Reference
AI Watermark Database
Multi-provider records with claim status, evidence, and inspection support stored separately.
Research
Document methods before headlines
AI Watermark Center publishes independent notes on methods, observations, limitations, sources, and verification dates. Studies are listed as collecting, dataset pending, complete, or retested. No fabricated sample counts.
Provider research
Source synthesis for Claude, OpenAI, Gemini, and Adobe Firefly lives in the database. That work is documentation research, not an experiment.
Provider research
Gemini / Nano Banana cluster
Search-intent pages for watermarks, SynthID, visible marks, and credentials.
Dataset pending
Dataset pending
Nano Banana Image Provenance Test: SynthID, Metadata & Content Credentials
This protocol registers how AI Watermark Center will collect owned Nano Banana image samples by Google product surface. No completed corpus is published. Parser fixtures from the image inspector are not research samples.
Dataset pending
OpenAI Image Provenance Test: C2PA, Metadata & Provider Verification
This protocol registers how AI Watermark Center will collect owned OpenAI image samples by product surface. No completed corpus is published. Parser fixtures from the image inspector are not research samples.
Active studies
Collecting
Does Claude Hide Watermarks in Unicode? Evidence & Testing
This study inspects owned Claude outputs for observable Unicode. It is not a Claude watermark detector. The 24-sample exploratory collection plan is open; no genuine samples have been published.
Completed studies
None completed yet.
Broader programme
Hidden Unicode testing
Document which non-printing characters appear in copied model output, how they survive common editors, and how they can be listed without guessing a model’s identity from their presence.
Metadata comparisons
Compare what metadata is present after typical export paths, and record what is stripped by screenshots, social uploads, and privacy cleaners.
C2PA preservation tests
Record whether Content Credentials survive specific export, convert, and re-save operations, with files, software versions, and dates.
Image export experiments
Log how compression, cropping, and format conversion affect inspectable credentials and metadata. Results will be published only with method notes.
Provider watermark status research
Collect official documentation first, then attempt independent inspection. Statuses stay unpublished until they have a verification level.
How research is judged
- MethodologyOfficial, verified, observed, and unverified — and what each allows.
- Editorial policyCorrections, sources, and human review for provider-specific facts.
- Hidden & invisible Unicode guideEducational companion to the Claude Unicode protocol.
- Hidden Unicode checkerTool used by the text-inspection protocols.
- AI watermark detectorWhat can be checked locally versus provider-side.
- Gemini hubSynthID, Nano Banana, and Content Credentials research cluster.