Baloney

51% of web traffic is now bots. 74% of new web pages contain AI content. People can't tell what's real anymore, and there's no easy and non invasive way to check.

Comment

Inspiration

In today's internet landscape, It's surprising to spend more that three minutes scrolling instagram reels without seeing an anthropormorphized dogs doing kickflips on skateboards, "fruit with a face eating fruit not with a face," and the countless ai comments promoting the same crypto scheme. While some of these forms of generative content bring harmless fun and practical use cases, people today have also begun to use these accessible tools to promote hate and disinformation online. With our generation having grown up on the internet, a place we grew up supposing would always be a window to the outer world, has today become a known "Slop Fest." In an effort to take a leap ahead of the rapid AI race, and experiment with newly published detection methods, we've developed Baloney.

What it does

Baloney is a Chrome extension paired with a web analytics dashboard that detects AI-generated content in real time. As you browse any website, the extension silently scans images and lets you highlight text for analysis using a 6+ signal ML ensemble.

How we built it

When developing Baloney, we began by purely brainstorming from the problem perspective exactly how we'd want to fix the problem. We wanted to introduce a non invasive way to detect ai generated content whenever a user prefers such a catered searching experience. Utlizing the Google Chrome Extension SDK, we discovered a way to gather interesting and revealing data about the immense use of generative content online without invading social media sites scraping restrictions by having users opt in. Our extension sends api calls from the collected media content and uses several research backed generative detection models. The analysis returned is stored in a protected supabase database, where when users return to the baloney.app dashboard, they recieve analyzed metric reports on the authenticity of their content.

The detection pipeline combines commercial APIs (Pangram 99.85% text accuracy, SightEngine 98.3% image accuracy), Google SynthID watermark detection (Gemini text + Imagen images), and statistical/frequency/metadata analysis. Frontend: Next.js 16 on Vercel with 17 API routes. Database: Supabase Postgres (7 tables, 11 views, 4 RPCs). Extension: Chrome MV3 with Grammarly-inspired UX.

Data Methods

  • 6+ independent detection signals per modality with dynamic weight allocation
  • Google SynthID watermark detection — Bayesian detection of invisible Gemini/Imagen watermarks
  • 207-sample evaluation benchmark across 15+ content categories
  • ROC curves (AUC 0.982), confusion matrix, ablation study proving ensemble > individual methods
  • Bootstrap confidence intervals, per-domain accuracy analysis
  • SHA-256 content hashing for crowd-sourced provenance tracking across platforms
  • Real-world dataset: every scan generates a data point (platform, content type, verdict, confidence, timestamp)

Challenges we ran into

We ran into several challenges when using our models trained on determining generative content when faced with real world data sets. Unlike an image downloaded from Gemini's website with an embedded SynthID metatag that automatically flags it, a screenshot or other altered file could yield false results at this stage in detection technology.

Accomplishments that we're proud of

We're really proud of our ability to learn and build something that addresses a real problem in this new, unknown, and unpredictble field. On top of the chrome extensions usefulness in detecting AI, it also comes equipped with features to completely blur and block out elements on a users page. In combining all of these frameworks, we were able to really consider and make choices on the user experience, app design, display and analysis of data.

What we learned

What's next for Baloney

With how proud of the initial concept we are, we hope to fully finish a production grade version of Baloney, and launched on the chrome web store. In collecting this information, we also note the value this data and our analysis have for social media companies, news outlets, educational institutions, and other entities that would want an accurate determination of AI.

Research opportunities: Misinformation velocity measurement, cross-platform content migration tracking, detection arms race benchmarking, information diet behavioral studies, and data protection insights for web safety standards.

AI Tools Used

Claude Code (Opus 4.6 + Sonnet 4.6) for validation, testing, and error elimination. Every architectural and product decision made by humans. Full disclosure: docs/AI_CITATION.md

Links

Built With

Share this project:

Updates

posted an update —

Thank You, MAD Data

First and foremost, thank you to the organizers, mentors, judges, and every participant at MAD Data 2026. Winning 1st Place was an incredible honor, but honestly, the best part of the weekend was being surrounded by people who share the same passion for "creating with code." Meeting the wonderful judges, volunteers, and sponsors gave us a glimpse into what working in the software industry looks like, and we had a blast every hour of it.

Winning the hackathon has inspired us to continue scaling Baloney. In coming up with the original concept, we noticed that the tools to generate AI content had outpaced the tools to detect it by an order of magnitude, and that nobody was building the consumer-facing solution. Building Baloney in 24 hours proved that the concept works. Winning proved that other people think it matters too. And the conversations we had with judges and fellow participants after the demo convinced us this is worth pursuing seriously.

What We've Built Since the Hackathon

We didn't stop when the hackathon ended. In the days since, we've been building Baloney for real users:

  • Chrome Web Store submission We're currently waiting to hear back on whether our extension submission was approved. We've taken extra attention into ensuring we don't violate and privacy or web scraping policies, and hope that our tool is released to the public soon.
  • Open-source community edition frozen at github.com/nategarelik/baloney under MIT license — you can audit every line of what the extension does

You Can Still Use It

The content analyzer is live right now. No extension needed, no signup required:

baloney.app/analyze — paste any text, upload any image, or submit a video to run it through our full detection pipeline. You'll see the verdict, the confidence score, and a breakdown of exactly which detection methods contributed and how much each one weighed in.

We built this to be transparent to the internet's users. We want to provide people with the ability to filter their social feeds, inform them about how to detect AI on their own, and bring back clarity.

A word of advice: AI detection is probabilistic, not definitive. Baloney gives you a confidence score and shows its work. Treat it as a strong signal, not proof. Use it alongside your own judgment, especially for anything high-stakes. We'd rather give you an honest 78% confidence than a dishonest 100%.

Where We're Going

We're incredibly proud of what we put together in 24 hours, and we're passionately scaling this into a professional product.

v1.0 — Public Launch (in progress) Chrome Web Store extension under review. Privacy policy, terms of service, and support pages live. Consent-first onboarding with explicit community sharing opt-in. Host permissions narrowed to 10 explicit platforms. The extension we demoed at the hackathon is being refined for real-world users.

The long-term vision hasn't changed: every user is a sensor. The more people who install Baloney, the more comprehensive and accurate the map of AI content across the internet becomes. That network effect is what turns a detection tool into an intelligence platform.

Stay Up to Date

We're building in public and we want the MAD Data community along for the ride:

Thank you again to everyone at MAD Data 2026. This is just the beginning.

— Nate & Ben

Log in or sign up for Devpost to join the conversation.