Berlin Group adopts working papers on LLMs and data sharing

LLMs are highly complex AI models for generating texts. They are trained with large amounts of data. The paper addresses risks to data protection and shows how these can be addressed with an effective understanding of how the technology works.

The Federal Commissioner for Data Protection and Freedom of Information (BfDI), Professor Dr. Louisa Specht-Riemenschneider, emphasizes: High-quality training data are required for training large language models. Where personal data are processed, this must happen in a transparent and lawful manner. There must be a comprehensive analysis for the appropriate handling of LLMs and effective supervision that constructively supports innovations. The Berlin Group's paper describes how this can be achieved.

Sharing data between organizations, especially internationally, can potentially lead to major data protection breaches and data abuse. The advice in the paper on data sharing is intended to contribute to the development of a trustworthy exchange and computing environment that can maximize the potential of safe and secure data exchange while maintaining data protection principles.

Professor Dr. Louisa Specht-Riemenschneider says: In order for data-driven innovations to benefit everyone, it is important to me to point out ways that enable data protection-compliant application and implementation of innovations. In particular, data protection-enhancing technologies, so-called privacy- enhancing technologies play a particularly important role, as the paper on data sharing shows.

The “Berlin Group” is an independent group of experts in the field of technological data protection and is chaired by the BfDI. In its work, the group benefits from its heterogeneous composition with participants from data protection supervisory authorities, government agencies, international organizations and non-governmental organizations, as well as from research and academia from all regions of the world.

Click here to access our working papers Large Language Models and Data Sharing.