Beyond Scalable Diversity: Foundation Models and Epistemic Justice

• Tuesday, December 1^st / 2–4 pm / C40.530

• This event is in English.
• This event is in-person with a hybrid option. Contact us to receive the Teams Meeting link.

• Registration is not necessary.

Foundation models, including large language models, are widely heralded as a transformative breakthrough in artificial intelligence. They are commonly presented as scalable, multimodal infrastructures capable of transferring knowledge across tasks and domains. In this contribution, I challenge their implied universality by approaching modelling as a performative practice: models do not merely represent a pre-existing world but configure what becomes intelligible, actionable, and governable. Examining language models, text-to-image systems, and a new generation of machine-learning “world models,” I show how uneven data resources and standardised architectures compress heterogeneity into dominant epistemic frames. 

Linguistic diversity makes this dynamic particularly visible. Language-modelling bias causes under-resourced languages to be represented, interpreted, and processed with less nuance, precision, and efficiency, flattening lexical and cultural distinctions through Anglocentric ground truths. Related distortions extend into visual models and ethnocentric representations of space and time. Drawing on anthropologist Anna Tsing’s distinction between scalable and meaningful diversity, I argue that adding languages, identities, modalities, or data to pre-trained architectures does not in itself transform the epistemic ordering they enact. Epistemic justice instead requires recognising diverse communication systems and knowledge practices as distinct ways of seeing, knowing, and world-making and enabling them to reconfigure model design and evaluation regimes fundamentally. 

Through illustrative cases in which pornographic platform taxonomies become embedded in foundation models and machine translation shapes migration decisions, I demonstrate that these stakes are not merely representational but socio-material such as when linguistic differences are silenced, distorted, or mistranslated in automated asylum procedures, people may be denied credibility, care, protection, resources, and rights. 

Prof. Dr. Paula Helm holds the Chair of Empirical Computational Ethics at the Center for Critical Computational Studies (C3S) and the Institute of Sociology at Goethe University Frankfurt. Working at the intersection of science and technology studies (STS) and responsible AI, her research spans diversity and language technology, platform economies, privacy, internet addiction, the infrastructural ethics of foundation models, and the use of AI in high-stakes social contexts. 

Her interdisciplinary profile is reflected in her active involvement in the German and European STS communities as well as the international FAccT and AIES communities. She has served as an area chair for ACM/AAAI AIES, is a member of ELLIS Europe, editorial board member of the Cambridge Forum for AI, and was spokesperson of the inaugural board of stsing e.V. Across previous and ongoing projects, she collaborates with communities and civil-society organisations, including Indigenous communities in the Amazon region and NGOs working on migration, language justice, and digital rights. 

Before joining Goethe University Frankfurt, Paula was Assistant Professor of Ethics and Data Science at the University of Amsterdam, where she led the Empirical Ethics Research Group at the Amsterdam Institute for Advanced Study and coordinated the MA track Cultural Data & AI. Her broader aim is to move AI ethics beyond abstract principles and public-relations commitments by grounding responsible AI in empirical inquiry, infrastructural critique, and the practices and concerns of affected communities, institutions, and engineering teams. 

cdcforum@leuphana.de