A collaboration of international researchers has developed a new AI model that is not only freely accessible and multilingual, but also focusses on safety. As part of this global initiative, Felix Friedrich (PhD student at the Artificial Intelligence and Machine Learning Lab led by Professor Kristian Kersting at TU Darmstadt) has concentrated on “safety tuning” (optimising safety). In doing so, he ensures that the AI not only corresponds to our values, but is also protected against misuse.

The US Biden-Harris Executive Order on Artificial Intelligence was used as an important compass and point of reference for optimising safety. This legal text places a clear focus on the safety and accountability of AI technologies and serves as a reference to establish a democratically legitimised standard for our values and safety.

Another special feature of the new AI model is its complete open-source availability and its ability to process multiple languages. “This step is crucial, as AI technologies are increasingly shaping our everyday lives and it is essential that they are accessible to a wide range of people from different languages and cultures,” says Friedrich.

The use of previously neglected languages such as Hindi, Vietnamese or Finnish in particular represents a fundamental improvement in accessibility. By disclosing the source code, the developers are enabling researchers around the world to learn from and build on the progress made, which will ultimately further advance the development of AI technology as a whole.

There is an important connection to the “Reasonable Artificial Intelligence (RAI)” project, which has been invited to submit a full application in the “Cluster of Excellence” funding line as part of the Excellence Strategy of the German federal and state governments. RAI has set itself the goal of developing and promoting AI technologies that are not only effective and efficient, but also prioritise ethical principles such as safety, fairness and transparency. The new AI model is a direct result of these efforts and emphasises the TU's commitment to responsible AI research and development.

Friedrich/bjb