The authors of the discussion paper argue for a nuanced view of technologies and tools that make generative AI more transparent and aim to discover and minimise distortions. They discuss dealing with bias as an example: Without an active attempt to counteract it, AI systems reflect the respective societal and cultural relations of their database and the values and inequalities contained therein. However, according to the authors, deciding whether and how to actively counteract this bias in the programming is no trivial matter. It requires technological and mathematical, as well as political and ethical expertise and should not be the sole responsibility of developers.
Strategies used to date to counteract the lack of transparency of generative AI also offer only a rather superficial solution. Users are often unable to understand how generative AI works. The still-new research field known as explainable AI develops procedures that aim to make AI-generated suggestions and decisions comprehensible retrospectively. However, the authors point out that the resulting explanations are also not reliable, even if they sound logical. It is even possible to deliberately manipulate explainable AI systems. The authors thus stress that generative AI should be used and developed with the utmost caution in cases where transparency is essential (for example in legal contexts).
They also elucidate the various possibilities for deception with respect to generative AI, for example when users are unaware that they are communicating with AI, or when they do not know what AI is or is not capable of. Users often attribute human capabilities such as consciousness and comprehension to AI. The quality, ease, and speed with which texts, images, and videos can now be generated creates new dimensions for possible misuse, for example when generative AI is used for propaganda or criminal purposes.
The discussion paper also addresses the issue of data protection. The success of generative AI is based partly on gathering and analysing users’ personal data. However, to date there is no convincing approach to ensure that users have the final say when it comes to the sharing and use of their data.
Publications in the series “Leopoldina Discussion” are contributions by the authors named. The Academy’s discussion papers offer scientists the opportunity to present thought-provoking ideas and to encourage and guide discussions flexibly and outside of formal working group processes.
The discussion paper was published in German on 17 October 2024 and an English translation is now available. It was created by the philosopher Professor Dr Judith Simon, Professor of Ethics in Information Technology at the University of Hamburg/Germany, the law expert Professor Dr Indra Spiecker gen. Döhmann, Professor of Digitalisation Law at the University of Cologne/Germany, and Leopoldina Member Professor Dr Ulrike von Luxburg, computer scientist and Professor of the Theory of Machine Learning at the University of Tübingen/Germany. The three researchers are members of the Leopoldina Focus Group “Digitalisation”.