This paper examines the rise of algorithmic harms from AI, such as privacy erosion and inequality, exacerbated by accountability gaps and algorithmic opacity. It critiques existing legal frameworks in the US, EU, and Japan as insufficient, and proposes refined impact assessments, individual rights, and disclosure duties to enhance AI governance and mitigate harms.
“... we argue there are good reasons for skepticism, as many of its key operative provisions delegate critical regulatory tasks to AI providers themselves, without adequate oversight or redress mechanisms. Despite its laudable intentions, the AI Act may deliver far less than it promises.”
This workbook addresses the challenge of defining AI fairness, proposing a context-based and society-centered approach. It emphasizes equality and non-discrimination as core principles and identifies various types of fairness concerns across the AI project lifecycle. It advocates for bias identification, mitigation, and management through self-assessment, risk management, and fairness criteria documentation.
The AI Act, initially overlooking multifunctional AI like foundation models, led to debates. Industry sought exemption, civil groups pushed for inclusion, foreseeing safety gaps and burdens on users. "General Purpose AI systems" (GPAIS) emerged in discussions, aiming to extend Act requirements to adaptable models, addressing operator responsibility. Current debate focuses on adapting the Act to cover these advanced AI, revealing its initial limitations. The paper will delve into this evolution, highlighting challenges and proposing policy adjustments for GPAIS regulation within the AI Act's framework.