated systems may narrow learning and overlook socio-cultural contexts.
- AI-tool effectiveness is only one influence on adoption; policy, governance, culture, infrastructure, teacher support, knowledge, confidence, ownership, workload, and ethics also matter.
- Fully automated systems may dehumanize learning by prioritizing information and declarative knowledge over tacit knowledge, wisdom, social competence, and emotional intelligence.
- About 5% of students independently engage with educational resources long enough to obtain statistically significant benefits.
- Teachers and learners may hold confirmation biases and unrealistic expectations, while AI-framed content can be judged less credible than equivalent content framed through educational psychology or neuroscience.
- Task-automation tools commonly rely on computational cognitive models that overlook socio-cultural learning beyond the individual mind.
4. Alternative Conceptualisations of Artificial Intelligence
Alternative AI conceptualisations use computational models to reshape human representations of thought and support learning in complex, constructivist environments. Analytics can inform teacher intervention, but prediction and intervention raise agency and normativity challenges.
- Internalized AI models can change humans’ representations of thought while preserving high human agency and control under relatively low automation.
- Constructivist learning research can use computational models to help learners develop competence through rich experiences and reflection.
- Researchers collect multimodal data including video, gestures, physical interactions, reflections, notes, and self-declared emotions in open-ended design problems.
- Multimodal features often produce better competence-prediction results than unimodal predictions.
- A multimodal dashboard was associated with teachers spending less time monitoring and more time scaffolding, alongside less student boredom.
5. The Challenge of Using AI as a Tool to Directly Intervene in Teaching and Learning
Directly intervening AI in teaching and learning faces challenges involving human agency, prediction in social contexts, and deciding what is good or bad in complex learning situations. AI models may instead support more precise descriptions of learning, but mechanistic measures can misrepresent its quality.
- Challenges of direct intervention: Direct intervention by AI threatens human agency and struggles with prediction and normativity in complex social learning contexts.Addressing these issues requires stronger alignment between human values and AI goals.
- Models for understanding: AI models can describe learning processes with greater precision without prescribing actions from predictions in complex social learning processes.This conceptualisation treats models as opportunities for thinking about learning.
- Limits of mechanistic measurement: Breaking collaboration into measurable components can prioritize faster or more precise completion even when those measures do not improve learning quality.Such measures may capture machine capabilities better than all aspects of human learning.
6. Value of Making Lived Experiences Visible to End Users
Computational models can make lived learning experiences visible through precise feedback and visualisations for students and teachers. Although users recognize these models as incomplete, the resulting awareness can support reflection, motivation, engagement, and regulation.
- Feedback for end users: Models can provide students and teachers with precise feedback about group interactions, supporting awareness, reflection, and motivation for future activities.Examples include speech-time percentages, interaction types, timelines, visualisations, and feedback.
- Making collaboration visible: Multimodal systems can detect speech, transcribe group discourse, label challenge moments, and return visualised feedback to students.The feedback includes further explanation of detected collaborative challenges.
- Value despite incompleteness: Users may value incomplete models because visible information increases awareness of their own activities and others’ behaviours.This awareness influences motivation, engagement, and regulation, including coregulation and socially shared regulation.
7. Value of Contributing to Learning Sciences Literature
AI models can formalize researchers’ concepts in more detailed and precise language, generating insights into complex learning processes. These insights may contribute to the further improvement of learning theory.
- Formalizing learning concepts: Computational models can clarify and communicate researchers’ concepts in detailed, precise, and formal language.This can generate potential insights into complex learning processes and advance learning theory.
- Contributing to learning theory: Models in multimodal learning analytics research have potential to contribute to learning theory.A literature review identified this potential in the relationship between learning theories and models.
8. Human-AI Hybrid Intelligence Systems
The paper identifies tightly coupled human-AI systems as a high-automation, high-human-agency approach intended to extend human cognition and competence. It also warns that automation and current interaction models can erode human competencies or limit genuine hybrid intelligence.
- Human cognition extended with AI: Tightly coupled human-AI systems aim to combine high automation with high human agency to extend cognition rather than merely improve task performance.The paper distinguishes this goal from current complementarity approaches that primarily match human and AI capabilities for productivity gains.
- Human cognition extended with AI: Delegating tasks to AI requires judgment because over-reliance may cause the atrophy of critical human competencies.The paper presents automation as appealing but cautions that preserving human capabilities must remain part of its evaluation.
- Risks of automation: Human-in-the-loop correction may make people converge toward AI-generated content instead of critically grounding judgments in their own understanding.The paper compares this risk with accepting the first search-engine suggestion as truth and calls for evidence-informed testing of automation over the long term.
- Alternative pathways: Internalized AI models are intended to fade away as human competence develops, whereas hybrid systems should strengthen competence as interaction increases.These are presented as distinct pathways: internalization changes human representations of thought, while extension builds a synergistic superstructure over human intelligence.
- Limits and future needs: Current AI systems require meanings to be clearly transferred before generating relevant outputs, limiting fluid interaction and shared meaning-making in hybrid systems.The paper therefore calls for more research and implementation of human-AI hybrid systems in education.
9. AIED is broader than the Design, Development, and Use of AI
AIED should extend beyond designing, developing, and using AI systems to include AI competency education and innovations that keep education relevant in an AI-influenced world. These changes include scaffolding human-AI interaction and redesigning assessment to evaluate learning processes, not only outcomes.
- AI in education requires educating people to use AI and data safely, effectively, and ethically, alongside designing and developing AI systems.
- 9.1 Educating People about AI: AI competency includes more than technical knowledge and application skills, requiring broader competencies for everyday interaction with AI.The UNESCO framework for teachers treats AI techniques and applications as only one of five main aspects of competence.
- 9.1 Educating People about AI: Tools and models should initiate continued research and practice that develops end-user competencies through appropriately scaffolded teacher and learner interactions.Hybrid intelligence does not emerge automatically merely by placing AI tools in end users’ hands.
- 9.2 Innovating Education Systems for an AI-driven World: Education systems must innovate, particularly in assessment, to remain compatible with tightly coupled human-AI hybrid intelligence systems.The paper calls for education to remain relevant in a world heavily influenced by AI.
- 9.2 Innovating Education Systems for an AI-driven World: Because many educational competencies are process-driven, assessments should evaluate how students develop competence rather than only the final products they produce.Essay writing is presented as teaching regulation, research, accuracy judgment, synthesis, and clear persuasive expression.
- 9.2 Innovating Education Systems for an AI-driven World: Analytics-based behavioural feedback had limited impact on already successful students but significantly increased engagement and performance among struggling students initially predicted to fail.The intervention used writing-engagement data such as time, edits, and added content to provide formative feedback on writing practices.
10. Concluding Remarks
AI in education should be understood through multiple conceptualisations, including its use to model learning and extend human cognition through hybrid intelligence. The paper calls for intentional, evidence-informed, human-centred educational futures that account for the benefits and unintended consequences of these approaches.
- AI can externalise, be internalised, or extend human cognition through tightly coupled human-AI hybrid intelligence systems.
- AI models can describe learning processes with greater detail and precision, supporting feedback, motivation, awareness, and theoretical understanding.The paper cautions that some learning aspects arise through lived experience and may not be fully explained by predictive models.
- Each AI conceptualisation may advance education while also producing unintended consequences that require careful consideration.
- Future hybrid-intelligence education should be intentional, evidence-informed, and human-centred, supported by empirical study of human-AI interaction and human competencies.
- The paper frames AI research as a continuing inquiry into human intelligence, including its augmentation, amplification, and the risks of entrusting cognitive competencies to AI tools.