Inspiration: As an avid debater, I often find it difficult to find reliable sources that give me the information I need. This reflects a larger problem that our society faces, especially when it comes to today's youth: bias in journalism. These biases do not necessarily come from incorrect information, but from unbalanced arguments that are opinions of the author. However, these "opinions" are often very difficult to differentiate from "facts" especially to a young audience. Enter "The Bias Bulletin".
What it does
"The Bias Bulletin" is a web app where you paste a news link, and it generates (using OpenAI) a “nutrition label” for the article:
- Reliability score: (how evidence-based / verifiable the writing is) with quick reasons
- Bias / framing score: (tone, framing, omission, balance) with quick reasons
- Fact vs Opinion: breakdown
- Left vs Right framing estimate: with % confidence, and it defaults to “center/unclear” when evidence is weak
- A short overall summary plus what to read next so you can balance your perspective instead of learing one perspective of a social issue.
How I built it
Because the timeline was really short, I used ChatGPT 5.2 as a coding assistant while building. The app itself is built with:
- Streamlit for the UI
- Trafilatura to extract readable article text from links
- OpenAI 5.2 to analyze the article using a strict JSON rubric so the outputs stay structured and consistent
ChatGPT helped me move faster by:
- Generating a first draft of the Streamlit layout and Python structure
- Helping debug errors (especially environment setup + JSON parsing)
- Improving the prompt + guardrails so the model gives more consistent, explainable outputs
I still made the final decisions on the feature set and UX, and I edited the code to match the project goal
Challenges I ran into
- First ever project
- Extracting article text was tricky.
- Figuring out how to determine bias (what rubrics I used to calibrate the AI model) was the most challenging part
- Attempting to eliminate AI bias
Accomplishments that I'm proud of
- It's a running website!
- Fact v. Opinion Section works!
- I can extract text from the link!
What I learned
I learned that building something that looks simple is not simple. Reliability and bias aren’t just numbers they’re judgment calls (that even humans routinely struggle to make!), and that means the app has to be clear about uncertainty and show "all steps". I also learned a ton about rapid prototyping: Streamlit is amazing , and structured JSON was a lifesaver.
What's next for The Bias Bulletin
If I keep building this, I’d add:
- More accurate right v. left leaning sources.
- A “compare two articles” mode (same event, different outlets)
- More transparency: a breakdown of bias types (framing, omission, loaded language, sourcing)
- A small database of “balanced” reading recommendations (so it can suggest follow-ups more reliably)
- “media literacy checklist” for teens
Log in or sign up for Devpost to join the conversation.