This being my Master’s thesis, I decided combine the work that I had completed for my Graduate Certificate in Ethics, and thus provide a philosophical defense and justification for the analysis of value within multimodal machine learning systems. As a result, my thesis became split into four distinct chapters:
- Chapter one seeks to clarify what is meant by moral value and how values can be intrinsic. Upon establishing this, I argue for the intrinsic value of certain actions and, by extension, interactions.
- Chapter two provides a defense against a common argument against moral realism, termed an Evolutionary Debunking Argument (EDA). In particular, I intentionally do not provide a comprehensive argument for moral realism, seeing as that is outside of the scope of this thesis; however, my defense provides a foundation for my argument within the next chapter.
- Chapter three extends the first two chapters by exploring the ways that the value of certain concepts (and in particular forms of media) can influence rational reflective agents. In particular, media can influence our reactions and judgments, two related but distinct states of rational reflection.
- Finally, chapter four provides empirical evidence of a lack of cross-modal consistency in the latent representation of moral value within vision-language models (VLMs). Furthermore, it also shows that many models display a degradation in their cross-modal consistency of moral value, as they scale. This leads to my closing my thesis by arguing for the prioritization of cross-modal consistency for abstract concepts such as moral value as multimodal models continue to scale.