Unspun AI

Inspiration

Recently, I was in my AP US Government and Politics class, where we were covering mass media and its effects on American society. After learning about the devastating effects of polarization, misinformation, and declining media trust, I began to see how deep the problem runs. We learned how social media algorithms, which are rapidly replacing traditional news sources, reward clicks over accuracy. Consumers get trapped in echo chambers that reinforce beliefs instead of challenging them. As a result, many young people hear one side of the story repeatedly, and that makes their political beliefs one-sided and eventually leads to straight-ticket voting—supporting a party label rather than policies. Gen Z often encounter news through social media, where misinformation circulates unchecked. This, coupled with the rise of sensationalized infotainment - mixing information with entertainment (focusing on celebrities or scandals instead of policy) to attract views - turn future elections into a dramatized left-versus-right game, weakening democracy. We learned how citizens retreat into ideological bubbles where ‘truth’ depends on the channel. Constant exposure to bias causes young people to disengage from politics, reducing civic participation. And we couldn’t just turn to ChatGPT because current LLMs inherit and amplify bias from their training data.

Fascinated, I researched how severe the issue is. The results I found astonished me: 80% of Americans say the country is divided on core values and principles. (Gallup, 2023) 72% of Gen Z struggle to tell real news from fake news. (Knight Foundation & Harvard, 2022) 64% believe social media spreads more misinformation than it prevents. (Pew Research Center, 2023) Only 26% of Americans trust mass media “a great deal” or “fair amount.” (Gallup, 2022) 6 in 10 U.S. adults say the news leaves them worn out. (Pew Research Center, 2021) 60% of young people claim that one of their biggest sources of news is word of mouth. (poll later taken by me)

My vision is to contribute to a future where news is presented to the public from a bird's-eye view. Unspun aims to bring people the news as it is, unfiltered by the biases of politics, religion, or race. It was developed to be a beacon of truth, not division, as is the frequent case in journalism today. With division at an all-time high, we must unite the country by presenting events without spin. As polarity rises, it’s our duty as citizens to eliminate bias from our news.

How we built it

In an era where misinformation and bias have clouded the truth and polarized American society, Unspun AI makes the dream of easy, impartial news a reality. The algorithm fetches the latest political news articles from several polarized and central news agencies. It then utilizes advanced NLP algorithms and our novel bias quantification package to measure ideological lean in each outlet. It then extracts core facts, balances all viewpoints, neutralizes sensational language, and uses an LLM to present short, neutral briefs in an interactive, scrollable feed. Behind the scenes, Unspun collects 80–90 major articles from outlets across the spectrum—from CNN to Reuters to FOX News, and groups similar articles about the same story using embeddings and linear algebra. From each group, the system derives a factbank - a list of facts agreed upon by a quota of articles and a net bias of |0.5| - plus liberal and conservative claims. We integrated a novel bias quantifier API that analyzes metrics like Fact Coverage, Opposition Coverage, Source Diversity, Loaded Intensity, and sentiment, for a final bias output between ±1. These scores are determined through cosine similarity of embeddings, sentiment analysis, and fact/quote diversity, ensuring transparency and clear logic in the pipeline. Unspun AI is unique in the sense that it not only flags bias and splits claims, but rather that it measures it numerically and provides readers with the full picture. Additionally, the pipeline is fully automated, so humans don’t introduce personal bias (they cannot match the neutrality of math). The rewrite LLM is only given the articles in the group as context, so no outside background info from the web is brought into the rewrites, as is the case with many existing LLMS like Chat GPT or Gemini. This bias-removal layer, combined with generative rewriting, allows Unspun to rewrite the core facts in a way that retains accuracy while minimizing spin, slant, sensationalism, and omission; an issue no mainstream aggregator currently solves. The user sees a clean, scrollable feed of easy-to-read bullet points optimized to engage younger generations like Gen Z and Gen Alpha. We prioritize transparency by providing the user with each original article that was used in the creation of the rewrite and its exact bias metrics, allowing users to verify how bias was measured and mitigated. Unspun AI ultimately transforms the way the next generation consumes information, rebuilding trust in the news and empowering readers with the truth.

Challenges we ran into

When I started building Unspun AI, I had absolutely zero full-stack development experience. I knew how to code in Python for ML, but I had never built an app that connected a backend, frontend, and API layer together. My biggest challenge was figuring out how to turn my random bias-quantifying scripts, scraper, and grouper into a usable consumer pipeline. I relied heavily on ChatGPT to guide me through setting up the backend server, creating API endpoints, and designing the React Native frontend based on my Figma files. Through this experience, I have gained so much confidence in what I can do with the help of modern AI tools (my exact prompt was “Help me set up a professional app structure using a python backend and javascript frontend. Educate me on the packages and frameworks we will be using. Walk me through it step by step as if I am a complete beginner and explain every step along the way”). Beyond that, integrating multiple components that were foreign to me such as the scraper, bias quantifying package, and all the APIs, into one smooth pipeline took a lot of debugging and optimization. The models were heavy and sometimes inconsistent, so I pivoted slightly on the logic, implemented caching and prompt control to improve speed and accuracy after doing some research online. Through these challenges, I learned how to think like a full-stack engineer, plan a scalable architecture, and turn abstract ML python code into a working app.

What we learned

Throughout my time working on Unspun, I grew in 3 main ways: personally, as a citizen, and as a developer. While building, I came across several roadblocks relating to feasibility, cost, etc. but I learned how to pivot the product while working under strict deadlines(v1 of Unspun looks a lot different than what it is today). If you were to ask me just a few months ago, I would have said that there is no way I could have developed a fully functional mobile app with backend ML logic from scratch, but this experience proved me wrong. The Congressional App Challenge challenged me to explore way outside of my comfort zone, forcing me to learn full-stack development, machine learning models, and design a working product from start to finish. I learned how to use ChatGPT as a real development mentor, guiding me through, servers, and front-end, and Git. More importantly, creating Unspun AI taught me that technology can have a real world purpose, and that with enough persistence and curiosity, even one student can create something meaningful for society. To me, Unspun was more than a mere coding project for a hackathon; it was an idea that I grew from a seed to a sprout, and I am yet to grow it to a tree. I didn’t just code Unspun, I built it.

What's next for Unspun AI

For Unspun AI 2.0, I plan to elevate both the intelligence and accessibility of the system. On the technical side, I would replace heuristic bias scoring with a neural network–based quantification model and build a more robust ML-driven extraction pipeline using multi-stage fact verification, cross-encoder and NLI models, and advanced event and entity normalization. This would allow the algorithm to detect not only political bias, but also religious, environmental, and social group biases with far greater precision. Feature-wise, the possibilities are endless! Unspun 2.0 would expand beyond text into a full multimedia experience: a daily, unbiased news podcast generated from Unspun’s top stories, AI-assisted visuals and short video summaries, and an interactive “Chat-with-the-News” feature for readers to explore complex topics. We also envision a browser extension that identifies and rewrites bias in real time, a visual bias heatmap for transparency, and user feedback loops that continuously train our models. In the future, my vision of the app is that it becomes similar to Duolingo, but for news, where users are incentivized to read unbiased news, inadvertently helping them grow in their political opinions and as civilians. Unspun 2.0 aims to make impartial information not only accurate, but engaging and universally accessible.