Every Claude API Response Now Carries an Invisible Watermark — What to Adjust Before You Ship

TL;DR: Every new Claude model embeds an invisible signal in text and C2PA metadata in file outputs. Presence proves Claude processed the content; absence proves nothing at all.

Since August 2, all new Claude models apply two types of provenance marks globally — not just for EU users:

• Text watermark: an imperceptible signal in generated text. Anthropic says it "may persist through some editing." It is not guaranteed to survive heavy rewriting, format conversion, or extraction pipelines. Detection is probabilistic, not definitive.
• C2PA file provenance: .svg, .png, and .jpg outputs receive signed Coalition for Content Provenance and Authenticity metadata. If your pipeline strips EXIF/metadata on pass-through, the provenance chain breaks silently.

What the mark says and doesn't say. A detected mark is evidence that Claude processed this content at some point — it does NOT mean the text is verbatim AI output. A missing mark tells you nothing: the mark may have been lost in editing, format conversion, or the content may predate August 2.

What to adjust if you build on the API:

1. Audit your post-processing steps. If you reformat, merge, or trim Claude output, test whether your pipeline strips the text watermark. Don't build logic that treats "no mark = human-authored."
2. Handle image metadata explicitly. Decide whether to preserve or strip C2PA metadata before distribution. Stripping is fine; inadvertently stripping it while claiming provenance is not.
3. Keep your own audit log. The watermark is corroboration, not a reliable substitute. Log which API calls produced which outputs at the application layer — the mark is backup evidence, not a complete record.
4. There is no opt-out. The API exposes no flag to disable watermarking. Plan around it, not against it.

Coverage extends across Claude API, Claude.ai, Claude Code, Claude Cowork, Claude Tag, and cloud partner deployments (AWS Bedrock, Google Cloud, Microsoft Foundry). Older models — anything released before August 2 — do not carry the mark.

Limit: Anthropic explicitly says detected marks are "not conclusive evidence" that Claude produced the content, and absence "cannot guarantee that AI wasn't involved." Treat the watermark as a transparency signal, not a detection system.

Sources: How Claude's text watermark works — Anthropic newsroom · Watermarks all Claude outputs globally — The Decoder · TechCrunch · EU AI Act context — The Next Web