Choose the right check
Published ; updated
Declawd explains watermark experiments and checks a small set of known carriers locally. A provider's production detector needs its own key, compatible configuration and access. Start with the content type and the question you need answered, because a text watermark, a signed file record and a pixel watermark need different checks.
Provider routes
Scroll horizontally to see all columns
| Provider | Content and carrier | Available check | Access and location | Declawd coverage |
|---|---|---|---|---|
| Anthropic | SynthID-Text variant in supported Claude text, separate C2PA in supported files | Public file checker, private-preview text detection | File checker described as local. Text access restricted to eligible organisations | Local PNG/JPEG C2PA inspection, without a compatible Claude text detector |
| SynthID in supported images, audio, video and Gemini app/web text | Public SynthID media verification | Provider service for supported media. Reviewed pages supply no public production text detector | Public SynthID-Text reference traces only, without Google's production keys | |
| OpenAI | C2PA and SynthID images, SynthID audio, textGrain text | Public media checker and Content Provenance API, approved text-detector access | Media checks submit content to OpenAI. Text access requires approval | Local C2PA inspection, without a compatible textGrain or SynthID media detector |
| Meta | Muse for production media, TextSeal2 and other published research | ContentSeal indexes the production description and separate research implementations | Root index has MIT terms. Each linked implementation and model needs its own licence and access review | Research source review, without a Muse detector |
| Adobe TrustMark | A pixel watermark with variant-specific models and payload encoding | Public matching Rust and Python implementations | Can run locally after model downloads. Code and model terms are recorded separately | Registered generated-image experiments with matching Q models |
OpenAI describes textGrain in EU ChatGPT and worldwide API opt-in. Anthropic's current table lists fourteen text-marked models. Sonnet 4.6 may lack a reliable mark when temperature, top_p or top_k has a non-default value, including a value inserted by a client library. A model name alone does not establish that a particular historical output was marked.
The reviewed Google overview names Gemini app and web text. A developer-forum participant corrected an API marking answer, but the current official pages reviewed here do not establish the API rollout history. The source review retains that unresolved distinction.
Different results
A C2PA parser can find an embedded record. Validation then asks whether the record binds to the selected file, whether the claim signature verifies and whether the signing credential matches the chosen trust policy. A provider-specific watermark check answers another question. Neither result establishes authorship, ownership or the truth of the content.
A matching TrustMark decoder can recover a payload from pixels after metadata removal. That result applies to the tested model, variant, encoding and images. A generic positive flag is weaker than exact payload recovery, and a remover can change pixels while leaving a watermark detectable.
Research terms
Meta's ContentSeal index describes Muse as proprietary production technology and links separate research systems. An MIT licence on the index does not license every linked model or expose the production system. OmniSeal Bench and Watermark Anything are archived, so keep them as historical references rather than maintenance candidates.
Public implementations let researchers select their own keys and run controlled tests. A repository name, a provider API key used for text generation or an operator-configured detector URL supplies no evidence of access to that provider's watermark detector. The repository review records selected claims and its capture gaps.
Local tools
Use Files for local C2PA checks, Clean to inspect listed Unicode structures and apply selected edits, and the SynthID lab for the fixed public token-trace reference. The method explains the prepared text experiment and its measured false positives and misses. These tools supply no general AI-writing verdict.