Sol Messing

Sol Messing

Social Science | ML


Georgetown University



Sol works at the intersection of social science and machine learning. He has led technical teams since 2006 with regular executive communication/public outreach. His work has been featured in the New York Times, NPR, CBS, Wired, FastCompany, TechCrunch, and other outlets and podcasts.

During the 2020 election cycle, Sol was Chief Scientist at ACRONYM, where his team drove high-level strategic investment by correctly modeling the electoral importance of Georgia, and leveraged hundreds of field experiments to build ML systems to produce real-time estimates of persuasive messaging impact–an amazing collaboration with Minali Aggarwal, Sylvan Zheng, Dan Fankowski and James Barnes.

At Facebook, Sol built prototype and production ML models whose offspring power ads, page-recommendations, groups, site integrity, and social science/ethical AI efforts. His team built a differentially private data platform and released a data set reflecting more than an exabyte of media data. He also worked on approaches that blend ML and experimentation (heterogenous effect estimation and contextual bandits).

He also founded the Data Lab at Pew Research Center, where his team used deep learning to produce algorithmic audits of Google Image searches and news photos on Facebook, used ML to study inauthentic and automated behavior on Twitter, and used NLP to understand the role of ideology and power structures embedded in how members of congress use social media to communicate.

Sol’s doctoral work at Stanford University consisted of research on social influence and social structure in digital environments and probably the largest study of algorithmic bias in social media, which was later published in Science and has thousands of citations. He also did work to uncover bias in visual portrayals of African American candidates in political attack ads.

His work on election forecasts sparked public debates about the ethics and consequences of models projecting the results of elections, and prompted FiveThirtyEight to change their user interface to address concerns raised by the work.

Sol is published in Science, the American Political Science Review, the Journal of Politics, Public Opinion Quarterly, Communication Research, Radiology and Pediatric Radiology. His book, The Impression of Influence with Justin Grimmer and Sean Westwood was published by Princeton University Press.


  • PhD Political Communication

    Stanford University

  • MSc Statistics

    Stanford University

  • Intl MBA (Policy emphasis)

    UC San Diego (GPS)

  • BA in Political Science

    UC Santa Barbara


Projecting Confidence

How the Probabilistic Horse Race Demobilizes the Public

Facebook Condor URLs Data Release

Largest ever social science data set, released under differential privacy

Impression of Influence

A book on how Members of Congress use Credit Claiming

Out for Justice

ML & decision support to optimize patrols




Media Clips


Recent Posts

Past vote data outperformed the polls. How did it go so wrong?

It’s becoming clear that the 2020 polls underestimated Trump’s support by anywhere from a 4-8 point margin depending on your accounting–a significantly worse miss than in 2016, when state polls were off but the national polls did relatively well.

It's the Forecasts, not the Polls

It’s becoming clear that the 2020 polls underestimated Trump’s support perhaps by even more than in 2016. As the Florida results poured in, many watching Nate Cohn’s election needles felt the same anxiety that crystalized after the 2016 version gave us the first hint that the polls were wrong.

Trump's chances are better than they look

According to the latest polling research, Trump’s chances of hanging on to power beyond 2020 look pretty dismal. Nate Cohn published an impressive battleground poll from New York Times/Sienna showing Biden ahead of Trump by at least six points in pivotal states.

Why Election Forecasting Matters

Do you remember the night of Nov 8, 2016? I was glued to election coverage and obsessively checking probabilistic forecasts, wondering whether Clinton might do so well that she’d win in places like my home state of Arizona.

Why you should care about privacy policy

In 2011, Sean Westwood and I ran a (IRB-approved) study using Facebook’s graph API to analyze participants’ entire ego network on the fly, and then randomly assign strong v weak ties to ostensibly endorse media content appearing on their news feed (stimulus).