DeepTxT

• 

  Excerpt from the Bash Reference Manual. Note that because our sample was so small, it's not entirely accurate

• 

  ChatGPT is very obviously flagged as AI - although since it tries to be human, our algorithm isn't entirely perfect

• 

  The full AI scan of the Bash Reference Manual

• 

  Unfortunately, as LLM's are trained on Wikipedia's data, occasionally DeepTxT will flag Wikipedia content as partially AI generated

• 

  Wikipedia is not always flagged, though

• 

  Checking just the body text of this news article

• 

  Checking this entire news article

Inspiration

The Dead Internet Theory suggests that, since around 2016, 90% of the content on the internet is either from bots or AI generated. Since the rise of AI in 2022, AI generated content has become increasingly common in all platforms: from Instagram presenting more and more absurd AI generated content in reels, to google providing the AI summary upon a search, to journalists using AI to write news articles. While some it may be funny like multilingual brainrot, it has led to many people falling prey to fake content, and becoming gradually mistrusting of the internet. Here at DeepTxT Ltd., we aim to help combat a portion of this problem within websites, by using our own algorithms in order to scan, analyze, and tell you exactly how much AI was used to create the text on the website you are currently viewing. DeepTxT will help you take back your own thoughts, allowing you to properly judge the content you consume, and ultimately help cultivate a more human online experience. And rest assured — this paragraph was ENTIRELY generated by a living, breathing, human being.

What we learned

Our original idea was bold and ambitious — trying to analyze every single bit of data on a webpage, whether that be text, images, audio, or videos, and determine what parts of a website were human or AI. Unfortunately, that was TOO ambitious, and we very quickly faced some insurmountable (rather, hard to surmount) obstacles that we felt could completely derail our ability to actually finish this product. As a team of beginners, we had to learn everything from scratch: whether that was frontend, backend, and how to use Github. Every road block we came across, we felt like we were ready to give up, it felt impossible... but we pushed on. We learned maybe the most important skill of them all: perseverance. We pushed through and finished DeepTxT, and while it may not be the product we imagined it would be, its absolutely a product that we are proud of.

How DeepTxT was built

Our frontend was built using TypeScript, React, and Vite. Our backend uses Python and FastAPI. It retrieves the text using BeautifulSoup, a python library, and plugs it into a complex algorithm to decipher whether how likely it was that it was made by AI. Our original model tried to use traditional AI checkers, such as ZeroGPT, but they ended up being too inaccurate for our tastes

Challenges

Our main challenge was scale. We had very big ideas for how we wanted DeepTxT to work; our original idea encompassed ideas such as analyzing the Instagram reels you were looking at in real time, and that just proved unrealistic for many reasons. Instagram, as it turns out, doesn't like when you try to scrape the reels as videos, and so we decided to look into it after this competition, and in fact videos in general don't seem to play well with AI content detection. Photos seemed easier, but still outside of the timeframe we had for this project.

Built With

• beautiful-soup
• fastapi
• python
• react
• typescript
• vite