Tushita Jha

Policy and academic discourse around AI often invokes concerns about AI systems’ effects on ”human autonomy”. A growing literature also highlights the priority for augmentation, extension, or amplification of human agency. How ever, as prior authors have pointed out, the concept of ’autonomy’is vague and heterogeneous; while on the other hand the character of ’agency’ that is recommended to be augmented remains in need of deeper philosophical attention. We outline four dimensions for evaluating how artificial assistants may affect autonomous human agency, and argue that these distinct dimensions suggest potentially different practical interventions. More specifically, we distinguish choice availability (the size of an agent’s available action set), volitional integrity (the fit between an agent’s actions and her evaluatively endorsed intentions), referential integrity (the accuracy of beliefs whose accuracy the agent takes to matter), and authenticity (the degree to which evaluative attitudes are the agent’s own). We then analyze the prior literature in light of our four dimensions, and argue that our framework allows for a richer analysis of the potential effects that artificial assistants may have on human autonomy. Then, we apply our framework to related discussions on epistemic agency, recommender systems and on human disempowerment from advanced AI, showing that each conflates analytically distinct agency-impairments. From the framework we derive five maxims for the development and regulation of artificial assistants, thereby centering norms of extension of agency.

As AI capabilities grow, the field faces a choice in how to define success. We argue that the primary objective of AI research should be Human Amplification -- aiming for systems that augment human abilities and preserve human agency. While the current focus on intelligent agents drives significant progress, treating replication of human agency as the ultimate metric of success emphasizes delegation over augmentation. We discuss how this presents distinct challenges regarding control and power dynamics. By contrast, anchoring AI development in the service of human agency opens up a much broader design space that blurs the distinction between non-agentic tools and sophisticated delegates. This shift directs research toward cognitive extensions that ideally spark a virtuous cycle of recursive human amplification, where advances in AI further empower us to build safer and more beneficial systems.