OpenAI：2023生成式语言模型和自动化影响行动：新兴威胁和潜在缓解措施（英文版）（84页）.pdf

1、Generative Language Models and AutomatedInfluence Operations:Emerging Threats and Potential MitigationsJosh A.Goldstein1,3,Girish Sastry2,Micah Musser1,Rene DiResta3,Matthew Gentzel2,and Katerina Sedova11Georgetown Universitys Center for Security and Emerging Technology2OpenAI3Stanford Internet Obse

2、rvatoryJanuary 2023Workshop Participants:Steven Adler,Shahar Avin,John Bansemer,Chris Bregler,Miles Brundage,SamGregory,Shelby Grossman,Ariel Herbert-Voss,Yacine Jernite,Claire Leibowicz,Connor Leahy,Herb Lin,Drew Lohn,Meg Mitchell,Amnon Morag,Alex Newhouse,Helen Ngo,Aviv Ovadya,Cooper Raterink,Yoel

3、 Roth,Bob Rotsted,Elizabeth Seger,and Raymond Serrato.Acknowledgements:We thank participants in the October 2021 workshop that we convened for inform-ing our understanding of various threats and mitigations.We also thank many workshop participantsfor providing feedback on a draft of this paper.For a

4、dditional feedback on the paper,we thank DeepeshChaudhari,Jeff Ding,Tyna Elondou,Shengli Hu,Daniel Kokotajlo,Gretchen Krueger,Pamela Mishkin,Ronald Robertson,Sarah Shoker,Samuel Wolrich,and Jenny Xiao.Josh Goldstein began working on theproject as a postdoctoral fellow at Stanford,and continued work

5、as a research fellow with GeorgetownCSETs CyberAI Project.Matthew Gentzel completed his contributions while contracting for OpenAI,andis now at Longview Philanthropy.Katerina Sedova completed her contributions to this project while shewas a research fellow with Georgetown CSETs CyberAI Project and b

6、efore she entered U.S.governmentservice.All errors remain our own.Lead authors contributed equally.arXiv:2301.04246v1 cs.CY 10 Jan 2023ContentsExecutive Summary11Introduction51.1Motivation.51.2Threats and Mitigations.61.3Scope and Limitations.71.4Outline of the Report.82Orienting to Influence Operat