美国数据创新中心（CDI）：2024生成式AI内容“标识”新规的局限性及改进策略研究报告（英文版）（13页）.pdf

1、 CENTER FOR DATA INNOVATION 1 Why AI-Generated Content Labeling Mandates Fall Short By Justyna Lisinska and Daniel Castro|December 16,2024 As artificial intelligence(AI)tools become better at creating high-quality contentincluding text,images,audio,and videocritics worry about potential misuses,such

2、 as to spread misinformation,perpetrate fraud,violate intellectual property(IP)rights,and create harmful deepfakes.Some policymakers are proposing a requirement for mandatory labels on all output generated by AI systems so users can distinguish between human-generated and AI-generated content.Howeve

3、r,mandatory labeling,particularly through watermarking,is neither a reasonable nor effective solution to the issues policymakers seek to address.Rather than singling out AI-generated content,policymakers should prioritize building trust within the digital ecosystem as a whole.INTRODUCTION Generative

4、 AI(GenAI)enables users to produce high-quality digital content such as images,text,music,and video.This technological advancement has enriched many creative possibilities,boosted workers productivity,and offered new tools for innovation.1 Most output from GenAI systems is beneficial and harmless,bu

5、t some policymakers are concerned about the technologys potential misuse,including to spread disinformation through fabricated content,violate IP rights from AI-generated imitations of existing works,and create harmful deepfakes,such as impersonations used to perpetuate fraud or exploitative content

6、 such as unauthorized AI-generated nudes of individuals.2 Policymakers have called for mandatory labeling of all AI-generated content;however,this approach has serious limitations.While labeling AI-CENTER FOR DATA INNOVATION 2 generated content,particularly through watermarking,may help users identi