Hello AoIR colleagues (with apologies for cross-posting),
I’m writing to invite you to Penn MEDIATED’s online convening, LLMs are the Third Wave of Automated Content Moderation. The convening will be held via Zoom on September 10 from 11-12:30 EST and you can register at this link. You can see the overview, agenda, and speaker lineup below—it has researchers (including own affiliated faculty Yphtach Lelkes and Danaé Metaxa) and technologists who will examine how LLMs are changing online platform’s content moderation, how they compare to prior automated systems, and what the emerging research in this space tells us. We’ll also hear from practitioners building a new ecosystem of custom LLM-based moderation systems.
Content moderation is critical to protecting users against harmful and illegal online content, although it may also restrict free expression. Over the past two decades, platforms have increasingly turned to automation to perform content moderation at scale. First through hashing algorithms and later through predictive machine learning methods, many large platforms have automated over 90% of content moderation actions. Today, online platforms are increasingly exploring using large language models (LLMs), as a new component of the content moderation technology stack. With online platforms conducting mass layoffs of trust and safety teams, while also increasing investments in their AI development, this convening will explore how LLMs fit in the content moderation landscape and what that means for platforms, policymakers, and users.
This convening brings together researchers, technologists, and practitioners to examine what this third wave of automated content moderation means in practice. We will explore how LLMs
are being deployed by platforms today, how they compare to prior automated systems, and what the emerging research in this space tells us. We will hear from experts who are building their own custom systems using LLMs for content moderation and researchers
who study their implications for information integrity. The conversation will also examine the policy environment shaping adoption: regulatory frameworks in the EU, Brazil and India, are setting new expectations for the speed and transparency of content moderation
decisions, incentivizing further automation by online platforms. We will share a pre-read working paper on this topic in advance of this event. [Register
at this link]
Agenda
Alex Engler - Welcome and Introduction
Alice Hunsberger - A trust and safety professional’s view
Brief Presentations
Dave Willner & Samidh Chakrabarti - A demo of Zentropi’s content classification tool and how it is different from other LLM-based content moderation approaches
Camille François & Juliet Shen - Overview of open-source tooling for LLM-based content moderation and the importance of cross sectoral collaboration in this space
Audience Q&A
Research Presentations
Professor Yphtach Lelkes and Neil Fasching share findings from their recent paper: Model-Dependent Moderation: Inconsistencies in Hate Speech Detection Across LLM-based Systems
Professors Danaé Metaxa and Sorelle Friedler discuss their recent work: Identity-related Speech Suppression in Generative AI Content Moderation
Audience Q&A
Discussion and Open Questions
Societal implications of LLMs in content moderation (Firestarter: Dia Kayyali
Policy developments are shaping LLM content moderation (Firestarter: Prithvi Iyer)
Closing Remarks
Alice Hunsberger, Head of Trust and Safety, Musubi
Dave Willner, Co-Founder, Zentropi
Samidh Chakrabarti, Co-Founder, Zentropi
Camille François, President, ROOST
Juliet Shen, Head of Partnerships, ROOST
Yphtach Lelkes, Associate Professor, Annenberg School for Communications, UPenn
Neil Fasching, Postdoctoral Researcher, UPenn
Danaé Metaxa, Raj and Neera Singh Term Assistant Professor of Computer & Information Science, UPenn
Sorelle Friedler, Shibulal Family Professor of Computer Science, Haverford College