AI Watermark & Provenance Status

Mechanism

Statistical text watermark + C2PA file provenance

Public verification

Announced, not yet released — Anthropic says detection by users and third parties is coming, with technical documentation to follow.

Checkable on this site

C2PA Content Credentials in files Claude generates (SVG, PNG, JPG) can be verified here in your browser. The statistical text watermark cannot be read by any third-party tool yet.

• Applies to Claude models launched on or after August 2, 2026; earlier models are being brought up to the same behavior.
• Generated text carries an imperceptible embedded watermark that does not change the meaning.
• Files Claude generates in supported formats carry signed C2PA provenance metadata.
• Anthropic notes a detected mark means content may have been processed by Claude — not that Claude authored every idea in it, and edits or translation can make marks undetectable.

Sources: Anthropic Help Center — How Claude marks AI-generated content

Mechanism

Invisible statistical watermark (per-modality) + C2PA adoption

Public verification

Ask the Gemini app to check images, video or audio ("Is this made with AI?"); rolling out to Search and Chrome. The dedicated SynthID Detector portal is limited to an early-tester waitlist.

Checkable on this site

SynthID itself cannot be detected locally by anyone but Google. What we can verify here is C2PA metadata, which Google is also adopting across products.

• Google reports over 100 billion images/videos and tens of thousands of years of audio watermarked.
• Text watermarking covers Gemini app and web output ("partial": not all Google text surfaces are documented).
• Partners including OpenAI, Kakao and ElevenLabs embed SynthID in their own generators.
• An enterprise "AI Content Detection API" exists on Google Cloud; there is no self-serve public API.

Sources: Google DeepMind — SynthID · Google — Identifying AI-generated media (May 2026)

Mechanism

C2PA metadata + SynthID watermark (images, supported audio)

Public verification

OpenAI's Verify tool (public preview) checks uploaded images for C2PA metadata and the SynthID watermark. There is no tool that detects ChatGPT text.

Checkable on this site

C2PA Content Credentials in ChatGPT/API images can be verified here in your browser. The SynthID layer cannot.

• OpenAI joined C2PA and, since May 2026, every image generated through ChatGPT, Codex and the API carries C2PA metadata plus a SynthID watermark.
• Supported audio output also carries SynthID.
• ChatGPT text output carries no deployed watermark — OpenAI researched text watermarking but has not shipped it. Metadata can also be stripped from images by platforms or edits.

Sources: OpenAI — Advancing content provenance (May 2026)

Mechanism

Open standard: signed provenance manifests embedded in files

Public verification

Fully public: any conforming tool can validate Content Credentials. This site does it locally in your browser; Adobe hosts an inspector at verify.contentauthenticity.org.

Checkable on this site

Full local verification — manifest parsing, signature validation and provenance chain, powered by the official CAI browser SDK.

• Not a hidden watermark: C2PA is signed metadata travelling with the file. Anyone can verify it; it can also be stripped.
• Adopters include Anthropic, OpenAI, Google, Adobe, camera makers (and, since 2026, phones such as Pixel 10 shipping Content Credentials from the native camera).

Sources: C2PA specification · Content Authenticity Initiative — c2pa-js

Mechanism

C2PA Content Credentials

Public verification

Content Credentials attached by Firefly and Creative Cloud apps are publicly verifiable with any C2PA tool, including this one.

Checkable on this site

Full local C2PA verification, same as any Content Credentials file.

• Adobe co-founded the CAI and attaches Content Credentials across Firefly outputs and Creative Cloud workflows.

Sources: Adobe Content Authenticity Initiative