Ayse Gizem Yasar (École Normale Supérieure (ENS)) has posted Manipulation as a Systemic Risk under the AI Act: Gaps and Enforcement Challenges on SSRN. Here is the abstract:
Manipulation is a growing concern in AI research and regulation. AI manipulation scenarios range from the use of generative AI in phishing campaigns to highly capable AI models going off the rails and inciting self-harm. Where general-purpose AI models are integrated into downstream AI systems, their manipulative features may be propagated across countless use cases. This is one of the reasons why harmful manipulation is considered a systemic risk under the EU’s AI Act. Indeed, the AI Act’s systemic risk provisions are seemingly capable of addressing manipulation concerns in GPAI models. However, a closer inspection of the regulation and supporting policy documents reveal two important enforcement challenges. The first is the formulation of AI Act’s systemic risk provisions and the policy guidance that operationalises these provisions. The systemic risk framework of the AI Act, with its focus on ‘high-impact capabilities’, is drafted in a way that creates confusion about the types of risks that the regulation is intended to address. Some of the most prevalent forms of GPAI-driven manipulation may therefore fall through the cracks. The second set of challenges emanate from the interaction between the AI Act and the Digital Services Act’s systemic risk regimes. The DSA’s list of systemic risks are expansive enough to cover at least some manipulation risks in GPAI models. It may be perceived as a backstop to address manipulation risks that the AI Act fails to catch. That said, the DSA should not be the first port of call to catch manipulation risks in AI models. It is first and foremost an intermediary regulation and was not designed in a world where platforms create their own content with AI models. The paper concludes with a brief discussion on how GPAI-driven manipulation may be adequately addressed as a systemic risk under the AI Act. The overall objective of this paper is to call attention to the immediacy of GPAI-driven manipulation as a systemic risk that should be addressed as comprehensively as possible under the EU’s digital rulebook.
To receive new posts from Legal Theory Blog by email, get a free subscription to Legal Theory Stack.
Lawrence Solum