Be(e)Chat – Safe and Responsible AI for Higher Education

Duration: January 2024 – December 2024
Status: Completed
Educational Level: Tertiary Level
Topic: Artificial Intelligence AI, Data Science for Education
Keywords: Data Science for Education, Artificial Intelligence AI

Initial Situation

The increasing integration of generative AI (GenAI) tools such as ChatGPT into higher education presents both opportunities and challenges. While such tools can support learning, writing, and teaching, their use in academic settings raises significant concerns around data privacy, transparency, bias, and institutional control. Commercial systems process user data externally, often lack transparency in data handling, and embed opaque biases in their training data. Moreover, subscription-based pricing models risk exacerbating inequities among students. In this context, educational institutions face a dilemma: how can they leverage GenAI’s educational potential while ensuring compliance with privacy laws, maintaining academic integrity, and upholding institutional autonomy? The Be(e)Chat project addresses this challenge by developing and evaluating a locally hosted, open-source chatbot system tailored to the needs of higher education institutions in the Canton of Bern. The system aims to provide secure, transparent, and ethically aligned GenAI access for students and teachers. Beyond its technical implementation, the project explores the broader pedagogical, ethical, and sustainability implications of locally deployed AI systems.

Objectives

Be(e)Chat aims to design, implement, and evaluate a secure, privacy-preserving, and locally operated GenAI chatbot for higher education. The objectives are to

1. provide an open-source alternative to commercial AI tools,
2. ensure institutional control and data protection,
3. assess system performance and ethical safeguards, and
4. create reusable resources—including hosting and maintenance guides— for the broader educational community.

Method

The project followed a participatory, iterative design and evaluation process. In an interdisciplinary workshop with students, lecturers, IT experts, and institutional stakeholders, requirements and risks were identified. These informed the development of a model-agnostic system architecture using “Open WebUI” and “Ollama” as open-source core components. A prototype was tested using a structured, qualitative evaluation framework across seven key functional areas—such as summarization accuracy, ethical guardrails, and web search reliability—based on 15 standardized test cases. In addition, load testing was conducted to assess system scalability and response times under concurrent usage.

Results

The Be(e)Chat prototype consists of a locally deployed, open-source chatbot system built on Open WebUI as user interface and Ollama as model inference engine. It integrates functions such as document upload and analysis, web search, output-based energy consumption statistics, ß-removal, and institutional authentication via LDAP. The system architecture allows the flexible integration of new and different large language models to ensure sustainability and adaptability. Quantitative evaluation using 15 standardized test cases across seven key functional areas yielded a 73.3 % compliance rate. Qualitative analysis highlighted strong performance in summarization, ethical guardrails, and data security, while revealing limitations in web search reliability and gender-inclusive text generation. Overall, the findings confirm the technical feasibility and pedagogical potential of secure, locally hosted GenAI systems as a transparent and privacy-preserving alternative to commercial AI services.

Implemented Translation

The Be(e)Chat system has been deployed institutionally within the Bern University of Applied Sciences, and made available to lecturers and students for educational use at the Department of Engineering and Computer Science (BFH-TI). In the course of several follow-up projects, educators begun integrating the chatbot into teaching and assessment contexts, for example, by building a “module-bot”, providing students with curated learning resources. All project deliverables, including the system documentation, evaluation framework, and deployment guides, have been published on the project website and/or as a publication at the EC-TEL 2025 conference. Future research directions include broader institutional rollout and partnerships with other higher education institutions in the Canton of Bern.

As a direct outcome of the project, BFH-TI decided to establish a dedicated local LLM inference service based on the Be(e)Chat architecture. For that, BFH has invested in energy-efficient high-performance hardware, contributing to a more sustainable application of AI. The published evaluation framework might be adopted by other researchers to benchmark the application of GenAI tools in higher education settings. In addition, we conducted a workshop at EC-TEL 2025 where participants applied the Digital Ethics Canvas (Hardebolle et al., 2023) to Be(e)Chat-based educational use cases and collaboratively identified potential vulnerabilities and mitigation strategies (see Media).

Participating Institutions