Claude now watermarks its text: what Anthropic’s move changes — and doesn’t — for your SMB’s AI content

On August 11, 2026, Anthropic announced that every piece of text Claude generates — across claude.ai, the API, Claude Code, Cowork and Claude Tag — now carries an invisible, machine-detectable watermark, applied worldwide since August 2. Supported image files (.png, .jpg, .svg) get signed C2PA provenance metadata instead. On August 15, follow-up reporting from Anthropic and the press added technical detail: the mark survives copy-paste and moderate edits, does not affect writing quality or readability, and a regular user cannot see it. The stated reason is direct: Article 50 of the AI Act requires providers of generative AI systems to mark their output in a machine-readable way so it can be detected as artificially generated.

How the mark actually works

Claude’s watermark is a production version of SynthID-Text, the technique Google DeepMind researchers described in a 2024 Nature study, Scalable watermarking for identifying large language model outputs. The method biases the model’s word choices according to a secret key held by the provider: any single word looks like a normal, plausible choice, but across a long enough passage the pattern becomes statistically detectable to whoever holds the key. It runs at inference time, does not require retraining, and — per the Nature paper’s own benchmarks on live traffic — has negligible impact on output quality or latency. That is precisely why it can be rolled out silently to every user rather than as an optional, visible flag.

What this changes for your SMB — and what it doesn’t

This is provider-side, after-the-fact marking: a way to detect that a text passed through Claude, useful to platforms, researchers or regulators running a check. It is not the disclosure Article 50 already requires from your business as a deployer whenever an AI talks directly to a customer — a support agent, a WhatsApp sales assistant, a client portal chatbot must still open the conversation by saying plainly it is an AI, invisible watermark or not. That obligation took effect August 2, 2026 and has nothing to do with this announcement; confusing the two is the mistake to avoid. If your company runs one of these customer-facing agents, the watermark changes nothing about what you owe your customers.

Where it does change the calculation is published content you never intended to label: blog articles, product sheets, regulatory-watch summaries, commercial proposals drafted with Claude, ChatGPT, Gemini or Mistral and put out under your company’s name without a word about how they were produced. Until now, an SMB could reasonably assume that lightly edited AI text was indistinguishable from human writing. That assumption is now false for Claude output, and likely soon for competitors following the same path — Google’s own SynthID already marks Gemini’s text and images. The stakes of that shift are not trivial: a peer-reviewed 2025 study in the Journal of Advertising, Disclaimer! This Content Is AI-Generated, found that disclosing AI involvement in an ad triggers consumers’ persuasion knowledge and measurably lowers trust in both the advertisement and the organization behind it — exactly the reaction a business risks facing on its own terms, proactively, versus having it forced on them later by a platform’s or a journalist’s detector.

Four concrete checks this week

• →List where your business publishes AI-drafted text without disclosure — blog posts, product sheets, regulatory-watch briefs, proposals — and decide your disclosure policy on your own terms rather than a third party’s.
• →For any customer-facing agent (support, WhatsApp sales, client portal), keep the explicit “you are talking to an AI” opening message — that duty stands regardless of any watermark, and predates this announcement.
• →Do not rely on invisibility to skip disclosure on sensitive content — reviews, news-adjacent posts, sales pitches — where trust matters most; detection tooling on major platforms is only going to get better.
• →Check each AI tool your teams actually use (Claude, ChatGPT, Gemini, Mistral): watermarking behavior, scope and persistence are not uniform across providers, so a single company-wide assumption will be wrong for at least one of them.