Why the AI Act Fails to Understand Generative AI

Presented at We Robot 2023, Boston University School of Law, Winner of the Best Paper Award

41 Pages Posted: 26 Dec 2023 Last revised: 15 Jan 2025

See all articles by Claire Boine

Claire Boine

Washington University School of Law; University of Ottawa Faculty of Law; Artificial and Natural Intelligence Toulouse Institute

David Rolnick

McGill University

Date Written: June 30, 2023

Abstract

The European Union's Artificial Intelligence Act (AI Act) represents a pioneering attempt to regulate AI technologies. However, this paper argues that the Act's framework is inadequate for addressing the challenges posed by generative and general-purpose AI systems. Through a critical examination of the Act's origins and development, this paper offers an account of how the EU's definition of AI shaped its perception of potential harms, leading to a misguided focus on dataset-related issues and an overreliance on the concept of "intended purpose" borrowed from product safety law.

The paper develops its argument in three parts. First, it traces the evolution of AI's definition in European institutions, demonstrating how the shift from viewing AI as autonomous agents to non-autonomous statistical tools influenced the regulatory approach. Second, it explains why the Act's original risk classification framework fails to adequately address generative AI, leaving most such systems unregulated and struggling to prevent various forms of harm. Finally, it analyzes the Act's late attempt to incorporate general-purpose AI systems, resulting in a complex regulatory framework that still fails to account for most harms from generative AI.

This analysis reveals that the AI Act, while groundbreaking, is in some ways obsolete before coming into effect. The paper concludes by proposing several solutions, including regulating systems over models, implementing universal risk assessments and red-teaming exercises for all generative AI systems, and establishing robust disclosure requirements. These insights offer valuable lessons for AI regulation efforts in other jurisdictions, emphasizing the need for adaptable approaches that can evolve alongside the technology they seek to govern.

Keywords: General Purpose AI Systems, AI Act, GPAI, Foundation models, AI law, AI regulation, Generative AI, EU law

Suggested Citation

Boine, Claire and Rolnick, David, Why the AI Act Fails to Understand Generative AI
(June 30, 2023). Presented at We Robot 2023, Boston University School of Law, Winner of the Best Paper Award, Available at SSRN: https://ssrn.com/abstract=4644701 or http://dx.doi.org/10.2139/ssrn.4644701

Claire Boine (Contact Author)

Washington University School of Law ( email )

St Louis, MO
United States

University of Ottawa Faculty of Law ( email )

2292 Edwin Crescent
Ottawa, Ontario K2C 1H7
Canada

Artificial and Natural Intelligence Toulouse Institute ( email )

University of Toulouse
Toulouse
France

David Rolnick

McGill University ( email )

1001 Sherbrooke St. W
Montreal, Quebec H3A 1G5
Canada

HOME PAGE: http://davidrolnick.com/

Do you have a job opening that you would like to promote on SSRN?

Paper statistics

Downloads
353
Abstract Views
851
Rank
176,133
PlumX Metrics