Study

Better Discourse will conduct a study of online behavior.

The study observes the interaction between biased content and online privacy behavior to better inform the public policy debate on strategies to counter hyper polarization.  The study has potential to inform the development of additional methodologies for detecting ideological bias in AI chatbots.

Background: Most Americans do not trust either the government or companies to handle their data responsibly.  Despite this, today’s digital information systems and hyper-networked markets challenge Americans to effectively manage their privacy.  Without comprehensive legislation, online privacy has become the responsibility of individual users who exhibit a large degree of variation in their online privacy behavior.  Algorithms, fueled by personal data, amplify divisions online and there are myriad harms that AI is positioned to exacerbate. 

The online information environment has become a key medium in which individuals receive information and form opinions and beliefs.  “Data is the new gold” in the online attention economy which incentivizes tailored content to optimize user engagement.  Polarizing and bias-confirming content from more ideologically extreme sources is algorithmically promoted relative to contextualized content from more moderate sources.  Americans are now hyper polarized, as reflected by affective polarization and purification of the political elites who represent each pole.  Observers suggest declining trust in public institutions and social cohesion put the future of liberal democracy in a precarious state.  

AI models draw from online sources to respond to chatbot queries.  By analyzing the sources used to formulate chatbot responses, it is possible to infer the ideological bias of the AI model and whether any measurable bias can be correlated with the bias of individual users.

Theory: Biased content is selectively delivered to individuals, depending on their bias, liberal or conservative.  Just as the results of conventional internet queries are different for each person, this project theorizes that AI chatbot queries use different sources for each person.  If personal data fuels the algorithms that confirm bias and amplify polarization, then reducing the availability of personal data would presumably reduce the accuracy and effectiveness of the process that matches biased content to the “appropriate” individuals.  Persons who exercise greater online privacy behavior and privacy consciousness would tend to exhibit less polarization and their AI chatbot queries would be informed by more moderate sources.

Better Discourse welcomes expressions of interest from academics and interns to collaborate in the conduct of this study. Reach out for more information on the study design.