{"ok":true,"records":[{"providerId":"anthropic","providerName":"Anthropic / Claude","aliases":["anthropic","claude","claude-sonnet","claude-opus","claude-haiku"],"textWatermark":{"status":"planned_not_public","method":"SynthID-Text-style statistical watermark","publicDetection":"planned","summary":"Anthropic has described Claude text watermarking and says a detection API is planned, but no public Claude verification adapter is connected yet.","limitations":["Claude watermark detection requires Anthropic verification access or compatible provider integration.","A future Claude watermark signal can indicate likely Claude involvement, not authorship proof.","Short, factual, code-heavy, lightly proofread, or heavily edited text can carry weak or missing watermark signal.","Rewrite and watermark-removal tools can weaken or remove statistical text signals, so missing signal must remain an unknown state."],"sourceUrls":["https://www.anthropic.com/news/claude-text-watermark","https://devby.io/news/anthropic-obyasnila-kak-budet-markirovat-teksty-claude"]},"fileProvenance":{"standard":"c2pa","status":"supported_when_present","summary":"Anthropic says Claude-generated files of supported types can attach C2PA content credentials, which should be reported separately from text watermarking."},"tamperResistance":{"level":"weak_against_rewrite","summary":"Claude text watermarking is described as a statistical wording signal. Substantial rewriting can weaken it; C2PA file credentials are a separate metadata/provenance layer that can be stripped or lost.","evidenceUrls":["https://youtube.com/shorts/htCLSpvxbhU","https://github.com/guillaumemeyer/watermarks-remover"]},"updatedAt":"2026-08-17"},{"providerId":"google","providerName":"Google / Gemini","aliases":["google","gemini","synthid"],"textWatermark":{"status":"research_only","method":"SynthID Text","publicDetection":"available","summary":"SynthID Text is documented for compatible generation pipelines, but it is not a universal detector for arbitrary pasted text.","limitations":["Detection only applies to compatible watermarked generation pipelines.","Results are probabilistic and can be uncertain.","A missing SynthID signal is not evidence that text is human-written.","Paraphrasing, regeneration, or metadata stripping can reduce detectable watermark/provenance signals."],"sourceUrls":["https://huggingface.co/blog/synthid-text","https://ai.google.dev/responsible/docs/safeguards/synthid"]},"tamperResistance":{"level":"weak_against_rewrite","summary":"SynthID-class text signals are statistical and should be treated as probabilistic. Media and file provenance can involve separate metadata or pixel-domain marks with different removal risks.","evidenceUrls":["https://youtube.com/shorts/htCLSpvxbhU","https://github.com/guillaumemeyer/watermarks-remover"]},"updatedAt":"2026-08-12"}],"updatedAt":"2026-08-21T14:04:34.168Z"}