Tech and Democracy

Dated Jan 27, 2020; last modified on Mon, 05 Sep 2022

Political Ads

Cambridge Analytica (CA) paid people to take in-app survey; mined FB profile data including friends' data; crafted tailored sensitive ads to sway-able voters. Elections are about emotions, not facts. Data science and social media can help us make sense of and manipulate the chaos.

An alternative argument. Political misinformation is:

  • Weak in high profile partisan races because pre-existing beliefs hardly change
  • Strong when people don’t have string pre-existing opinions, e.g. misinformation about voter ID laws causing people to stay home from the polls

CA also dug/created dirt using shell companies. Guises: investors offering shady deals, fake IDs and websites, students at local universities, tourists.

Re: . There are regulations that political ads should be labelled as such. What’s the issue here?

There is a significant disagreement between what ad platforms, ordinary people, and advertisers consider political. It seems important to consider social issue ads as political, especially in the call for regulation of political advertising.

\(\Delta\) identifies a problem of definition.


There exist datasets for how legislators vote in Congress. Intellectual puzzles abound, e.g. # dimensions needed to meaningfully predict voting, party loyalty of members, etc.

Re: . What does the web have to do with law-making? I can see petitions, polls and requests for comments, but not much beyond…

learn a model to predict the adoption or rejection of proposed law edits in the EU’s law-making process, which is an instance of a peer-production system. Features investigated: Member of the European Parliament features (e.g. nationality, political group, gender), rapporteur feature, dossier features (e.g. reports vs. opinions), text features (e.g. “any other”, “human rights”, “positive impact”, “resettlement”, “greenhouse gas”, “fisheries”), clusters in latent features (e.g. “environment and communication”, “defense and protection”, “investment and development”, and “business and innovation”).

[Definition] Peer production / mass collaboration is a way of producing goods and services that relies on self-organizing communities of individuals. Non-profit examples: Project Gutenberg, Wikipedia, Linux, Mozilla. For-profit examples with partial peer production: Delicious, Etsy, Goodreads, TripAdvisor, Yelp, etc.

\(\Delta\) offer a more quantitative answer to “this bill will never pass, don’t bother”.

Civic Engagement

“Fixing democracy” is hard. For example, an app that lets citizens vote on issues. First, selection bias because only a fraction of people will participate. Furthermore, people tend to be rationally ignorant, mostly caring about a few issues.

Individuals don’t hold power, groups do. Helping people connect with others and teaching them how to carry out effective advocacy together is hard. It’s not a technology problem. It is a societal and institutional challenge.

Re: . I expect that the peer-pressure aspect of it makes it outperform other forms of voter outreach, e.g. canvassing.

added randomization to help them unobtrusively assess the causal effect of their users' messages on voter turnout. However, there were statistical challenges to assessing the effect. addressed these challenges using additional data and found that friend-to-friend mobilization efforts have statistically significant treatment effects that are among the highest in the get-out-the-vote literature.

\(\Delta\) quantify how much effective friend-to-friend mobilization efforts are.

Re . Don’t know what to expect. What are “opinion dynamics”?

provide a linear-time approximation algorithm (using \(\ell_{2}\) norms of some vectors instead of matrix multiplication and inversion with proved error bounds) for quantifying internal conflict, disagreement, controversy, and polarization. This makes computing such indicators feasible for graphs that have millions of nodes.

\(\Delta\) I didn’t know about quantitative indicators for seemingly qualitative features like internal conflict. For example, cites ’s definition of internal conflict:

For a graph \(\mathcal{G} = (V, E, w)\), its internal conflict \(C_{I}(\mathcal{G})\) is the sum of the differences between internal and expressed opinions over all nodes:

\(C_{I}(\mathcal{G}) = \sum_{i \in V} (z_i - s_i)^{2}\).


Re: . The recommended objects are either other users or topics. Topics make more sense in a political context. Don’t want to venture too far to be undesirable to the user, nor too close to be an echo chamber. Galef addressed a similar problem qualitatively on why blindly listening to the other side tends not to work.


set out to increase the diversity of recommendations in the users and political content graph. They propose an estimator for ideological positions of users and content, and then based on these estimators, propose random-walk based recommendation algorithm. Touted benefits: more ideologically diverse recommendations, no dependence on pre-existing labels (e.g. X is a conservative news outlet), effectiveness in recommending diverse long-tail items.

\(\Delta\) I didn’t know what to expect. seems like a sample implementation of the ideas put forth by Galef.


  1. So You Want to Reform Democracy. Joshua Tauberer. . Nov 22, 2015.
  2. Channel 4 News: Cambridge Analytica Uncovered: Secret filming reveals election tricks. . Mar 19, 2018.
  3. America’s misinformation problem, explained: It’s better - and worse - than you think. Sean Illing. . Nov 6, 2017.
  4. Understanding the Complexity of Detecting Political Ads. Vera Sosnovik; Oana Goga. The Web Conference, 2021, pp. 2002-2013. . .
  5. Interpretational Challenges with Ideal Point Models. John Myles White. . Jan 19, 2019.
  6. War of Words II: Enriched Models of Law-Making Processes. Kristof, Victor; Aswin Suresh; Matthias Grossglauser; Patrick Thiran. The Web Conference, 2021, pp. 2014-2024. . .
  7. Peer production - Wikipedia. . Accessed Jan 6, 2022.
  8. Assessing the Effects of Friend-to-Friend Texting onTurnout in the 2018 US Midterm Elections. Aaron Schein; Keyon Vafa; Dhanya Sridhar; Victor Veitch; Jeffrey Quinn; James Moffet; David M. Blei; Donald P. Green. The Web Conference, 2021, pp. 2025–2036. .
  9. Impactive | All-in-One Digital Organizing Suite. . Accessed Jan 7, 2022.
  10. Fast Evaluation for Relevant Quantities of Opinion Dynamics. Wanyue Xu; Qi Bao; Zhongzhi Zhang. The Web Conference, 2021, pp. 2037–2045. . .
  11. Quantifying and minimizing risk of conflict in social networks. Chen, Xi; Jefrey Lijffijt; Tijl De Bie. ACM SIGKDD International Conference on Knowledge Discovery & Data Mining, Proceedings of the 24th, pp. 1197-1205. 2018. .
  12. Random Walks with Erasure: Diversifying Personalized Recommendations on Social and Information Networks. Bibek Paudel; Abraham Bernstein. The Web Conference, 2021, pp. 2046–2057. .