73rd Helmholtz Open Science Online Seminar

Background

Artificial intelligence (AI) is having an increasingly significant impact on science and is finding its way into a growing number of research workflows. At the same time, AI models and the processes by which they are developed are often only partially transparent and are therefore frequently perceived as a "black box". To ensure reproducibility, transparency and trust in AI-assisted research, open science principles such as transparency and openness must be applied consistently, both when conducting research with AI and when developing and providing AI models themselves.

But when is an AI model actually open and reproducible? According to the Open Source AI Definition, several components must be openly available for a model to qualify as open source: in addition to the model parameters (e.g., the weights), this includes the complete source code, for instance for data processing and training, as well as detailed information about the training data. Many models described as open, however, only release their weights ("open weights"). The lack of training code and information about the training data makes reproducibility more difficult.

In this event, we explore to what extent open AI models promote the reproducibility and transparency of research, and the challenges researchers face when developing, publishing, and using such models. Following an introduction, we will approach the topic from three perspectives: reusability in AI, the open publication of AI models, and quality assurance.

Program