Methodology, in plain language.
A forensic check is only valuable if its boundaries are well-understood. We separate verifiable cryptographic provenance from heuristic model signals, and display technical limitations alongside every result.
1. C2PA Credentials
We parse cryptographic JUMBF boxes according to the C2PA 2.1 specification, verifying certificate chains and digital signatures from certified signing authorities.
2. Metadata Signatures
We inspect EXIF, XMP, and PNG text chunks for known generator tags from Midjourney, Stable Diffusion, DALL-E, Adobe Firefly, and consumer camera sensors.
3. Invisible Steganography
We scan for 26+ zero-width Unicode characters, bidi overrides, tag characters, and specialized spacing patterns (such as ChatGPT o3-mini U+202F spaces).
Image & Visual Verification
When an image is submitted, the backend runs a sequential inspection pipeline:
- C2PA Manifest Extraction: Reads signed provenance assertions. If valid and signed by a trusted certificate authority (e.g. Adobe, Truepic, OpenAI), this is the strongest signal available.
- EXIF & XMP Parsing: Checks for software markers, generation parameters, and metadata discrepancies.
- Frequency Domain & Neural Analysis: For uncredentialed images, our calibrated computer vision model (based on
OwensLab/commfor-model-384) evaluates pixel-level frequency artifacts characteristic of diffusion pipelines.
Text & Steganographic Analysis
Text analysis combines deterministic string inspection with transformer classification:
- Invisible Unicode Layer: Deterministically locates zero-width spaces (
U+200B), non-joiners (U+200C), joiners (U+200D), word joiners (U+2060), and narrow non-breaking spaces (U+202F). If found, we identify the exact character positions and carrier distribution. - Stylistic Classifier: Evaluates sentence perplexity and burstiness using fine-tuned RoBERTa checkpoints (
fakespot-ai/roberta-base-ai-text-detection-v1). We report probability ranges and uncertainty margins rather than binary verdicts.
Website & URL Quality Scanning
Our website scanner fetches public HTML pages to inspect content structure against observable search engine quality patterns:
- Visible vs. Hidden Content: Compares visible text against hidden DOM elements (off-screen positioning, zero opacity, display none).
- Repetition & Link Density: Analyzes keyword stuffing, unnatural heading repetition, and affiliate link saturation.
- Copy Quality: Passes body paragraphs to our text engine to identify machine-generated text sections.
What “No Signal Detected” Actually Means
A negative result means that our supported tests did not locate verifiable credentials, recognized metadata, or abnormal statistical patterns. It does not prove human authorship.
Watermarks can easily be stripped by taking a screenshot, converting file formats, running compression algorithms, or rephrasing text with human edits.
Scan files using this multi-layered methodology
Experience transparent evidence reporting with full technical attribution.