Transparent methods · Local analysis
Methodology and limitations
GenDetect is currently an explainable front-end demo. It combines observable signals into a screening index; it does not run a trained and calibrated classifier of content origin. This page documents the implemented scope, score meaning, and evidence that still requires human review.
01 · Implemented signals
Each input type follows a different browser analysis path. This page describes signals the code actually reads and does not present unimplemented model capabilities as product features.
After normalizing whitespace, rules compare vocabulary diversity, repeated three-token sequences, average sentence length, punctuation frequency, unusual-character rate, and sample length. Short text can run but offers less context.
Open the AI text detectorThe browser uses Canvas to sample color buckets, average brightness, luminance variance, and adjacent-pixel edge density, plus dimensions and file size. The current implementation does not read EXIF, content credentials, or camera models.
Open the AI image detectorRules read file size, duration, resolution, average bitrate, extension, and MIME type. The current version does not inspect video frame by frame or identify speech, lip movement, or people.
Open the AI video detector02 · Score meaning
The interface adds the weights of triggered rules and clamps the result between 0 and 100 so signals are easier to compare. It is not a calibrated probability or a statistical confidence interval.
Each unusual signal adds a different weight, such as high repetition, limited color diversity, or an extreme bitrate. Several signals together mean the input deserves more review, not that its origin has been identified.
Short text adds a small rule weight, while very short video or video with an unreadable duration adds more. When duration is missing, current code also derives bitrate as zero and triggers the “very low bitrate” rule. These weights expose demo behavior; they do not establish a higher origin probability.
The current result cannot identify a generation model, author, or editing process. Similar statistical patterns may appear in templated human writing, compressed media, and AI-generated content.
03 · Known limitations
Any detector that sees only the final artifact faces lost information and distribution change. The cases below can create both false positives and false negatives.
Human rewriting, translation, screenshots, filters, crops, platform compression, and repeated transcoding all alter original signals. They may hide generation traces or create patterns that look unusual.
Generation systems keep changing, and people can deliberately adjust rhythm, noise, or encoding parameters. Fixed rules cannot cover generation methods that have not yet been observed.
Code, poetry, tables, brand templates, low-light photos, illustrations, and unusually encoded video may naturally cross thresholds. A high score does not prove AI generation, and a low score does not prove human authorship.
04 · Verification and privacy
Reliable conclusions come from provenance, timing, and editing history, not one score. The current implementation keeps input on the device and commits to disclosing future capability changes.
Preserve complete text, the original file, publication time, author information, and edit history. Screenshots and reposts discard provenance signals that could support verification.
Use content credentials, reverse-image search, trusted-source versions, audiovisual continuity, and contextual facts. Important decisions require human review and a chance for the person being assessed to explain.
Detection currently runs locally in the browser, and the front-end code does not send submitted content to a backend API. If a future backend model or processing service is introduced, this page and the privacy notice must be updated together.