Unbias.me

Logo

Inspiration

In today's digital landscape, individuals are frequently exposed to news, opinions, and information that may contain biases or lack diverse perspectives. Most AI models either summarize content or classify it as "true" or "false," failing to engage users in critical thinking. The idea for this AI agent was sparked by the growing need to help users sift through the vast, often overwhelming flow of information encountered online

What it does

This project aims to develop an end-to-end AI agent that helps users think critically about any topic by prompting them to explore alternative perspectives, potential biases, and underlying assumptions. Unlike fact-checking models, this AI does not label information as right or wrong but instead raises awareness and encourages independent analysis. The primary goal is to help users critically analyze vast amounts of information, reducing the risk of misinformation influencing global events like elections, wars, or public trust in institutions.

How we built it

We built this product by integrating web scraping, natural language processing (NLP), and large language models (LLMs) to analyze and critique online content effectively. Using Python as the primary language, we leveraged libraries like BeautifulSoup to scrape and clean text data from various websites. To enhance user interaction, we developed a browser extension that allows users to extract and analyze content directly from web pages in real-time. Initially, we worked on fine-tuning custom LLMs using a GPU provided by NVIDIA to detect biases and misinformation in the extracted content. However, due to technical constraints, we transitioned to using OpenAI’s gpt 3.5 for more efficient and scalable model execution.

Challenges we ran into

One of the primary challenges we faced was dealing with GPU dependency issues on the NVIDIA hardware. Resolving dependency mismatches and ensuring compatibility with machine learning libraries like TensorFlow and PyTorch proved time-consuming. To stay on track, we opted to switch to OpenAI’s API, which offered a more reliable and scalable solution for running our LLMs. Another significant challenge was content cleaning during the scraping process, as many websites include distractions such as ads, embedded videos, timestamps, and unrelated metadata. These elements interfered with our model’s ability to generate meaningful critiques. We had to implement advanced content filtering techniques to ensure only relevant text was fed into the LLM. Additionally, integrating the browser extension presented its own technical hurdles, particularly with ensuring seamless communication between the extension and the backend analysis engine while maintaining real-time feedback capabilities.

Accomplishments that we're proud of

One of our biggest accomplishments was developing an AI-powered browser extension that promotes critical thinking and helps users recognize potential biases in the content they consume. We created a functional chatbot using Streamlit that processes web content in real-time and prompts users with thought-provoking questions rather than providing direct answers. Another key achievement was fine-tuning language models to detect bias, missing context, and logical inconsistencies across various domains, including finance, health, and politics.

What we learned

Throughout this project, we deepened our understanding of how language models can be fine-tuned for specific domains and trained to encourage critical thinking rather than just delivering factual information. We also learned how challenging it can be to maintain neutrality in AI systems and how important it is to ensure user privacy while offering personalized experiences. Implementing sentiment analysis taught us how nuanced bias detection can be. Finally, building a real-time content extraction system within a browser environment highlighted the importance of balancing technical performance with a smooth user experience.

What's next for Unbias.me

The next steps for Unbias.me include refining the model’s ability to handle more complex forms of misinformation, including multimodal data such as images, graphs, and videos. We also plan to enhance the browser extension by integrating a memory function that helps track user preferences while ensuring complete privacy. Another major focus will be expanding domain-specific knowledge bases, especially in rapidly evolving areas like health misinformation and political discourse.