Axiological Foundations of Generative AI Systems: An alignment beyond Human Values

SIN RESUMEN

The main objective of this research is to critically examine and strengthen the axiological (values-based) foundations of GAI systems. In particular, the project aims to identify the values (explicit or implicit) that current state-of-the-art generative models operate under, evaluate the alignment of these values with widely endorsed human and societal values, and develop a framework or set of guidelines to improve value alignment in future generative AI systems.

OE1: Determine and map out the explicit design principles or value guidelines that have been used in the development of systems like ChatGPT, Gemini, and Claude (e.g., safety rules, content policies, RLHF reward models). OE2: Empirically evaluate the behavior of these AI systems in ethically salient scenarios to infer the values that appear to guide their outputs. OE3: Compare the models’ responses and underlying policies with established human values and ethical frameworks. Identify areas where the AI’s values seem misaligned or insufficient OE4: Gather insights from key stakeholders –including philosophers, linguists, management scholars and sociologists– on what values they believe GAI should embody.

Investigacion basica

Disciplinario

CONCURSO ANUAL DE INVESTIGACIÓN

• 2 — Administración

• 10 — Ciencia computacional

• 52 — Filosofía práctica (moral, social y política)

Ciencias sociales - Otras ciencias sociales - Otras ciencias sociales

PONTIFICIA UNIVERSIDAD CATÓLICA DEL PERÚ