verifAi

Inspiration

While observing our peers during the opening ceremony, we noticed most were using chatGPT. This sparked a discussion about the reliability of information in both their immediate research and broader educational contexts. Generative AI is increasingly integrated into educational tools, offering personalized learning experiences that adapt to individual styles and paces. It's used to create engaging educational content, including practice questions and interactive materials. However, ensuring the accuracy and reliability of AI-generated information is essential, particularly in educational settings where misinformation can spread easily. By responsibly leveraging generative AI, we can enhance learning experiences and ensure students are exposed to accurate and reliable information to address the UN’s sustainable development goal 4: quality education.

What it does

verifAi is a web extension where a user copies and pastes AI-generated text to have its accuracy confirmed or denied. It does this by parsing the text sentence by sentence, to ensure validity on a smaller scale. However, it also extracts the main facts from the text to ensure that overall concepts are correct and returns these facts and their sources to the user in an integrated environment.

How we built it

verifAI is a web extension built using NextJS, CSS, and HTML, using hooks, such as useEffect and useState, and will be compiled into a Chrome extension.

The parsing and calculations were implemented using javascript and the fact-checking was implemented using the Google Gemini API

Challenges we ran into

None of our team members had significant previous experience with AI, so our main obstacles stemmed from unfamiliarity and lack of experience. For example, we encountered regional limitations with Google Gemini API, the Language Learning Model we intended to use, as we were unable to obtain an API key in Canada. Thankfully, we were able to solve these issues with creative problem-solving, research, and support from mentors.

Accomplishments that we're proud of

The team is very pleased with our working prototype! We’re extremely proud of our perseverance and dedication throughout these 24 hours, and we accomplished way more than we ever expected Let us go through the timeline of this hackathon in 3-hour increments.

Hours 1-3: We brainstormed and were able to settle on an idea. Talking to a mentor clarified our direction.
Hours 3-6: Tiger got a local host chrome extension working!
Hours 6-9: Honestly a bit of a slump... But we got bubble tea!
Hours 9-12: This is when things started getting serious. We sought more help from a mentor and progress started rolling again
Hours 12-15: Our lovely friends let us stay over at their place and the slide deck was started.
Hours 15-18: Tiger (the responsible one) goes to bed. Maggie and Steph start to go crazy. 
Hours 18 - 21: zzz
Hours 21-24: Final push! The slide deck is complete and the devpost is almost done too. Final touches on the extension

What we learned

No one on our team had substantial previous knowledge with natural language processing models, so we learned a lot of new terminology and concepts. For example, initially, we intended to verify information by scraping relevant search results from Google and calculating their semantic similarities. However, a mentor (thanks Sina!!) directed us to research Retrieval Augmented Generation (RAG). RAG is a natural language processing technique where the model first retrieves information from a database, and then generates text based on the retrieved content. Using the Gemini API, we were able to incorporate previously existing RAG resources into our project.

Additional things we learned: Some of our group members had no experience with APIs, git, and nodejs. Throughout this hackathon, we really pushed ourselves outside of our comfort zones by experimenting with new platforms and technologies

What's next for verifAi

There were several features that we brainstormed, however we were unable to implement them during this hackathon due to time constraints. The first was sorting sources by date in order to eliminate outdated sources. Secondly, we noticed Chat GPT and Gemini both avoid controversial topics. This poses quite a problem as these seem to be the topics lacking nuanced discussion in modern times. We plan to address this by parsing sentences by who, what, where, when, why, how, and verifying individual facts; this separates objective facts from subjective opinions. Moreover, verifAi can be scaled to confirm more than just texts from generative ai, but human generated text as well. This includes sites like twitter, quora, and reddit, for which user entries can be checked for validity.