NewsAgent

Inspiration

After reading papers [1], [2] and [3], we realized that there was a lot of news companies showing many criticial issues around the world only from one side of the story, often ignoring voices that offer a contrasting narrative. This leads to polarization, misinformed opinions, and an incomplete understanding of key events. By limiting perspectives, audiences miss out on the full context and nuance necessary for informed decision-making, ultimately deepening societal divides and eroding trust in the media. Audiences are increasingly seeking a fuller picture, one that goes beyond the siloed narratives presented by traditional outlets.

What it does

We made an AI agentic news platform, which generates balanced short news articles in form of audio broadcast, similar to podcasts or radio news. It does so by aggregating and synthesizing information from a wide range of sources and perspectives. Our platform not only curates content from various reputable viewpoints but also generates insightful articles that capture the nuances behind every story. The NewsAgent can be endlessly traversed, allowing the user to dive deep into any rabbit hole they wish, uncovering the rich diversity of opinions and facts that shape today's news.

How we built it

For the information retrieval we implemented a news aggregator called Feedly, ensuring that the sources there show contrasting opinions on current global issues. Then we implemented Chat GPT o1 language model to aggregate the obtained information from different sources and provide summaries which emphasise the issue of different points of view in the situation and look for contrasting opinions in the sources. For the real voice integration we implemented ElevenLabs API to both translate user voice to text understandable by machine and to make the system actually read the news with a voice. We used React, Next.js and Node.js to develop the front-end and back-end components of our application and implemented Lovable to generate responsive, creative UI for the page. Additionally we used Vercel to deploy our page on the server and Vercel AI SDK to orchestrate the AI tools and agents.

Challenges we ran into

The issue with removing bias from the LLM models. The issue with careful selecting the relevant sources that both are reputable and are evenly distributed in terms of points of view.

Accomplishments that we're proud of

Delivering an efficient system that takes input from user's voice to generate a reliable output showing news about different viewpoints on many topics, whether there are contrary viewpoints in the news sources. We managed to develop an AI agent flow of AI agents working together to: 1. mine relevant information from the source pages. 2. provide summaries of different views included in media articles on some topics, shows summaries and attempts to explore the reason why the stakeholders hold their opinions. Application can be explored by using your own voice to generate news and go in depth about existing generated articles. The AI agent also answers with real speech.

What we learned

More about AI Agents and about the current limitations of LLMs; more in depth understanding of web development.

What's next for NewsAgent

Reliability scores for each generated article, more sophisticated recommendation system for the articles.

[1] = Unveiling the Hidden Agenda: Biases in News Reporting and Consumption. [2] = Newsalyze: Effective Communication of Person-Targeting Biases in News Articles. [3] = The Media Bias Taxonomy: A Systematic Literature Review on the Forms and Automated Detection of Media Bias.