Abstract
As generative artificial intelligence systems such as ChatGPT become increasingly embedded in professional and academic communication, questions of openness, collaboration and accountability gain renewed urgency. Language plays a central yet often underestimated role in science-policy relations, particularly in multilingual and multi-level governance contexts such as higher education. The UniTermGPT project examines how ChatGPT handles university terminology across different German language varieties (Austrian, German, Swiss and South Tyrolean German) and addresses the implications for science-policy interfaces, professional practices and responsible artificial intelligence governance.
Open innovation in science research emphasizes that societal impact depends not only on technological capability but also on the conditions under which knowledge is produced, (re-)used and circulated. In higher education, university terminology is politically and institutionally situated: it reflects national legal frameworks, institutional structures and policy traditions. Terms designating study-related or organisational concepts often differ substantially across countries (even within the same language and despite the harmonization efforts within the European Higher Education Area). When these terms are used inaccurately, inconsistently or are “normalized” by large language models, important distinctions risk being erased, potentially affecting interpretation and causing misunderstandings.
UniTermGPT addresses this challenge by combining open research practices with collaborative knowledge production across disciplinary and professional boundaries. The project compiles an open corpus of university texts from Austria, Germany, Switzerland and South Tyrol and extracts and compares terminology across these university systems. These data are contrasted with existing terminological resources and used to evaluate how ChatGPT processes language-variety-specific terminology under different prompting conditions. Selected texts are translated into English and back into German varieties (and vice versa), with outputs annotated and assessed by expert translators and terminologists (working in the university domain).
Going beyond basic research, UniTermGPT places a strong emphasis on societal impact by explicitly linking empirical research to questions of governance, professional practice and policy relevance. The project foregrounds collaboration as a prerequisite for meaningful openness, positioning translators and terminologists not merely as annotators but as co-producers of knowledge about ChatGPT’s performance and limitations. Their role extends across the research process (from jointly defining annotation methods and tools to co-developing policy briefs and practical recommendations) thereby embedding professional expertise directly into the production of policy-relevant insights. Such collaboration aligns with the principles of open innovation in science by treating translators not as end users, but as co-creators of application-oriented academic knowledge.
A key output of the project is a policy brief that translates empirical findings into recommendations for policymakers and technology providers, among others. The brief argues for responsible integration of large language models into text generation and translation workflows (in a university context) that foreground the importance of system-bound terminology and language varieties as well as the relevance of language professionals in an artificial intelligence era.
The project also speaks directly to the profession of translation and terminology work. Rather than framing large language models as a disruptive replacement, UniTermGPT demonstrates that professional expertise and domain knowledge are increasingly valuable. Translators and terminologists play a crucial role in curating data, quality assuring (domain-specific) outputs and technology development. The policy brief advocates for continued investment in these professions, recognizing them as essential actors in maintaining language diversity in the digital age.
By publishing data, methods and findings openly, the project contributes to transparency in a domain where large language models are increasingly used but rarely scrutinized in policy-sensitive contexts. Beyond the university context, the insights generated by UniTermGPT are transferable to other policy-relevant domains characterized by language-variety-specific and system-bound terminology, such as law, healthcare and public administration. In this way, the project illustrates how open and collaborative research can inform responsible use of large language models, strengthen science-policy relations and enhance the societal impact of academic knowledge.