Empowering Community-Driven Determination of Values for Language Models

Abstract

Emerging technologies like Artificial Intelligence and Large Language Models are often developed in Western contexts and carry implicit values, from developer choices or underlying training data, which are not adequately representative of the diverse contexts in which they are deployed. The resultant misalignment from the lack of engagement with non-Eurocentric value paradigms results in inadequate, and potentially harmful outcomes that impact these unconsidered communities. To codify fundamentally subjective human values therefore necessitates the elicitation of these nuances through the inclusion and involvement of these very communities.

This thesis argues that participants’ lack of familiarity with new technologies like Artificial Intelligence impacts their engagement and contribution to participatory processes of AI development. This thesis also helps demonstrate how grounded theory approaches can be leveraged to contextualize awareness-building efforts that can potentially empower community participation by addressing such familiarity gaps.

This two-fold objective of (i)eliciting community-relevant attributes for language model alignment (ii)through the necessary familiarization of the technology in question is demonstrated through the means of sample case studies. A grounded participatory process CALMA (Community-aligned Axes for Language Model Alignment) is designed and evaluated through these cases to illustrate this contextualized alignment exercise. Learnings from this comparative case study are then extended to explore avenues for communities and institutions to adopt similar techniques that center the voices of the final users.