/  STEAM in AI   /  research paper   /  STEAM in AI Research: Building a Depolarization GPT AI

STEAM in AI Research: Building a Depolarization GPT AI

A study conducted by Cheng R. Yi in the STEAM in AI: Data + AI Ethics Intensive under the guidance of Mentor Dhruv Diddi, Senior Software Engineer at Turo.

Title: Building a Depolarization GPT AI

Author: Cheng R. Yi

Affiliation: DataEthics4All Foundation, STEAM in AI


This paper aims to introduce an innovative AI program designed to foster depolarization by presenting diverse ideas through various communication styles. The study explores the potential of shifting individuals’ perspectives by gradually nudging them towards alternative viewpoints. Leveraging an AI program, we generate summaries with varying levels of polarity on polarizing subjects, ultimately aiming to reduce polarization and encourage the acceptance of diverse perspectives. This research contributes to the ongoing efforts to combat polarization and promote healthier social discourse.


Polarization is a pressing societal issue exacerbated by the homophily effect, where like-minded individuals tend to congregate, creating echo chambers that reinforce existing opinions. Such polarization negatively impacts public opinions, societies, and democratic processes. However, studies suggest that gentle nudges can lead to depolarization and mitigate echo chamber effects, emphasizing the importance of avoiding radicalization in this process.


Drawing insights from a study conducted by the Indian Institute of Science Education and Research, we incorporate random nudges into echo chambers to reduce group polarization. While it may not fully depolarize the entire group, this approach fosters an environment with globally diverse opinions, conducive to thought-provoking conversations. The sensitivity of nudges is crucial, as overly strong nudges may lead to radicalization, achieving counterproductive results.

Furthermore, we present mathematical strategies to create and maintain groups with socially diverse opinions, highlighting the ephemeral nature of polarizing viewpoints. To achieve these goals, self-belief and moderation of social influence are vital components.


We hypothesize that social depolarization is achievable through the use of AI and Large Language Models (LLMs) to articulate ideas in ways that are more palatable to diverse audiences. We introduce a -5 to 5 scale, where positive values represent pro-topic sentiments, and negative values signify anti-topic viewpoints. This scale offers a nuanced assessment of polarization, capturing crucial information effectively.


In our project, we employed various technologies to attain our objectives. Specifically, we utilized OpenAI’s GPT through the Langchain library. By creating a Chat object and using chat.predict(), we elicited responses from GPT. Employing prompt engineering, we posed questions like, “Imagine the following passage is an article about a topic. Rate its polarization on a scale of -5 to 5, where 5 represents {topic[0]} and -5 represents {topic[1]}. Assume a critical perspective is polar. Respond with only an integer.” Subsequently, we provided an input, typically an excerpt from a polarizing article. We also utilized Typer to establish a user-friendly interface for testing all functions. Additionally, we adapted each function for potential integration with FastApi as a backend for a forthcoming frontend interface.

Assessment and Measures (Results):

To evaluate our results, we conducted simulated user sessions through OpenAI’s Chat-GPT. We measured user polarization by querying the LLM’s opinion on various topics. In the future, we plan to expand our trials to include real users and assess changes in their polarization through pre- and post-survey responses.

Discussion and Future Work:

Looking ahead, we aim to broaden the range of topics covered. Currently, our program focuses on three topics: Earth’s sphericity, climate change, and GMOs, chosen for their high polarization levels and scientifically supported viewpoints. In the future, we aspire to develop a more generalized depolarization program that does not necessitate topic separation. Additionally, we intend to enhance user-friendliness by implementing a frontend chatbot and incorporating features to track conversation history during usage.


1. Pal, R., Kumar, A., & Santhanam, M. S. (2023). Depolarization of opinions on social networks through random nudges. arXiv:2212.06920v2 [physics.soc-ph]

2. Sobkowicz, P. (2023). Social Depolarization and Diversity of Opinions—Unified ABM Framework. Entropy (Basel), 25(4), 568. doi: 10.3390/e25040568.  

3. Wu, Y., Li L., Yu Q., and Zheng Y. (2023). Strategies for reducing polarization in social networks. Chaos.