Truth is a
technical problem.

Known Watermark Detection

Certain AI publishers embed disclosure markers directly in their images — ALLOD, for example, places a visible watermark in the top-left quadrant. We scan pixel data in that region before any API call, catching these cases instantly at zero cost. A confirmed watermark returns a 99% confidence verdict without further analysis.

99%

Watermark verdict confidence

Satire & Parody Recognition

Known satire publications — The Onion, The Babylon Bee, and dozens more — produce content that is human-written but intentionally fabricated. We maintain a curated domain list and flag these sources immediately, returning a "Real (human-written satire)" verdict rather than misclassifying them as AI.

100%

Known satire domain coverage

Physical Improbability Check

Real photographs cannot depict physically impossible events. A chimpanzee piloting a helicopter, a dinosaur in a modern city, or a human with functional wings cannot be photographed — only generated. We run a focused AI vision pass asking a single question: "Is this physically possible?" Impossible scenes are flagged as AI at 94%+ confidence before full analysis runs.

94%

Improbability detection confidence

C2PA Provenance Manifest

The Coalition for Content Provenance and Authenticity (C2PA) defines a cryptographic standard for embedding tamper-evident creation records in media files. Adobe, Microsoft, OpenAI, and Google all support it. When a C2PA manifest is present and identifies an AI generator, that is definitive proof — no heuristic analysis needed. We check for this before any other signal.

99%

C2PA verdict confidence

EXIF Metadata Forensics

Authentic photographs carry EXIF metadata: camera make and model, lens data, GPS coordinates, shutter speed, ISO, and creation timestamps. AI-generated images typically lack this provenance chain entirely. Missing EXIF on a photorealistic image is a significant AI signal — we extract raw JPEG bytes before any API call and pass the result as a weighted flag to our analysis model.

76%

EXIF signal weight

Canvas Statistical Analysis

Three browser-side heuristics run in parallel before the AI model sees the image: noise floor uniformity (AI images have unnaturally consistent noise across regions), red/blue channel correlation (real sensors produce characteristic color noise patterns), and edge sharpness uniformity (AI generators produce suspiciously uniform edge transitions). A canvas score ≥65 overrides the AI model verdict entirely.

88%

Canvas heuristic accuracy

GPT-4o Vision Analysis

Our primary detection layer sends the image to GPT-4o with a forensic prompt covering background geometry, reflection consistency, photorealistic group shot anomalies, lighting direction coherence, and shadow accuracy. The model returns a structured verdict with per-signal confidence scores. Canvas and EXIF context are embedded in the prompt so the model can factor them into its reasoning.

94%

GPT-4o image detection rate

Ensemble Override Logic

No single signal is infallible. Our final layer reconciles all upstream signals: a canvas score of 45–64 combined with a high-confidence "Real" verdict from GPT-4o is downgraded to "Uncertain." A canvas score ≥65 overrides GPT-4o entirely. C2PA and improbability checks short-circuit the pipeline when definitive. The result is a calibrated confidence score, not a raw model output.

97%

Ensemble detection accuracy