Inspiration
I was scrolling through reels when I saw this video link link. I only realized it was AI because the creator said so at the end. The moment flipped my excitement about the tech into concern about how easy misinformation is becoming. I had always thought I could tell what was real, but this proved otherwise. I showed it to my teammate and he missed it too, which sparked the idea for a helper that rides along while you browse and flags likely AI content. I ran the concept by a friend doing a PhD in AI to explore approaches, pitched it to my partner, and that is how ScrollSafe began.
What it does
ScrollSafe is a Chrome extension that analyzes the short-form content you're watching and detects if it contains AI-generated content. The system provides real-time feedback through color-coded badges that appear alongside videos, giving you instant credibility assessment without interrupting your viewing experience. Currently, the prototype works with YouTube Shorts, but we plan to extend it to Instagram Reels for the full MVP.
How we built it
We started with the core principle that the tool should be non-invasive and not distract from the content itself, but serve as a gentle reminder by your side. This led us to design the badge system that sits unobtrusively in YouTube's interface. We built the frontend Chrome extension first, creating the badge overlay system and video detection logic. Then we developed the backend API with a three-stage detection pipeline: instant cache checks, fast heuristics analysis on video metadata, and deep AI-powered image analysis using specialized detection models from Hugging Face. We implemented advanced image preprocessing with center cropping to optimize image analysis for our specific use case. The system uses a hybrid approach where heuristics provide the baseline detection and AI analysis enhances confidence scoring, with intelligent decision logic that prioritizes heuristics for non-AI content while leveraging AI to strengthen AI detection. We integrated everything with proper error handling and user experience considerations, then deployed to AWS with SSL certificates for production use.
Challenges we ran into
We worked under a tight deadline. We found out about the hackathon a week before the cutoff and had four days to build a usable prototype.
DOM manipulation proved trickier than expected. In a live environment, interface elements do not load deterministically, and sub-second delays affect code that attaches to them. Our badge insertion into YouTube’s action container frequently executed before the targets existed. We addressed this by using the MutationObserver API to monitor DOM changes and reattempt placement until the required nodes were available.
Our initial AI integrations faced significant hurdles. We tried multiple vision models on Hugging Face, including CLIP and other general-purpose classifiers, but ran into API compatibility issues and inconsistent results. The breakthrough came with specialized AI detection models such as haywoodsloan/ai-image-detector-dev-deploy, which better fit the task. We also realized raw model outputs were not enough; accuracy depended on proper preprocessing of images extracted from each video. We implemented downloading those images, center-cropping to 9:16, and JPEG optimization to deliver clean, well-formatted inputs. This preprocessing pipeline proved critical for reliable detection.
Handling fast user scrolling created edge cases we hadn't anticipated. Users can scroll through multiple videos in seconds, so we had to implement smart debouncing and request cancellation to avoid wasting API calls and prevent race conditions.
Implementing proper concurrency was crucial for making our API integration non-blocking. We used ThreadPoolExecutor to handle multiple simultaneous requests without blocking the main application thread.
Accomplishments that we're proud of
Completing a fully functional, presentable prototype in such a short timeframe feels incredible. We built a complete full-stack system with real AI detection capabilities, proper deployment infrastructure, and a polished user experience. We're particularly proud of solving the YouTube Data API integration challenge, successfully integrating a sophisticated AI detection pipeline using specialized vision models from Hugging Face, and developing intelligent hybrid logic that balances AI analysis with heuristics for optimal accuracy. We're especially proud of implementing the "Doom Scroller," a background service running on cloud infrastructure that continuously searches the internet for viral short-form content, performs deep AI analysis on multiple frames using advanced detection models, and caches the results. This will create an extensive database of pre-analyzed content that our live detection system can reference first, dramatically improving response times and accuracy.
What we learned
- DOM in live systems: timing and observation are key to reliable element targeting.
- Concurrency: non-blocking APIs and parallel requests improve responsiveness.
- Data preprocessing: center cropping and image compression greatly boost AI accuracy.
- Model choice: specialized AI-detection models outperform general classifiers for this task.
- Platform integration: addressed YouTube specifics, including rate limits and API reliability.
What's next for ScrollSafe
We're excited to build out the remaining features. We also plan to expand to Instagram Reels and other platforms, making ScrollSafe the go-to tool for content authenticity verification across all major social media platforms.
Log in or sign up for Devpost to join the conversation.