Linguistische Berichte Heft 287

The outputs of modern large language models (LLMs) are profoundly shaped by their input, reflecting the ethical standards and biases of their human counterparts. As illustrated by Mi-crosoft’s Chatbot Tay in 2016, autonomous AI systems can pose significant risks when they fail to differentiate between the nature of data they consume and (re)produce. Therefore, it is crucial to critically examine the advantages and disadvantages of generative AI on both indi-vidual and collective levels. This paper offers a historical and conceptual overview by contex-tualizing the evolution of generative artificial intelligence from a linguistic perspective and exploring its potential and challenges. It further aims to build theoretical understanding and elucidate the mechanisms behind LLMs like ChatGPT and some of their undesirable behaviors, such as hallucinations. While the integration of AI into our lives and work is inevitable, pro-moting AI Literacy – the ability to use AI technologies ethically and effectively – is pivotal to this discourse. By combining linguistic theory and practical insights, this study sheds light on the transformative role of artificial intelligence in contemporary society.

The current study examines the effect of Dependency Length Minimization (DLM) on clausal extraposition in German, focusing on the positioning of relative clauses (RCs) and finite noun complement clauses (NCCs). Drawing on evidence from newspaper corpus, acceptability judgement, and elicited production data, we explored whether DLM effects are consistent across clause types and modalities. While our results confirm earlier findings showing a bias towards keeping extraposition distance short and effects of clause length, we found that RCs and NCCs differed in their probability of being extraposed, reflecting their distinct syntactic properties. We also found differences among our three data types. While acceptability judge-ments showed an overall preference for integrated clauses, our corpus data show an overall preference for extraposition for NCCs and no clear preference for RCs. The results from our sentence production tasks revealed a strong avoidance of extraposition for RCs but a preference for extraposition for NCCs. Taken together, our findings indicate that the DLM constraint interacts with a dependent clause’s grammatical status and also applies differently across data types. On the methodological side, our findings highlight the importance of considering multi-ple types of data to establish how and to what extent word-order preferences are determined by processing-efficiency biases such as the DLM constraint.

In languages with grammatical gender such as German, the gender feature of third-person singular anaphoric personal pronouns is usually assumed to be controlled by the linguistic antecedent in discourse. In predicational copula sentences of the form “[this/the N1]NP1 is [a N2]NP2” a single discourse referent is usually made available for anaphoric pronominal refer-ence. The gender of the anaphoric pronoun then usually agrees with NP1. However, gender control by NP2 has been claimed to be possible as well in some cases (Löbel 2000 and Mueller-Reichau 2008). We call this phenomenon gender shift. While gender shift has been discussed in the literature on German, the empirical and theoretical picture is still unclear. In this paper, we present two experiments (a production experiment and a follow-up rating exper-iment) that aim to investigate gender shift in German. We show that a) gender shift is indeed available in German, however b) it is usually less acceptable and produced less often as com-pared to other configurations, c) except in what we call, following Mueller-Reichau (2008), classificatory constructions. While our analysis widely follows ideas suggested in Mueller-Reichau (2008), we suggest that a broader theory of re-classification is needed to account for the gender-shift phenomena under investigation. We suggest a possible avenue for such a theory inspired by the study of sign languages.