PoliTalk

\With 85% of youth getting news from social media, PoliTalk's extension detects political bias in real-time so you can think critically about what you watch. Scroll smarter, not blinder.

Comment

Inspiration

We live in an era where social media platforms like TikTok have become the primary news source for millions, especially for Gen Z. It's critical to understand the bias of the content we consume, especially as social media is flooded with uninformed lies and misunderstandings. We noticed short-form content often contains subtle political messaging that viewers may not immediately recognize. We wanted to create a tool that empowers users to critically engage with their content by providing real-time analysis of political bias, helping them understand the perspectives they're being exposed to and make more informed decisions about the media they consume.

What it does

Our Chrome extension analyzes TikTok, Instagram Reels, and YouTube Shorts in real-time and classifies their political bias on a spectrum from Extreme Left to Extreme Right, including Left, Center, and Right positions. The extension:

  • Detects when you're on TikTok, Instagram Reels, or YouTube Shorts
  • Automatically opens, screen records, and processes the video content
  • Transcribes the audio to text using ElevenLabs
  • Feeds the transcription to Google Gemini to detect key terms that indicate political leaning
  • Finds relevant news articles on the topic from historically neutral news channels
  • Uses Roberta-Ideology-Classifier to detect the political bias
  • Displays a visual scale showing where the content falls politically

How we built it

We built a multi-step pipeline that processes short-form videos, extracting audio, analyzing content, and classifying political bias all in real-time. Here's how we made it work:

Chrome Extension:

  • Detects when you're on TikTok, Instagram Reels, or YouTube Shorts
  • Automatically triggers screen recording when videos are detected
  • After the processing pipeline, displays the bias results and more directly in the interface

Processing Pipeline:

  • Screen recording captures the video and audio content automatically
  • We use ElevenLabs API to transcribe audio to text
  • We then send that transcription to Google Gemini API to:
    • Identify politically charged keywords and phrases
    • Find relevant news articles from neutral sources
  • Finally, we use Roberta-Ideology-Classifier to assign the final bias classification
  • The results are then sent back to the Chrome extension for display

Tech Stack:

  • ElevenLabs for audio transcription
  • Google Gemini for keyword analysis and news retrieval
  • Roberta-Ideology-Classifier for bias detection

Challenges we ran into

  • Researching how to access short-form videos through web scraping or screen recording, eventually deciding on screen recording based on versatility and adaptability
  • Struggling with Gemini 2.5 Flash API keys to create a keyword finder using Google AI Search
  • Inputting the Hugging Face RoBERTa Ideology Classifier into our system which was difficult as our first time using a pre-trained model
  • Building our first Google Chrome Web Extension using a Manifest-based architecture
  • Dealing with numerous merge conflicts due to our separate roles and frameworks
  • Integrating ElevenLabs Scribe V2 speech-to-text model later in the coding process caused several programming issues and backtracking
  • Trying to find a clean UI/UX interface which looked professional and informational

Accomplishments that we're proud of

  • Gaining experience in implementing API Keys
  • Learning how to build a new software application type: a Chrome Web Extension
  • Combatting merge conflict through resolution, inspection and backtracking
  • Experimenting with pre-trained models, learning how to implement external models into our programs
  • Being able to expand our initial idea onto other social media sites due to our versatile implementation

What we learned

  • How to build an AI pipeline by integrating transcription, classification, and generative reasoning tools
  • The value of adaptable technical decisions, including pivoting to screen recording for broader platform support
  • How to securely manage API keys and work with multiple external APIs
  • How to build a Chrome Web Extension using a Manifest-based architecture
  • How collaboration and UI/UX design are critical to creating a functional, user-friendly product

What's next for PoliTalk

In the future of PoliTalk, to finetune our model, we would train our own custom model specifically on a broader social media database. This would work more accurately for our variety of social media bases.

We could also make PoliTalk work for more social media platforms which would be simple to input as it has a very versatile system of grasping data.

PoliTalk could also be more informative with more modes and information points, making a more versatile tool.

Built With

  • a-pre-trained-model-trained-on-a-dataset-of-10
  • capturestream-api
  • chrome-extension-api
  • elevenlabs-api
  • express.js
  • gemini2.5-api
  • high-accuracy-transcription-of-social-media-audio-content.-hugging-face-roberta-ideology-classifier
  • javascript
  • node.js
  • react
  • roberta-ideology-classifier
  • tailwind-css
  • typescript
  • used-to-infer-ideological-themes-and-belief-patterns-from-transcribed-audio.-gemini-2.5-api-serves-as-the-core-reasoning-engine
  • vite
Share this project:

Updates