The Value Foundation of Chatgpt

The objective of this research paper is to analyse and differentiate the value foundation of the large language model ChatGPT in more detail. Previous studies on this topic assume a rather progressive value foundation of ChatGPT, which is caused by the training data used to train the LLM. In order to investigate the value foundation in more detail, a vignette experiment was conducted with ChatGPT in which the programme was asked to evaluate a statement on the division of labour in the household in the role of different people with varying socio-demographic characteristics. In order to show a concrete value foundation and thus indirectly a bias of ChatGPT, the results of the experiment are compared both descriptively and by means of multivariate linear regressions with real survey results of the International Social Survey Programme from 2012. The results of this paper suggest that ChatGPT is subject to a progressive bias. Furthermore, it could be shown that ChatGPT is neither able to correctly predict other value foundations nor to differentiate attitudes within the progressive value stance. Thus, this paper shows that ChatGPT does not objectively evaluate all values and attitudes equally, which results in the user being influenced by the programme.