Building a Better Social Media Feed: Integrating Democratic Values into AI

Chenyan Jia

By: Chenyan Jia and Michelle Lam

Have you ever wondered what values are encoded in the algorithms that curate your social media feed? Do you ever worry about whether this feed content has the power to sway and influence the opinions of your friends and family?

It’s becoming increasingly clear that the artificial intelligence (AI) systems curating our social media feeds are playing a large role in political polarization. Recent studies point to a worrying trend: social media seems to be ramping up partisan animosity (Milli et al., 2023; Törnberg, 2022). The very platforms we use to connect and share are potentially deepening our negative feelings and behaviors towards those with different political views. Algorithms are not neutral: they have the ability to amplify political posts (Huszár et al., 2021) and can potentially erode our trust in democracy itself (Lorenz-Spreen et al., 2023).

What if we could instead craft artificial intelligence for social media platforms that actively works to reduce partisan animosity? Our research team tackles this question in our latest project, which was conducted on CloudResearch Connect. This work is forthcoming in the Proceedings of the ACM: Human-Computer Interaction and will be presented at the ACM Conference On Computer-Supported Cooperative Work And Social Computing (CSCW 2024).

Clicks, Likes, and Partisan Animosity

The content that users see on social media platforms is typically determined by engagement metrics such as clicks, views, likes, and comments that improve user satisfaction and increase revenue for the social media platform. However, partisan animosity may not be detectable with user engagement signals like these—in fact, algorithms that maximize engagement can sometimes amplify anti-social behavior.

Rather than primarily relying on engagement metrics, we investigated how we could integrate democratic values into social media feed ranking algorithms. Political science research by Voelkel et al. 2023 proposed eight anti-democratic variables, which indicate the extent to which individuals are opposed to core principles of democracy. Based on this research, we used manual methods and then a large language model (LLM) to identify the anti-democratic attitudes reflected in social media posts (see below). This approach allowed us to rank a social media feed based on the extent to which posts promote anti-democratic attitudes.

Definitions and example measurement of eight anti-democratic attitude outcome variables in the political science literature.

From Manual Ranking to AI Ranking: Three Studies to Reduce Partisan Animosity

In our first study, we manually ranked social media posts according to their anti-democratic attitude scores and presented 1,380 Connect participants with altered social media feeds. Participants who viewed social media feeds that either removed anti-democratic content or had anti-democratic posts appear toward the bottom of their feed (downranking) reported reduced partisan animosity (compared to an unaltered social media feed), and importantly, the altered feeds did not influence participants’ reported experience and engagement.

Website Interface of Democratic Attitude Feeds. Participants in different feed conditions were exposed to different interfaces. (Left) Example posts towards the top of the Downranking condition where anti- democratic information is ranked towards the bottom of the feed. (Center) Example posts in the Remove-and- Replace feed in which anti-democratic posts are replaced with pro-democratic posts. (Right) Example posts in the Content Warning feed where anti-democratic posts are blurred with content warnings and users must click the post to see the information.

After demonstrating that altered social media feeds effectively reduce partisan animosity, we wanted to test whether a large language model (LLM) could identify anti-democratic social media content as well as humans. We chose GPT-4 for this task and used a method called zero-shot prompting. This method helped us create a model that understands democratic attitudes using the same 8 anti-democratic items used in Study 1. Our efforts were successful: the model’s evaluations closely matched human judgments. This suggests that such AI could effectively assess social media content on a large scale without requiring human input.

In our final study, we put our model to the test. We compared it head-to-head with the manual approach from Study 1. We recruited 558 participants from Connect and showed them one of three types of social media feeds: one downranked manually by humans, another downranked by GPT-4, and an unaltered feed as a control group. Just like in our first study, participants who browsed through either of the downranked feeds (whether AI-generated or manually curated) showed significantly less partisan animosity than those who saw the unaltered feed. What’s more, the AI-generated feed was just as effective in reducing partisan animosity as the manually curated feed, and it did so without decreasing user engagement.

(Left) Social media tends to orient objective functions around observable variables like user engagement, which may drive partisan animosity. (Center) Our societal objective function method models social scientific constructs as algorithmic objectives. We translate anti-democratic attitude variables to qualitative codebooks for manual ratings and LLM-based algorithmic ratings that produce an anti-democratic attitude model for political social media posts. (Right) We re-rank social media feeds to mitigate anti-democratic attitudes and observe reductions in partisan animosity.

What’s Next for AI and Social Media?

Our research highlights the potential for AI to promote democratic values and reduce political friction on social media. As we edge closer to the 2024 US presidential race, what will undoubtedly be a contentious political period, our method offers a promising option for platforms to address a potentially dramatic rise in partisan animosity.

Going further, can we expand our approach to tackle a wider range of other societal values? We believe our societal objective functions approach should be used to experiment with a wide range of other societal values—from mental health to self-expression to environmental sustainability. As we expand our scope, it will be critical to study the impact of feeds on real-world social media sites with field studies and longitudinal deployments with diverse communities.

Our work embeds the social scientific construct of anti-democratic attitudes into a social media AI objective function. This societal objective function method presents a novel strategy for translating social science theory to algorithmic objectives, which opens up new possibilities to encode societal values in social media AIs.


Huszár, F., Ktena, S. I., O’Brien, C., Belli, L., Schlaikjer, A., & Hardt, M. (2022). Algorithmic amplification of politics on Twitter. Proceedings of the National Academy of Sciences, 119(1), e2025334119.

Jia, C., Lam, M. S., Mai, M. C., Hancock, J., & Bernstein, M. S. (2023). Embedding democratic values into social media AIs via societal objective functions. arXiv preprint arXiv:2307.13912.

Lorenz-Spreen, P., Oswald, L., Lewandowsky, S., & Hertwig, R. (2023). A systematic review of worldwide causal and correlational evidence on digital media and democracy. Nature Human Behaviour, 7(1), 74-101.

Milli, S., Carroll, M., Pandey, S., Wang, Y., & Dragan, A. D. (2023). Twitter’s Algorithm: Amplifying Anger, Animosity, and Affective Polarization. arXiv preprint arXiv:2305.16941.

Törnberg, P. (2022). How digital media drive affective polarization through partisan sorting. Proceedings of the National Academy of Sciences, 119(42), e2207159119.

Voelkel, J. G., Stagnaro M. N., Chu, J., Pink, S. L., Mernyk, J. S., Redekopp C., Ghezae, I., Cashman, M., Adjodah, D., Allen, L., et al. 2023. Megastudy identifying effective interventions to strengthen Americans’ democratic attitudes. (2023).

Related Articles