How does generative AI facilitate faculty interaction and role transition in biostatistics?

What the study actually tracked

The paper asks how GenAI integration reshapes the teaching team itself, not only student learning. The setting was a one-semester undergraduate biostatistics course at a single medical school in Taiwan, taught across two academic years (2024 and 2025) to nursing, pharmacy, and medical biotechnology students. Statistical concepts were embedded in clinical scenarios and organized around the PPDAC problem-solving cycle (problem, plan, data, analysis, conclusion/communication).

The evidence came from two streams. Teachers' reflection records captured how the team described its own practice, and 4,697 LINE messages sent during the study period captured who talked to whom. The messages were coded for teaching relevance, educational research relevance, and GenAI relevance by one author and reviewed by two others, with disagreements settled by discussion. Qualitative coding, two LLM-assisted semantic workflows, and social network analysis were then combined.

The design matters for reading every result below. This was a single-course, single-institution action-research study with a small teaching team, and the quantitative network model contained eight active nodes. The authors themselves label the network estimates exploratory rather than confirmatory.

GenAI moved from tutor to co-creator

In the first year, students used GenAI mainly as a personal tutor for technical support, and instructors described their own role as resource providers who designed lesson plans, offered AI kits, and gave positive guidance. The stated aim was lowering technical barriers so students knew how to solve problems.

In the second year, the framing changed. Instructors required comparative analysis and iterative inquiries so students would check whether AI answers were consistent with other sources, and statistics was taught as a complete problem-solving process rather than a set of tool operations. Teachers also used GenAI to synthesize student inquiries and draft responses that human faculty then reviewed and refined.

This is the concrete meaning of the paper's phrase "human–GenAI co-learning": GenAI produced material, and faculty judged and corrected it. The authors describe the shift as double-loop learning, in which the team revisited its underlying assumptions about teaching and assessment rather than only optimizing delivery. That label is the authors' interpretation of the reflection records, not a separately measured outcome.

Two LLM readers converged on the role shift

The semantic analysis used two LLM workflows with deliberately different styles. Gemini Pro reported a qualitative, holistic reading, while Claude CoWork reported a quantitative, keyword-based approach that converted messages into term-frequency–inverse-document-frequency features and applied k-means clustering. Gemini named five clusters, including administrative logistics, GenAI integration, and educational research; Claude named six, including critical thinking, team growth and dynamics, and clinical application.

The two outputs did not match cluster for cluster. The research team reviewed and compared them in consensus meetings and merged them into three themes: AI-mediated curriculum redesign, faculty role transition, and evidence-informed educational research. Both LLMs independently flagged the same two signals: growing emphasis on higher-order thinking and a shift in the teachers' role from knowledge providers to facilitators of critical thinking.

The authors treat the LLM outputs as preliminary analytic aids rather than definitive findings, and the paper notes that the actual data-access method was not defined by the LLM. The convergence is therefore supporting evidence for the role-shift theme, not an independent measurement of it.

Overall talk was dense; GenAI talk was selective

The social network analysis treats each teaching-team member as a node and each observed message exchange as a directed tie, weighted by interaction frequency. Four measures summarize the structure: density is the share of observed ties among all possible ties, reciprocity captures mutual exchange, transitivity reflects triadic closure, and centralization indicates how much communication depends on one or a few actors.

Across all conversation types, the faculty network was highly connected: density 0.875, reciprocity 0.939, transitivity 0.850, and centralization 0.143. In plain terms, most members talked directly with one another, exchanges usually went both ways, and the network did not hinge on a single central figure.

The topic-specific networks looked different. Teaching-related conversation had density 0.595 and centralization 0.472. GenAI-related conversation was more selective still, with density 0.411, reciprocity 0.696, transitivity 0.549, and centralization 0.429. These are different measures on different subnetworks, so the GenAI numbers should not be read as a decline in team cohesion; they describe a narrower, more clustered discussion that formed around particular needs and expertise.

GenAI ties formed rarely but held strongly

The paper's summary finding is that GenAI-related ties were less likely to form but, once formed, were strongly reciprocal and locally clustered. The density and centralization figures support the first half: fewer possible GenAI connections were actually used, and that discussion concentrated around particular members. The reciprocity of 0.696 and transitivity of 0.549 support the second half, showing that the connections which did exist tended to be mutual and to close into small triangles.

The authors used bootstrap temporal exponential random graph models to probe which mechanisms governed tie formation, including reciprocity, triadic closure, lagged network dependence, and individual attributes such as gender and medical professional background. Overall communication showed positive lagged network dependence, with an estimate of 1.868. Because the network was small, the authors present all of these estimates as exploratory.

The practical reading is that GenAI knowledge did not spread evenly through the team. It circulated through a subset of members who had the relevant expertise or task need, which is why the authors argue that cross-disciplinary exchange needs deliberate structure rather than being left to emerge.

Leadership shifted as the curriculum matured

The message logs show that roles were not fixed. The course director was initially prominent in defining the problem, clarifying learning goals, and encouraging a less hierarchical team structure. The statistics educator introduced educational theories and frameworks into the design, then shifted toward mentoring, framing, and supporting evaluation of GenAI integration. The nursing educator moved from curriculum implementation toward leading collaborative educational action research.

Message volumes and topic mixes changed alongside these roles. For example, the course director's GenAI-related share of messages rose from 6.6% in year one to 29.9% in year two, and the statistics educator's rose from 6.1% to 22.1%. Group broadcasts were common throughout, which the authors read as a sign of information transparency.

The authors conclude that GenAI integration can catalyze new forms of faculty leadership through expertise sharing and collaborative problem-solving. That is their interpretation of the logs and reflections; the study did not test whether this leadership pattern improves student learning, and the authors explicitly call for further work on short- and long-term student outcomes and on equity of AI access.

Where this evidence stops

Several boundaries qualify the conclusions. Communication was captured on a single messaging platform, so informal interactions, face-to-face discussions, and individual AI experimentation fall outside the data. Coding of GenAI-related discourse depended on human and LLM interpretation, which may miss nuance, indirect sentiment, or latent tensions. The authors suggest fine-tuned models for formal classification and sentence-level transformer embeddings for affective and pragmatic dimensions as possible improvements.

The most decision-relevant point for curriculum leaders is a comparison rather than a single number. A team can show high overall density and reciprocity while its GenAI discussion remains concentrated in a few reciprocal, clustered ties. The paper's own framing puts the challenge there: introducing GenAI into learning tasks is not the same as building faculty networks that convert AI-related experimentation into shared pedagogical capacity.

That distinction is where this study's evidence is strongest and most specific. The measured contrast between overall communication (density 0.875, centralization 0.143) and GenAI-related communication (density 0.411, centralization 0.429) suggests that leaders should look at topic-level network structure, not just overall team cohesion, when judging whether GenAI expertise is actually circulating.

Sources used in this answer