Reality Check

• 

  Logo

We were inspired by how easy it has become to fool people with AI-generated content. Deepfakes, AI voices, and synthetic videos are now everywhere on social media, often spreading misinformation or manipulating viewers without them realizing it. As everyday users ourselves, we saw how difficult it is to tell what is real and what isn’t. We wanted to create a tool that helps people feel more confident about the media they consume and ultimately make social platforms a safer place to take in information.

Reality Check is a browser extension that analyzes videos directly from a user’s screen and predicts the likelihood that the content was generated by AI. As someone scrolls through platforms like TikTok or YouTube Shorts, the extension captures the video content being watched, processes it through our detection model, and displays a percentage score showing how likely the video is AI-generated. The goal isn’t to replace human judgment, but to provide a real-time signal that helps users question suspicious content before trusting or sharing it.

We started by training a machine learning model using a dataset of both AI-generated and authentic videos. This involved collecting examples, extracting visual and audio features, and experimenting with different model architectures to distinguish subtle artifacts left by generative systems. Once we had a working model, we integrated it into a Chrome extension. The extension captures video frames and audio from the page, sends them through a processing pipeline, and returns a prediction score to the user interface. Our stack combined Python for model training and inference, Node for orchestration, and Chrome extension APIs for real-time video capture and interaction with web pages.

One of our biggest challenges was that we were new to training models from scratch. We had to learn data preprocessing, feature extraction, evaluation metrics, and how to iterate on model performance in a short amount of time. Another major challenge was figuring out how to capture videos from live web pages and feed them into our model efficiently. There were many ups and downs, from model outputs that made no sense to integration issues between the extension and backend, but each obstacle pushed us to understand the full stack of AI deployment better.

We’re proud that we built an end-to-end product, not just a model. Reality Check moves from research to real user impact by running detection in real time while someone is browsing. We’re also proud of tackling a meaningful problem. This tool has potential uses in education, journalism, digital safety, and everyday social media consumption. It empowers users to think critically about the content they see rather than passively accepting it. Most importantly, we proved to ourselves that we can learn unfamiliar technologies quickly and turn an idea into a working system within a short timeframe.

We learned how complex AI deployment is compared to model training alone. Building datasets, handling noisy inputs, optimizing inference speed, and designing a usable interface all matter just as much as model accuracy. We also learned the importance of iteration and how small improvements in data quality and pipeline design often had bigger impacts than changing the model itself. Beyond the technical side, we gained experience collaborating under uncertainty, dividing research tasks, and communicating across different parts of the stack.

We plan to continue improving the accuracy and robustness of our detection model by expanding datasets and incorporating more advanced multimodal techniques. Our vision is to extend Reality Check beyond video into images, audio, and text, creating a unified system that helps users evaluate all forms of media authenticity. We also want to improve performance so detection happens faster and works across more platforms.

Ultimately, we believe tools like Reality Check can help people feel safer and more secure online by giving them transparency into the content they consume while also encouraging a healthier relationship with digital media.

Built With

• databricks
• html
• javascript
• json
• python
• visual-studio