Two different meanings of “misuse”

Abstract

The term “AI misuse” encompasses two distinct threat models requiring separate analysis and mitigation. The first, “democratization of offense-dominant capabilities,” involves weaker actors gaining access to AI to cause harm that was previously beyond their capacity, such as designing bioweapons. This risk can be addressed with technical solutions, like training models to refuse dangerous requests. The second threat, “power concentration risk,” involves already-powerful actors using AI to consolidate their control, potentially leading to totalitarianism. Technical fixes are less effective against this threat, as the misuse is often authorized by the system’s legitimate owners; solutions instead require interventions like increased transparency and robust governance. Current discourse on AI misuse often conflates these two problems in a motte-and-bailey fallacy, focusing on technical solutions for the democratization risk while using rhetoric that suggests the more severe power concentration risks are also being addressed. This conflation is dangerous, as it leads to overinvestment in technical fixes for the former and an underinvestment in countermeasures for the latter. – AI-generated abstract.