Voluminous yet Vacuous? Semantic Capital in an Age of Large Language Models

LLMs risk eroding "Semantic Capital" — the collective knowledge in our digital ecosystem — which poses broader social and epistemic challenges beyond typical AI safety concerns, requiring us to reconsider how we interact with and perceive these technologies.

A "deontological tipping point" may be approaching in our AI-driven information landscape, meaning we need to move beyond simply following AI ethics guidelines and regulations to actively understand the spectrum of social epistemic risks LLMs introduce to our shared knowledge.

The paper maps specific semantic risks using Luciano Floridi's taxonomy, identifying how LLM capabilities and limitations create particular vulnerabilities in our collective knowledge systems — though the abstract does not detail which specific risks pose the greatest concern.

Human-AI collaboration should augment rather than replace human intelligence, suggesting the goal should be designing LLM interactions that enhance human knowledge and decision-making rather than substitute for human judgment and expertise.

Tracking and controlling LLM impacts remains challenging, indicating that despite the urgency of addressing these risks, we currently lack adequate mechanisms to monitor how these models affect our information ecosystem and semantic capital.