Technical analysis of C2PA, AI watermarking and verifiable provenance signals.
This library explains which technical traces a file may carry about its origin, how those signals can be checked and where their limits begin. It focuses on reproducible findings instead of probability claims about human or machine authorship.
The long-form articles examine standards, research and provider mechanisms with linked sources. A missing signal never proves human creation; it only means that the documented inspection did not establish a supported provenance signal.
Compare C2PA manifests, invisible watermarks and perceptual fingerprints: what each proves, which transformations each survives and where each fails.
The three approaches answer different questions. C2PA answers 'who signed what, and what happened to it'. A watermark answers 'did this come from provider X'. A fingerprint answers 'have I seen this content before'.
C2PA is rich and verifiable by anyone, but fragile — most email clients, messaging apps and CMS platforms strip it, and a screenshot destroys it entirely.
Learn how AI watermarks are embedded and detected in text, images, audio and video, why they fail, and what a negative result actually means.
An AI watermark is a deliberate, keyed statistical bias injected at generation time — not a visible logo and not metadata. Removing metadata does not remove it.
Every text watermarking scheme has the same three parts: a seed derived from preceding tokens, a sampling rule that biases token choice, and a scoring function that measures the bias afterwards.
A technical review of Google SynthID: tournament sampling, Bayesian detection, open-source components, verification limits and removal attacks.
SynthID is two very different systems under one brand. SynthID-Text is peer-reviewed in Nature, open-sourced, and running in production. SynthID for image, audio and video is proprietary with no published paper, no ROC curves and no independent benchmark.
The modality with the strongest science has the weakest public accountability. Google lets the public query image, audio and video marks through Search, Chrome and Gemini — but text detection has never shipped to anyone outside Google.