From co-construction to co-evolution between human actors and AI in learning - [Thesis]

This work is part of the field of Computer Environments for Human Learning (EIAH) and analyzes the role that artificial intelligence can play in the personalization of learning situations.

The general aim of the manuscript is to understand how to design systems capable of adapting pedagogical environments to learners' characteristics, while taking into account teachers' constraints and practices.

The author explicitly situates her research at the interface between computer science, knowledge engineering and educational science. She reminds us that IALT is an intrinsically interdisciplinary field that aims to "provide models and tools to support learning and teaching in digital environments" (p. 2).

Replacement or support

The manuscript defends a central thesis: artificial intelligence applied to education should not be conceived as an automation mechanism replacing human pedagogical activity, but as a set of devices likely to support the activity of teachers and enrich learning environments. In this respect, the author emphasizes that her work aims to "propose not automatic systems that replace teachers, but tools that can be adapted to their practices", making it possible to generate or recommend pedagogical activities adapted to learner profiles (p. 6).

The manuscript begins with a historical perspective on the field of AIE and the relationship between artificial intelligence and learning. The evolution of the field is presented as a succession of attempts to solve a fundamental pedagogical problem: adapting teaching to the individual needs of learners.

• The first programmed teaching systems were based on branching pathways that allowed the learner's progress to be modulated according to his or her answers. The advent of computer science and then artificial intelligence gradually led to the development of more complex systems capable of integrating knowledge about the subject being taught, the pedagogy and the learner.
  
  Intelligent Tutoring Systems occupy a central place in this genealogy. These systems are based on the idea that learning can be supported by explicit models of knowledge, making it possible to diagnose a learner's cognitive state and suggest appropriate activities.
  
  
• However, the author also points to the emergence of alternative approaches favoring open learning environments in which the learner retains greater autonomy. Educational recommendation systems are an illustration of this evolution: they "recommend resources, leaving the learner the choice of accepting or rejecting the recommendation" (p. 4).

Based on this contextualization, the scientific contributions presented in the habilitation are organized around two main lines of research.

• The first concerns the modeling of pedagogical knowledge to assist teachers in the design of personalized learning environments.
  
  The work mobilizes knowledge engineering methods to formalize pedagogical expertise and make it usable by computer systems. This approach has led to the design of meta-models for representing pedagogical activities, learner profiles and skills objectives.
  
  These models are integrated into authoring tools enabling teachers to parameterize adaptive learning environments. The approach developed aims to support rather than automate pedagogical activity. In particular, it can generate or recommend pedagogical activities based on learner characteristics and the constraints of the teaching context.
  
  Particular emphasis is placed on integrating the skills-based approach into these adaptive environments. The models we have developed make it possible to use skills repositories to calculate learning profiles and guide pedagogical recommendations. This orientation reflects the desire to bring computer models closer to the pedagogical frameworks used in educational institutions.
  
  
• The second line of research concerns the exploitation of learning traces produced by digital environments.
  
  Learning platforms today generate a growing volume of data describing learners' interactions with educational resources. This data is an important source of information for understanding learning processes and supporting pedagogical decisions. The work presented here combines knowledge engineering, data mining and machine learning to analyze these traces and extract useful models for instructional design.
  
  This research has led to the development of platforms for analyzing learning data without requiring advanced programming skills. They also facilitate the capitalization and sharing of analyses produced by users.

The author emphasizes that these systems combine two complementary approaches: models derived from pedagogical expertise (top-down approach) and models discovered from data (bottom-up approach).

On what basis are recommendations made?

Another important contribution concerns the transparency of artificial intelligence systems used in educational environments. The decisions taken by these systems, such as pedagogical recommendations, are often based on algorithmic mechanisms that are difficult for users to interpret. The work presented here explores the possibility of introducing explicable AI mechanisms to make the decisions produced by artificial intelligence engines comprehensible.

Prospects for Human-IA interactions

The manuscript concludes with the presentation of a research program structured around several scientific perspectives. These include

• the development of explicable AI mechanisms integrated into the personalization cycle of learning environments,
• the design of hybrid AI architectures combining symbolic approaches and machine learning, as well as
• the study of co-evolution between human actors and intelligent systems in educational environments.

The latter perspective is one of the main threads running through our work. It is based on the idea that digital learning environments should not be conceived as mere technological tools, but as socio-technical devices in which teachers, learners and computer systems interact.

From this perspective, the personalization of learning appears less as an algorithmic problem than as a question of articulation between computer models, pedagogical knowledge and educational practices. The habilitation thus proposes a conceptual and methodological framework for thinking about the integration of artificial intelligence in learning environments, while taking into account the pedagogical and institutional realities of the educational world.

Reference

Lefevre, M. (2026). From co-construction to co-evolution between human actors and AI within the EIAH personalization cycle. Habilitation to direct research in computer science, Université Claude Bernard Lyon 1.
https://hal.science/tel-05504496v1