Ever been fooled by AI?
AI-generated text, images, and videos are flooding every social platform you use. Most of it is invisible. Some of it is designed to deceive. Baloney gives you the power to see what's real and what isn't — right inside your browser.
The Problem
AI-generated deception is no longer hypothetical. It's happening right now, to real people, with real consequences.
[Example: Mnookin email deepfake incident — a university president targeted by an AI-generated voice clone used to authorize fraudulent wire transfers.]
[Example: Neetcode tweet about AI-generated content dominating feeds — developers noticing the majority of online discussion is now synthetic.]
[Example: AI-generated video deception — deepfake videos used to spread misinformation during elections, financial scams, and social engineering attacks.]
What is Baloney?
Baloney is a multi-modal AI content detection platform shipped as a Chrome extension and a web dashboard. It analyzes text, images, and video in real time — so you always know what you're looking at.
Text Detection
Highlight any text on a page. Our cascading pipeline returns a verdict in under a second.
Image Detection
Images are auto-scanned with a discrete colored dot. Hover to see the confidence; click for full analysis.
Video Detection
Poster frames and captured keyframes are run through the image pipeline for per-frame verdicts.
How It Works
Text Detection — Cascading Pipeline
Models run sequentially. Each stage can exit early with a high-confidence verdict, saving latency and cost.
SynthID
Google's watermark detector. If a watermark is found, we return a high-confidence AI verdict immediately and stop.
Pangram API
Best-in-class commercial text detector. Excellent at catching subtly edited AI text that evades open-source models. If high confidence, use result and stop.
4-Method Ensemble (Fallback)
RoBERTa + ChatGPT Detector + MiniLM Embeddings + Statistical Features. Only runs if stages 1-2 are inconclusive or unavailable. Weighted fusion produces the final verdict.
[More details coming — placeholder for in-depth model descriptions]
Image / Video Detection — Cascading Pipeline
Same early-exit strategy. Video additionally uses multi-frame extraction (poster + keyframes) before routing through this pipeline.
SynthID
Google's watermark detector for images. If a watermark is found, we return a high-confidence AI verdict immediately and stop.
SightEngine Generative AI
Commercial API purpose-built for detecting content from Kling, Sora, Veo, DALL-E, Midjourney, Stable Diffusion, and other frontier models. If high confidence, use result and stop.
Reality Defender (Deepfake)
Only triggered on ambiguous SightEngine results. Specialized deepfake detection for a second opinion on manipulated media.
4-Method Image Ensemble (Fallback)
ViT + SDXL Detector + FFT/DCT + EXIF. HuggingFace fallback if APIs are unavailable.
[More details coming — placeholder for in-depth model descriptions]
Technologies & Models
API Services
SynthID
Google DeepMind
Watermark detector for text and images. First stage in both pipelines — if a watermark is found, we return a high-confidence AI verdict immediately.
Pangram
Pangram API
Best-in-class commercial text detector. Excellent at catching subtly edited AI text that evades open-source models.
SightEngine
SightEngine API
Generative AI image detector purpose-built for content from Kling, Sora, Veo, DALL-E, Midjourney, Stable Diffusion, and other frontier models.
Reality Defender
Reality Defender API
Specialized deepfake detection. Triggered on ambiguous SightEngine results for a second opinion on manipulated media.
Open-Source Models
RoBERTa OpenAI Detector
openai-community/roberta-base-openai-detector
Fine-tuned RoBERTa for GPT-family text detection. Primary open-source text signal.
ChatGPT Detector
Hello-SimpleAI/chatgpt-detector-roberta
Dedicated ChatGPT-output classifier tuned for conversational AI patterns.
MiniLM-L6-v2
sentence-transformers/all-MiniLM-L6-v2
Sentence-level embeddings that capture semantic patterns unique to LLM outputs.
AI Image Detector
umm-maybe/AI-image-detector
ViT-based classifier trained on real vs. AI-generated image pairs.
SDXL Detector
Organika/sdxl-detector
Specialized for Stable Diffusion XL outputs — catches the latest diffusion artifacts.
Open-source models are hosted on HuggingFace. Click any card to view the model page.
Detection Results
[TBD]
Scans Run
[TBD]
Accuracy
3
Modalities
[TBD]
Average Latency
Error Analysis
We prioritize minimizing false positives (Type I errors) — incorrectly labeling human content as AI. A false accusation erodes trust far more than a missed detection.
Confidence Floor
No verdict is issued below a 60% confidence threshold. Content that falls below this floor is marked Inconclusive rather than making a shaky call. Users can trust that when Baloney says “AI-generated,” the system is genuinely confident.
Bayesian Posterior Adjustment
Raw model outputs are adjusted using Bayesian posterior reasoning. In plain terms: we factor in how common AI content actually is on each platform. A 70% model score on a platform where only 5% of content is AI means the real probability is much lower than 70%. This dramatically reduces false positives in low-prevalence environments.
Type I (False Positive): Human content flagged as AI — we minimize this aggressively. | Type II (False Negative): AI content missed — acceptable at the margin; users can always re-scan manually.
Limitations
No detection system is perfect. Here's what we can't reliably detect — and we think honesty about this matters more than marketing claims.
Short text under ~50 words lacks enough signal for reliable classification.
Heavily human-edited AI text blends signals and may read as human-written.
Screenshots of AI-generated text bypass our text pipeline entirely.
Brand-new generative models not present in our training data may evade detection until we retrain.
Safety & Ethics
No PII Collected
We never collect personally identifiable information. User IDs are anonymous session tokens.
Server-Side Processing
All detection runs on our server. No model weights or inference happen on your device.
Open-Source Models
Every model we use is publicly available on HuggingFace. No black boxes.
No Permanent Storage
Content is analyzed in memory and discarded. We store verdicts and metadata, never the content itself.
What's Next
Baloney started as a hackathon project. Here's where we're taking it.
Developer API
A REST API so any app can run AI content detection. Pay-per-scan pricing. Ship detection into your own product.
Enterprise Dashboard
Organization-wide AI content analytics. Track AI exposure across teams, domains, and content types.
More Models
Continuously adding detectors as new generative models emerge. Fine-tuning on the latest GPT, Claude, Gemini, and Sora outputs.
Browser-Native Integration
Working toward deeper browser APIs for seamless, zero-install detection. Manifest V3 sidepanel is just the start.
Get Started in 3 Steps
Install Extension
Add Baloney to Chrome from the extension page. One click, zero config.
Browse Normally
Visit any social media platform. Baloney quietly scans content in the background.
See AI Everywhere
Colored dots on images, underlines on text, and a full analysis sidepanel on click.
AI in Development
Baloney was built with the assistance of AI coding tools. We believe in full transparency about AI usage in our own development process. For a complete disclosure of AI tools used, see our AI Citation document.