At the Aequitas Workshop on Fairness and Bias in AI (Oct 2024, Saint Jacques de Compostelle, Spain), Thierry Poibeau, research partner in the ASTOUND consortium, presented a thought-provoking position paper on bias, subjectivity, and normative frameworks in Large Language Models (LLMs). The paper, titled “Bias, Subjectivity and Norm in Large Language Models,” challenges the prevailing assumptions around bias mitigation in AI and proposes a more nuanced, transparent approach.

In current AI discourse, there is a tendency to treat bias as an error to be eliminated. However, this paper argues that biases in LLMs are not only inevitable but also reflective of broader societal values and norms. Attempts to retroactively “de-bias” models often overlook the deeper question: whose norms and whose fairness are being encoded?

Rather than striving for an elusive neutrality, the authors advocate for increased transparency in how models are filtered and aligned with use-case-specific expectations. This means recognizing that LLM outputs are shaped by multiple layers of subjective choices—from training data curation to deployment context—and that responsible AI should foreground these complexities.

This paper underscores ASTOUND’s broader mission: to develop contextually aware AI systems that are both socially intelligent and ethically grounded. By engaging in the ethics of model development, we move closer to creating AI that not only performs well, but does so in ways that are understandable, fair, and accountable.

Read the full article here: https://cnrs.hal.science/hal-04838836v1