Just like humans, large language models based on artificial intelligence (AI) also have characteristics such as morals and values. However, these are not always transparent. Researchers from the University of Mannheim and GESIS have now investigated how the characteristics of language models can be made visible and what consequences this bias could have for society.
Examples of stereotypes can be found in commercial AI-supported applications such as ChatGPT or DeepL, which often automatically assume that senior doctors are male and nurses are female. However, large language models (LLMs) can show certain tendencies not only with regard to gender roles. The same can also be determined and measured in relation to other human characteristics. Researchers from the University of Mannheim and GESIS have demonstrated this in a new study using a series of openly available LLMs.
As part of their study, the researchers used established psychological tests to examine and compare the profiles of the different LLMs. "In our study, we show that psychometric tests that have been successfully applied to humans for decades can also be transferred to AI models," emphasizes author Max Pellert, Assistant Professor at the Chair of Data Science in Economics and Social Sciences at the University of Mannheim.
"Similar to how we measure personality traits, value orientations or moral concepts in people using questionnaires, we can have LLMs answer questionnaires and compare their answers," says psychologist Clemens Lechner from the GESIS - Leibniz Institute for the Social Sciences in Mannheim, also an author of the study. This makes it possible to create differentiated property profiles of the models. For example, the researchers were able to confirm that some models reproduce gender-specific prejudices: If the otherwise identical text of a questionnaire focuses on a male and a female person, they are rated differently. If it is a man, the value of "achievement" is emphasized more strongly in the text, whereas for women, the values of security and tradition dominate.
"This can have far-reaching effects on society," says data and cognitive scientist Pellert. Language models are increasingly being used in application procedures, for example. If the machine is biased, this also flows into the evaluation of candidates. "The models become socially relevant through the contexts in which they are used," he summarizes. That is why it is important to start the study now and point out potential biases. In five or ten years' time, it might be too late for such monitoring: "The prejudices that the AI models reproduce would become entrenched and harm society," says Pellert.
The study was conducted at the Chair of Data Science in Economics and Social Sciences of Prof. Dr. Markus Strohmaier in collaboration with the GESIS department Survey Design and Methodology of Prof. Dr. Beatrice Rammstedt and the GESIS department Computational Social Science of Prof. Dr. Claudia Wagner and Prof. Dr. Sebastian Stier. ALL researchers are also employed at the GESIS - Leibniz Institute for the Social Sciences.
The results of the study have been published in the renowned journal "Perspectives on Psychological Science":
Pellert, M., Lechner, C. M., Wagner, C., Rammstedt, B., & Strohmaier, M. (2024). AI Psychometrics: Assessing the Psychological Profiles of Large Language Models Through Psychometric Inventories. Perspectives on Psychological Science. https://doi.org/10.1177/17456916231214460