Detection limitations
No detector is a truth machine
Scores describe the configured providers’ response to the submitted material. They are not proof of authenticity, manipulation, authorship, fraud, or intent.
Images and metadata
Metadata may be removed by legitimate platforms, rewritten by editors, copied between files, or absent from screenshots. Its absence is not proof of synthetic generation.
Text
AI-written-text detection is unreliable and must not be used as proof of authorship, academic misconduct, or fraud. Writing style varies with editing, translation, disability, templates, and subject matter.
Video
Video AI detection is probabilistic. Re-encoding, compression, resizing and the processing social-media platforms apply on upload can materially change how confident a result is, so the same clip can score differently once it has been shared.
Content Credentials and other provenance metadata are often removed during transcoding or platform processing. Their absence does not mean a video was made by a person.
Treat a video result as one signal alongside others, not as definitive proof of how the video was made.
Audio
Compression, transcoding, background noise, accessibility tools, and ordinary editing can resemble synthetic signals.
Development simulation
Deterministic fixture providers are visibly labelled. Their scores exist only to exercise the product workflow and must not be interpreted as detection.