Skip to content

Simple Unmark guide

AI Watermarks vs AI Detectors vs Provenance Metadata

Published
Last updated
Sources checked
Primary references: 6

Short answer

A watermark is a keyed signal deliberately added while text is generated, checkable only with the key. An AI detector has no key and guesses from surface statistics, so it produces false positives and negatives. Provenance metadata such as C2PA is signed information attached to a file, not to prose, and it is easily stripped.

These three are constantly treated as interchangeable, including by tools that sell against them. They differ in who adds the signal, who can check it, and how wrong it can be.

Choose the privacy boundary before every rewrite.

Private is the lowest-cost path and stores no text in our application database. Confidential verifies Google Confidential Space in your browser and encrypts to a key the workload created for that one request, so our web backend receives no text — though DeepInfra still performs the rewrite. Confidential AI, with the model inside a confidential GPU, is coming soon.

Compare the three modes and their limits

Watermarks: added on purpose, checked with a key

A text watermark is inserted during generation by the provider, using a secret key, and verified by recomputing that key over the text. Where a provider deploys one, the check is principled rather than a guess.

The practical catch for everyone else is access. Google publishes its approach and a reference implementation. Anthropic offers a detection API in restricted private preview for eligible organizations. OpenAI's textGrain detector, announced on 5 October 2026, is limited to approved researchers and expert organizations. These access limits mean we cannot run the providers’ tests on your text.

Detectors: no key, only guesses

A general AI-writing detector infers authorship from patterns in the text rather than checking a provider’s watermark key. It can produce false positives and false negatives. A 2023 study of several GPT detectors found that they frequently misclassified non-native English writing; that result concerns the tested detectors and dataset, not every current product.

OpenAI retired its own text classifier in 2023 over accuracy. A detector score is not evidence of authorship, and a watermark result and a detector result are not the same kind of claim at all.

Provenance metadata: signed, and attached to files

C2PA Content Credentials are cryptographically signed statements about how a file was made, travelling with the file. OpenAI documents C2PA plus SynthID watermarking for images and supported audio, with a public verification tool; Anthropic attaches signed C2PA provenance to supported file types.

Two consequences follow. Metadata is strong evidence when present, because it is signed. It is also easy to lose, because copying text out of a file, screenshotting or re-encoding discards it, which is why watermarking exists as a second layer.

Watermark rewriting and detector humanizing

Tools sold as AI humanizers try to change a detector's guess by altering surface style: sentence length, rhythm, common phrases. A watermark is checked with a key against word choices, so the two problems overlap only partly.

A style tweak that moves a detector score can leave much of a watermark in place. A thorough rewrite that changes most word choices can still be flagged by a style-based detector, just as human writing sometimes is. Neither outcome can be promised, and a tool that claims both is claiming two measurements it cannot make.

Which one applies to your situation

Pasted prose has no metadata to check, and may carry a watermark you cannot test for. A downloaded image or audio file may carry both a watermark and signed metadata. A detector applies to anything and proves nothing on its own.

  • Worried about a statistical watermark in text you are editing, including ChatGPT textGrain: rewriting is the documented lever. OpenAI's own test shows synonym replacement already weakens it.
  • Worried about stray invisible characters: scan and remove them, which is exact and free.
  • Facing a detector accusation: a score is not proof, and cleaning text does not make it proof of the opposite.
  • Handling files: provenance metadata and file-level watermarks are outside what a text tool can touch.

Sources

Published by Simple Unmark. Our Methodology explains how we source claims, what we test, and what we do not claim.