A Digital Humanism perspective on providing language resources to CLARIN in an age of AI commodification: The case of UniTermGPT

Abstract

The accelerated uptake of large language models (LLMs) has intensified the demand for highquality language resources, while simultaneously reshaping the political economy of language data. Research infrastructures such as CLARIN play a central role in advancing open science by providing FAIR-compliant access to language resources, yet they also operate within a context increasingly characterized by AI-driven commodification and extractive data practices. This paper adopts a Digital Humanism perspective to examine the ethical implications of providing language resources to CLARIN in this context, using the UniTermGPT project as an example. UniTermGPT compiles and annotates German-language university corpora and terminology across Austrian, German, Swiss and South Tyrolean varieties to address limitations of large language models in handling language-variety-specific terminology. By discussing the types of data produced by UniTermGPT, their integration into CLARIN, this paper discusses the challenges associated with openness, attribution and potential downstream reuse in opaque AI training pipelines. By situating FAIR and CARE principles within a broader Digital Humanism framework, the paper argues for ethical-by-design approaches to language resource provision that make human labor visible, preserve linguistic diversity and address power asymmetries between public research infrastructures and commercial AI actors.