What You Can Learn from 4k Conversations Between Israelis and Palestinians (www.betterconflictbulletin.org)

Adam Becker has run roughly 4,000 one-on-one and small-group conversations between Israelis, Palestinians, and global participants since October 7th and is channeling the lessons into HeadOn (formerly Dugree), an AI-driven platform for “difficult conversations.” His hands-on experiments—A/B testing de-escalation moves in live chats and video calls—revealed consistent patterns: small humanizing gestures (brief personal interruptions, showing mundane domestic scenes) dramatically reduce hostility, text-only exchanges tend to become toxic, and the identity of who speaks often matters more than the exact content. Becker recruited thousands of student “interns,” mixed in outsiders from around the world, iterated on guardrails, and concluded this is fundamentally a data problem: with enough labeled interactions you can design better conversational experiences. For the AI/ML community this is a rare, high-value dataset and a set of concrete design constraints. Multimodal signals (video/audio context) appear crucial for de-escalation, suggesting models should incorporate visual and paralinguistic features, not just text. The project points to practical tasks—predicting escalation, recommending micro-actions, personalizing moderator interventions—and underscores risks: toxicity, context-specific guardrails, and the need for human-in-the-loop oversight. HeadOn’s work shows scalable, evidence-driven conflict-mitigation systems are feasible, but they require ethically curated data, robust evaluation metrics for “de-escalation,” and careful deployment to avoid amplifying harm.