Chat Is This Real

Inspiration

Over the past two months, the number of AI-generated videos and images has skyrocketed on social media. Many of us felt overwhelmed, questioning whether the breaking news was true or false. Feeling powerless, we took action and conceived the idea of creating a detector to distinguish between AI-generated images and videos and real images.

What it does

Our program takes an image/video as input and performs operations on it to determine whether or not it's AI. Then it will finally return a number based on how confident it is of AI-use involvement.

How we built it

AI image generation models train classifiers to be able to detect them. The model's goal is to become so powerful that a classifier has a 50% success rate in guessing whether an image is AI or real. So, making a regular classifier would not suffice; instead, we opted for first operating a Fast Fourier Transform on the image, converting it to a new image. We then looked for inconsistencies in the new image that were only present in AI-generated images.

Challenges we ran into

The first challenge we ran into was resource-related; we are a small team of teenagers, and image-generation companies are behemoths compared to us. The way they train their models in the first place is by using classifiers. However, we were able to beat this by coming up with clever ways to detect AI use, instead of attempting to brute-force the problem.

Our second challenge was time management problems. This was our team's second hackathon, and we weren't used to the amount of work we had to do in such a short amount of time. However, we were able to conquer this problem with discipline, with some of our teammates pulling two all-nighters.

Accomplishments that we're proud of

Our proudest accomplishment is being able to best AI-image generation companies that have spent billions of dollars and years in the industry. Our model was able to reach an 86% testing accuracy,

What we learned

Hackathons challenge you to make a real, tangible product from your skills. While many on our team excel at math, coding and pitching, making something we're proud of in two days seemed genuinely impossible. However, being able to go through stress, struggle and setbacks, and still make something we're proud of, was truly an amazing experience.

What's next for Chat Is This Real?

While our video-checking software is still above-average and powerful, we really want to improve on it, so we can reduce the number of false positives it receives. Next, we want to add a lot more quality-of-life features to this model, for example, adding an extension that, when used, will detect any AI images/videos on the screen. This will make checking for AI significantly easier and reduce the friction of moving files between websites.