Humboldt-Universität zu Berlin - Mathematisch-Naturwissenschaftliche Fakultät - Institut für Informatik

Human–AI Analytical Configurations for Learning Analytics in Complex Face-to-Face CSCL Environments

 

Foreseen reviewers:

 

Prof. Dr. Gergana Vladova

E-mail: gergana.vladova@hu-berlin.de

        Department of Computer Science, Humboldt-Universität zu Berlin

 

Prof. Dr. Niels Pinkwart

E-mail: niels.pinkwart@hu-berlin.de

        Department of Computer Science, Humboldt-Universität zu Berlin

 

Prof. Lei Wang

E-mail: wanglei@tongji.edu.cn

        Department of Control Science and Engineering, Tongji University

 

Abstract:

 

Artificial intelligence (AI) is increasingly used in learning analytics (LA) to process complex educational data at greater scale and with higher levels of automation. Yet learning processes such as collaboration, regulation, and problem solving are not directly observable; they become analytically meaningful through representation, theoretical operationalisation, contextual interpretation, and validation. This raises a central challenge for AI-assisted LA: how can computational participation be organised without weakening meaningful interpretation, theoretical grounding, and human responsibility?

 

This thesis investigates how human–AI analytical configurations can support meaningful, interpretable, and theoretically grounded LA in complex face-to-face computer-supported collaborative learning (CSCL) environments. It conceptualises configurations as alternative organisations of analytical work across human and computational actors, characterised by the relationship between automation and human interpretive control. Four configurations are examined: a human-centred analytical baseline, delegated computational analytics, multi-stage analytical integration, and prompt-based large language model analytics.

 

The findings show that increasing automation redistributes rather than eliminates analytical and interpretive work, shifting human involvement toward construct operationalisation, validation, contextualisation, and review. Human-centredness therefore depends not simply on human presence, but on preserving meaningful interpretive control and contextual judgement. The findings further show that learning-theoretical constructs are configuration-dependently enacted and may be transformed through computational representation and generation, creating construct drift as a potential epistemic risk. These findings are synthesised into a system-level design perspective for AI-assisted LA. Rather than identifying a universally optimal configuration, the thesis argues that meaningful, interpretable, and theoretically grounded LA depends on coordinating automation, representation, interpretation, validation, and human interpretive control within particular analytical tasks and contexts.

 

***

 

Kind regards,

Gergana Vladova