Public opinion significantly influences the conduct of society. Traditional survey-based methods for gauging public opinion do have certain drawbacks, though. Huge language models like GPT3, PaLM, ChatGPT, Claude, and Bard have been developed, raising concerns about how AI can comprehend and adopt attitudes based on human language. MIT and Harvard University’s most recent research builds on earlier natural language processing software developments that attempt to condense enormous datasets for better decision-making. They offer a fresh approach to examining media diet models, altered language models that imitate the viewpoints of particular subpopulations based on the media they consume (such as radio, television, or the internet), and how it can be used in forecasting public opinion.

Also Read: Learning the Basics of Deep Learning, ChatGPT, and Bard AI

What is Media Diet?

The types and quantity of media a person regularly consumes are called their “media diet.” It encompasses all media types, including social media, news websites, TV shows, movies, books, and podcasts. Consuming a balanced mix of informative, instructive, and pleasant media while avoiding exposure to harmful content that can harm mental health or well-being is considered part of a healthy media diet. A media diet aims to foster a healthy and long-lasting connection with media that can promote personal development.

Media Diet Models: Demonstrating Predictive Power

The researchers showed that media diet models are successful across various media forms, exhibit predictive power, are resilient to question framing, and include predictive signals even after considering demographics. Additional research revealed that these models are sensitive to how much attention people pay to the news and how their effects change based on the type of question posed.

Developing a Model for Media Diet

The process of creating a media diet model involves three steps. Researchers must first create or utilize a language model to anticipate missing words in a document. They mostly use BERT, a trained model, in their study. Second, altering the language model by using a media diet dataset to train it. This dataset contains articles from multiple media sources that span a specific period. Broadcasting of transcripts and online news on TV and radio by the researchers. This change enables the model to incorporate new data while updating its internal knowledge representations. Third, posing the questions to these models to see whether the answer distributions are representative of populations with various eating habits depending on the media they consume. They analyze responses to survey questions by querying the media diet model.

Public Opinion Forecasting

Academics employ regression models in forecasting public opinion based on polling data. The polling data on COVID-19 and consumer confidence come from statewide polls. Lastly, the researchers use the nearest neighbor approach to track the source media diet datasets from which they obtained projections for a particular survey question.

Importance of Media Diet Research

Three interrelated problems reinforce the value of media diet study. First, when discussing selective exposure, we’re discussing the widespread systemic bias where individuals gravitate toward information supporting their preexisting beliefs. Second, the term “echo chamber” describes an atmosphere that amplifies and reinforces shared views among people with the same viewpoints. Thirdly, content curation and recommendation algorithms surface items based on users’ prior behavior, reinforcing their worldviews. They refer to this as creating a “filter bubble.”

Also Read: This is How Experts Predict the Future of AI

Our Say

This ground-breaking study from MIT and Harvard demonstrates how language models might assist in public opinion forecasting based on media consumption using natural language processing. It promotes improved decision-making across various businesses by illuminating societal challenges and addressing pressing personal difficulties. Additionally, it also helps in the understanding of selective exposure, echo chambers, and filter bubbles. These findings significantly impact government, economic strategies, and public health. It offers a considerable improvement in understanding human sentiments and supports decision-making across several industries by providing a new method of anticipating public opinion.