Openness in and with AI

This presentation introduces the topic of openness in and with AI, with a focus on open-source AI in research contexts. It examines the motivations for open AI systems, analyzes different dimensions of openness, and presents practical tools for using open models in research pipelines.

The first part covers the core ideas behind open source and the specific problems of proprietary AI systems for research – particularly regarding reproducibility, transparency and data privacy. The second part analyzes various aspects of openness in AI models (weights, training data, code, documentation, licensing) and discusses the phenomenon of "open-washing". The third part presents concrete services and tools for using open models, including the "ChatAI" LLM by GWDG (Gesellschaft für wissenschaftliche Datenverarbeitung mbH Göttingen) and Ollama for running models locally.

Topics covered:

• Openness and AI: motivation, advantages over proprietary systems, limitations
• Aspects of open AI: open weights, open data, open code, documentation, licensing, open-washing
• Open-source AI in practice: ChatAI (GWDG), Ollama, best practices for LLM use in research pipelines

Code:

• Python experiments with open LLMs and synthetic data that were used during the presentation as well as installation instructions for Ollama can be found on GitHub: https://github.com/UB-Mannheim/research-skills/tree/main/2026_FSS_Openness-and-AI