Journal of Artificial Intelligence Research

Social media makes it look like the public is deeply divided. But what if most of the public just isn’t speaking? In a new paper with Atrisha S. just published in the Journal of Artificial Intelligence Research, we show that the polarization we observe online can emerge even when nobody’s opinions have changed. We call it rational silence. Most research on polarization assumes something is pushing people’s views apart: echo chambers, filter bubbles, algorithmic radicalization. Our model shows something different. We don’t change anyone’s opinions. We just let individuals weigh the costs and benefits of speaking up. When rhetoric heats up, two things squeeze moderates out. Allies speaking loudly substitute for your own voice, reducing your incentive to add to the chorus. And intense rhetoric from opponents shrinks the reward you get from expressing your view. Either way, moderates lose the reason to speak. The people who remain are those whose views are extreme enough that the reward from expression still outweighs the cost. It gets worse. We show that ideological media organizations, partisan outlets and political influencers, amplify the effect by signaling that the other side is more extreme than it really is. That makes speaking up feel even less worthwhile for moderates and pushes more of them into silence. Platform recommender systems, optimizing for engagement, then sort people into communities where the loudest voices dominate. Here’s what I worry about. Policymakers and legislators increasingly look to social media to gauge where the public stands. If expressed opinion is systematically skewed toward the extremes, our democratic institutions are navigating with a broken compass. And AI models trained on internet data inherit the same distortion, producing outputs that reflect who spoke up rather than what opinions people actually hold. For every company building AI products on top of that data, this is a problem. We identify practical interventions. Platform moderation that accounts for the intensity of the underlying opinion, not just the intensity of the rhetoric. And recommender strategies that prioritize participatory content over ideological content for users with strong views. One surprise from the model: blanket moderation that raises the cost of rhetoric for everyone can backfire, discouraging moderates more than extremists. The fix isn’t changing what people think. It’s changing who gets heard. https://lnkd.in/ekCnpq6d

I basically never post here, but making an exception today to boast about my recent paper with Charlie Blake published in Journal of Artificial Intelligence Research. A lot of my friends are not on other social media and I would like them to know about it. It is about how winnable patience games (aka solitaire games) are. Three things make it special for me. One is that it's about Patience which is the game my mother taught me to love and we spent many happy hours playing together (possibly happier for me as I wouldn't let her move until I had worked out the best play). Second is that it uses lots of the work I've done throughout my career so feels a bit like putting it all together at last. Third thing is that it is in a journal which has been 100% free to both authors and readers ever since its inception in 1993 - and I also published the third paper in it so have published more than 32 years apart in the same journal. https://lnkd.in/e-q2jrGS

I am pleased to share that our paper, “Label-Aware Pseudo-Training Sample Generation for Text Classification”, has been accepted and published in the Journal of Artificial Intelligence Research (JAIR), a well-recognized Q1 journal in the field of artificial intelligence. The article is available here: 🔗 https://lnkd.in/df_Zc544 I would like to express my sincere gratitude to my advisor, Dr. Seyed AbolGhasem Mirroshandel for his guidance and support throughout this work. I am also thankful to Prof. Owen Rambow for his valuable insights during the research process. In this paper, we introduce a label-aware, embedding-space data augmentation approach that injects learnable artificial tokens into input sequences. Rather than generating new surface-level text, the method operates within the latent space of a frozen LLM to produce diverse, coherent, and label-consistent training samples. Our experiments across multiple benchmark datasets show consistent improvements over standard fine-tuning, highlighting the effectiveness and scalability of this augmentation strategy.