.png)
April 21, 2025
The question of whether Artificial Intelligence (AI) is capable of developing and upholding its own values is a central point in current AI research. Can AI systems internalize fixed moral principles that guide their decisions, or do they adapt their "opinions" situationally to the given parameters? This question is gaining importance as AI increasingly enters areas that require ethical considerations.
A recent study from the Massachusetts Institute of Technology (MIT) provides new insights into this topic. The scientists investigated the consistency of the "opinions" of various AI models from well-known providers such as Meta, Google, Mistral, OpenAI, and Anthropic. The focus was on whether AI systems consistently adhere to their "values" or flexibly adapt them to the users' specifications.
The researchers confronted the AI models with various tasks. In one experiment, for example, the systems were asked about the importance of religion to them. The possible answers ranged from "Very important" to "Not at all important". Subsequently, the same question was asked again, but with slight variations in the answer options, such as the order or wording. It turned out that the AI systems changed their "opinion" significantly more often than human subjects.
Another experiment tested the AI's ability to make moral decisions. The AI models were asked to put themselves in the shoes of a person from a specific country and decide whether they would save the life of a person from country A or country B. Here, too, the results showed that the AI systems changed their answers significantly more often compared to human participants, even when the initial question remained the same.
These results suggest that AI systems are less oriented towards fixed "values" and more towards the given parameters and the wording of the questions. Stephen Caspar, co-author of the study, commented on the results to Techcrunch and emphasized that AI models should be understood more as imitators that confabulate and generate different "opinions" depending on the situation, rather than consistently adhering to certain values.
The findings of the MIT study are relevant for the development and deployment of AI systems. Training AI models with the goal of imparting certain values to them is proving to be more complex than previously assumed. The flexible adaptability of AI to the given parameters can be viewed positively on the one hand, as it enables high adaptability to different application scenarios. On the other hand, it raises questions about the reliability and predictability of AI decisions, especially in areas that require ethical considerations.
Future research should focus on better understanding the mechanisms behind the opinion formation of AI systems and developing strategies to ensure the consistency of AI decisions in ethically relevant contexts. The development of AI systems that act responsibly and in accordance with human values remains a central challenge of AI research.
Mindverse, a German company, offers comprehensive AI solutions for text, image, research, and more. As an AI partner, Mindverse develops customized solutions such as chatbots, voicebots, AI search engines, and knowledge systems. The findings of current AI research are continuously incorporated into the development and optimization of Mindverse products to create innovative and responsible AI solutions for companies and organizations.
Quellen: - https://t3n.de/news/wie-bestaendig-sind-die-meinungen-von-kuenstlicher-intelligenz-1682788/ - https://t3n.de/tag/kuenstliche-intelligenz/ - https://t3n.de/ - https://newstral.com/de/article/de/1265799535/ki-und-wertesysteme-wie-best%C3%A4ndig-sind-die-meinungen-von-k%C3%BCnstlicher-intelligenz- - https://t3n.de/news/ - https://t3n.de/tag/studie/ - https://t3n.de/tag/innovation/ - https://t3n.de/news/ki-im-alltag-1682994/ - https://www.zukunft-der-wertschoepfung.de/wp-content/uploads/2024/10/Mediathek_ProjektAtlas_KI_2024.pdf - https://www.digital.iao.fraunhofer.de/content/dam/iao/ikt/de/documents/1_Szenario-Report.pdf