GenDetective: AI Content Detector

GenDetective is a multi-modal AI content detector that uses statistical and stylometric analysis as well as LLM reasoning to identify AI-generated images, videos, and text.

Comment

Inspiration

The rapid growth of generative AI tools has made it increasingly difficult to distinguish between human-created and AI-generated digital content. Images, videos, and text produced by AI models are now widely shared online, sometimes without disclosure, which raises concerns around misinformation, academic integrity, fake news and digital trust. This project was inspired by the current situation in the world that is whether the image, video or text I am seeing is real or synthetically generated? So to solve this we made GenDetective which is a user friendly multimodal chrome browser extension which detects whether the input (Image, video, text) is real or AI generated.

What it does

GenDetective is a browser extension that helps users determine whether images, videos, or text are likely AI-generated or human-created, directly from their browsing window. Users can upload media or paste text inside the extension popup, and GenDetective performs forensic analysis to return:

  • A clear classification (AI-generated, likely AI, likely real, real, or inconclusive)
  • A confidence score
  • Human-readable explanations describing why a piece of content was flagged

How we built it

GenDetective consists of 2 main components:

Browser Extension (Frontend)

  • Built with HTML, CSS, and JavaScript
  • Provides a clean, modern UI for uploading images, videos, or text
  • Communicates with the backend via REST APIs

FastAPI Backend

  • Handles all computational analysis and machine learning inference
  • Uses separate pipelines for image, video, and text detection

Detection Technique Used:

  • Image analysis: frequency spectrum (FFT), noise patterns, color entropy, edge coherence, metadata inspection with Gemini API reasoning.
  • Video analysis: temporal consistency, optical flow variance, face symmetry, texture smoothness with Gemini API reasoning.
  • Text analysis: TF-IDF character n-grams and stylometric features with logistic regression The backend returns structured, explainable results that are shown inside the extension.

Challenges we ran into

  • Avoiding false positives, especially for edited or compressed media.
  • Handling large images and videos while keeping the extension responsive
  • Deciding when to return “Inconclusive” instead of forcing a confident label
  • Limited API calls

Accomplishments that we're proud of

  • Built a fully functional local browser extension, not just a backend demo
  • Implemented multi-modal detection (image, video, and text) in one system
  • Designed an explainable AI pipeline with forensic justifications

What we learned

  • AI-generated content often differs from human content in minor statistical ways.
  • Browser extensions impose real constraints on performance and UX
  • AI has become too real, no single feature can reliably detect AI

What's next for GenDetective: AI Content Detector

  • Expanding evaluation datasets for better benchmarking.
  • Using advance models for detection.
  • Enhancing text detection using hybrid ML + large language model analysis.
  • Deployment of backend on cloud services and actually launching the extension.

Built With

Share this project:

Updates