Watermarking and the AI API: what developers need to know

If your product calls the Claude API, the text it returns to your users carries Anthropic's watermark[1] — the API is explicitly in scope, alongside the Claude platform, Claude Code, Claude Cowork and Claude Tag[1]. There is no documented opt-out[1]. For most products this changes nothing, but there are three cases where it does.

Does this affect my app?

Usually not

The mark does not change the meaning, quality or readability of the output[1] and does not require anything from you. If you are building a chatbot, a summariser, or anything where users know they are talking to an AI, you can stop reading here.

Three cases where it does

1. You resell output as human-written. Ghostwriting tools, content mills, "undetectable" copy generators. The mark travels with copy and paste[1], so your product's core promise now has a technical expiry date attached to it.
2. Your users are graded or audited. Education, academic writing, legal drafting. Your users can be marked for text they wrote themselves and merely ran through your tool, because a mark reflects processing rather than authorship[1]. Tell them that up front — it is a support burden if you do not.
3. You publish output at scale. If you generate public-facing text, assume it is identifiable as machine-processed and disclose accordingly. Article 50 puts obligations on deployers too, not only on model providers[7].

Things you might be assuming that are not established

• That you can test for it. Probably not. Anthropic's watermark detection API went into private preview on 1 September 2026[10], but it is scoped to organisations that EU law obliges to check, plus enterprises under the same compliance obligation, by application through a form[1]. Building a product feature on it is not an option unless you are in that set; Anthropic says access will widen over time[12]. Until then you cannot verify whether a given response carries a mark, or whether your post-processing removed it.
• That the file checker covers your text. It does not. Anthropic's free Claude Content Checker reads C2PA credentials on image, video and audio files and takes no text input at all[1][11]. If you generate files with Claude it is genuinely useful; if you generate prose it answers nothing.
• That short outputs are marked. Very short outputs are named by Anthropic as a case where content may be unmarked[1], and detection works poorly on small samples, growing more confident as passages get longer[10]. If your product emits one-line responses, behaviour is likely inconsistent.
• That generated code is marked the same way prose is. It is not. Where an exact output is required the watermark is not applied, so code carries generally less of it, and what there is sits in comments and arbitrary naming choices[10]. The same holds for factual passages, where the correct word is the only word.
• That the mark says something about your users. It does not. Anthropic states the watermark carries no identifying information and cannot be traced to a person, organisation or chat, and that nothing in the mark or its key would let anyone recover that[10]. It also says the watermark changes nothing about output ownership or a user's rights under Anthropic's terms[10].
• That going through a cloud platform opts you out. It does not. Anthropic says embedded watermarks apply when supported models are accessed through AWS, Google Cloud or Microsoft Foundry, and that watermarking happens at the model level regardless of the product or surface[1]. Files follow a stated rule rather than a vague one: signed provenance metadata is added when Claude creates a file, so it applies only where the platform offers Claude's file generation features — which Anthropic names as the Claude apps and the Claude Platform API, including Claude Platform on AWS and Claude in Microsoft Foundry[1]. On the API, its developer documentation says the credential is attached to supported image, video and audio files produced in the code execution sandbox when you download them through the Files API, with no change to your requests or response handling; text files, PDFs and office documents are not signed[13].

What to do

Add a line to your privacy policy or terms noting that AI-generated output may carry provenance marks. If your users could be assessed on the output, say so where they will see it, not in a footer. That is the whole of the work for most teams.

Last verified against primary sources: 2026-09-18