Switzerland in 2050. A Conversation with Gemini for the weekly Magazine Tagesanzeigermagazin Zürich
Claudio Bucher and I did not employ strictly scientific methods in this study, as we are not AI experts, but merely enthusiastic users. AI is neither an artificial expert nor can it foresee the future. AI is a product of our time that is evolving at a rapid pace and, accordingly, is shaping how we experience, assess, and influence reality. It draws on a vast archive of human knowledge with which it has been trained. This knowledge is a product of the past—at best, the
very recent past.
In this respect, an AI prediction says at least as much about the present as it does about the future. What makes studies of this kind interesting, however, is that AI processes this knowledge in a much shorter time than we can. Through this process of statistical research in the AI’s library, scenarios can be developed, for example, for Switzerland in the year 2050.
In our case, AI refers to so-called large language models that have been trained to mirror, generate, and process human language. As is well known, artificial intelligence is not infallible. The system is influenced by the way questions are phrased and by the specific questions asked. The system is also not (yet) completely stable. Even if I repeat the same question, I may receive differing answers. The system does not yet have a reliable long-term memory. In extensive dialogues, it may take inconsistent positions because it does not remember its previous statements. For our study, we tested various models (Claude, ChatGPT, Gemini) and observed that they provide very similar answers. This speaks to their relative consistency. We ultimately focused on Gemini and, after various tests on key societal parameters such as society, politics, the economy, and
culture, decided to forego detailed prompts.
Instead, we asked the model about the ten most significant and radical changes in Switzerland in 2050 compared to today. To assess the likelihood of the answers, we repeated the central question several times, thereby challenging the model regarding the plausibility of its
answers. This resulted in a large body of material. We removed only repetitions from the dialogue and deleted overly lengthy explanations, but made no substantive changes to the response texts.
The AI speaks “verbatim.”
