Smitha Milli is a Postdoctoral Associate at Cornell Tech. They received their BS and PhD in Electrical Engineering & Computer Science from UC Berkeley where they were supported by an NSF Graduate Research Fellowship and an Open Philanthropy AI Fellowship. The goal of their research is to create algorithms that help bridge diverse perspectives and facilitate constructive conflict. To do so, they employ a range of methodologies ranging from randomized trials akin to those in the social sciences, to the creation of novel machine learning algorithms, to economic game-theoretic analyses of socio-technical phenomena. Beyond academic outlets, Smitha’s work has been discussed in live television on ABC7 News, in articles published by Tech Policy Press and the Knight First Amendment Institute on the effects of social media, and in testimony to the House Financial Services Committee.

AG 1, AG 2, AG 3, INET, AG 4, AG 5, D6, SWS, RG1, MMCI   Info

AG Audience

English

60 Minutes

Saarbrücken

Machine learning algorithms curate much of the content we encounter online. However, there is concern that these algorithms may unintentionally amplify hostile discourse and perpetuate divisive 'us versus them' mentalities. How can we re-engineer algorithms to bridge diverse perspectives and facilitate constructive conflict? First, I will discuss results from our randomized experiment measuring effects of Twitter’s engagement-based ranking algorithm on downstream sociopolitical outcomes like the amplification of divisive content and users’ perceptions of their in-group and out-group. Crucially, we found that an alternative ranking, based on users’ stated preferences rather than their engagement, reduced amplification of negative, partisan, and out-group hostile content. Second, I will discuss how we applied these insights in practice to design an objective function for algorithmic ranking at Twitter. The core idea to the approach is to interpret users' actions in a way that is consistent with their stated, reflective preferences. Finally, I will discuss lessons learned and open questions for designing algorithms that support a plurality of viewpoints, with an emphasis on the emerging paradigm of bridging-based ranking.

Claudia Richter

+49 681 9303 9103

--email hidden

Cisco In-House

passcode not visible

logged in users only