AI4Democracy Series 2: Depolarizing and moderating social media with AI

In recent years, the digital public sphere—primarily hosted on social media platforms—has become a key element of democratic processes. As these platforms increasingly influence political discourse and public opinion, concerns about the risks they pose to democracy have grown. A key concern is the role of AI algorithms in exacerbating political segregation and polarization. The new AI4Democracy paper "Depolarizing and moderating social media with AI: Tools and guidelines leveraging representation spaces" provides actionable insights into addressing this problem.

Understanding the Risks

AI algorithms, which filter and curate the content we see online, can inadvertently reinforce political segregation by promoting selective exposure to information. This lack of diversity in the content users encounter can deepen political divides and reduce trust in democratic institutions. Traditionally, efforts to mitigate this risk have involved prescribing diversity in content recommendations. However, this normative approach raises several questions: How much diversity is enough? Who decides what constitutes appropriate diversity? And how do we ensure that enforcing diversity doesn't itself exacerbate polarization?

A New Approach to Algorithmic Moderation

Pedro Ramaciotti (Sciences Po – CNRS) proposes an innovative solution that bypasses these normative dilemmas. By leveraging advanced AI techniques and spatial models of politics, Ramaciotti suggests that we can design algorithms to be agnostic to political information. This means creating AI systems that do not rely on political data when making recommendations, thereby reducing the risk of political segregation without the need to prescribe specific levels of diversity.

Key Insights and Actions

1. Producing Markers for Hidden Semantics: AI service providers should develop markers to detect political and other sensitive information within AI-generated data representations. These markers help identify and measure political biases in AI systems without violating data protection regulations like GDPR.
2. Self-Assessment of AI Systems: Providers should regularly self-assess their AI systems to detect and address unintended political biases. This involves using predefined markers to evaluate how AI algorithms process political information, ensuring compliance with regulations and promoting transparency.
3. Evaluating Alternative Learning Procedures: Continuous evaluation and adjustment of AI learning procedures can help mitigate the risks of political segregation. By experimenting with different algorithms and embedding techniques, providers can neutralize unwanted political biases while maintaining accuracy.
4. Designing for Openness: AI developers should facilitate collaboration with regulators and researchers by designing systems that allow external scrutiny. This enhances transparency and accountability, aligning with regulatory requirements.

Addressing Challenges

1. Geometrical Complexity of Learned Representations: AI systems often encode information in complex, non-linear geometries. Decoding these structures to mitigate political biases is challenging, requiring advanced methods to interpret these encodings.
2. Processing of Sensitive Data: Analyzing political information can risk violating data protection regulations. Solutions include collaborating with vetted researchers who have exemptions or using non-sensitive proxies to assess political dimensions indirectly.
3. Feature Alignment and Loss of Information: Ensuring AI systems do not process political information may lead to the loss of other relevant data, potentially reducing accuracy. Balancing the exclusion of sensitive data with retaining useful information requires careful design.

By addressing these challenges and implementing the proposed actions, AI service providers can mitigate the risks of political segregation and polarization, contributing positively to democratic processes.

For those interested in reading the report, please access the AI4Democracy page and unlock the download of the paper.