Four things people often call an “AI signature”
Portable metadata is information attached to a file. It may include a generator name, XMP field or C2PA Content Credential. A C2PA credential is signed and can be validated by a compatible tool; a plain text generator field is not the same thing. See Content Credentials explained.
An embedded watermark is part of the image, video, audio or text signal. A vendor-specific detector may recognize it even when ordinary file metadata is gone. SynthID is an example. See invisible AI watermarks.
A platform label is a decision and display made by a website or app. It may be informed by a file signal, a user disclosure or records held by that service. You cannot infer which source it used just by looking at the label. See why an AI label may remain.
A classifier score is an estimate based on patterns in content. It is not a signed record of origin. A classifier can be wrong, and its coverage depends on how it was trained and tested. NIST's overview of synthetic-content transparency discusses these approaches and their limits.
Which question are you trying to answer?
If you want to know what travels with a particular file, inspect its metadata. If you want to validate a signed history, use a C2PA validator. If you want to check a specific vendor watermark, use that vendor's detector where one is available. If you need to make a high-stakes judgment about origin, use multiple independent sources and context; no single negative result establishes that a file was made by a human.
Free AI Wiper serves the first question for supported files and text. Its test method states what it checks and what it leaves outside scope.