Claude's Invisible Watermark Just Hit Its Older Models. Your AI-Written Copy Now Carries a Receipt You Can't Read.

On Friday, Anthropic told its users that starting Tuesday, November 3, the invisible text watermark it baked into its newest Claude models earlier this year would extend to four of its most widely used older ones: Claude Mythos 5, Opus 4.7, Opus 4.6, and Sonnet 4.6. Three of those switched on this week; Opus 4.6 slipped to November 16, where it will be joined by Opus 4.5, Sonnet 4.5, and Haiku 4.5 in a second wave. If AI helps write your website copy, product descriptions, estimates, or customer emails, that text now carries a mark you cannot see, cannot check, and cannot opt out of — installed to satisfy a European transparency law ahead of its December deadline, and applied worldwide anyway. Here is what actually changed, where the receipts have holes, and what a small business should do about it this week. None of it requires a lawyer.
What Actually Changed on November 3
Anthropic announced its text watermark in August, in response to the EU AI Act's transparency requirements. The initial rollout covered only its newest models; last Friday's notice extended it to the older, most-deployed workhorses. The mark, in Anthropic's own description, is "an imperceptible pattern in word choice" — a statistically detectable bias in which synonyms the model picks, built on a variant of Google DeepMind's open-sourced SynthID method. It adds no visible characters, does not change meaning, pricing, or response formats, and, per Anthropic, "contains no information about the user, their organization, or their conversations with Claude." Because it is embedded in the model itself, it follows the model everywhere it is hosted — Amazon Web Services, Google Cloud, Microsoft Foundry — and customers cannot turn it off. Every Claude model released after August 2, 2026 has shipped with the mark on day one; this wave is the retrofit of the older fleet. Users who reach deep into the model picker can still select 2023-era Opus 3, which is not covered — for now.
The Asymmetry: Everyone Can Check Your Copy Except You
Here is the part a business owner should sit with. The watermark is invisible, and there is no user-facing checker: Anthropic's detection is restricted to organizations recognized under EU law — regulators, law enforcement, journalists, researchers — plus companies obliged to verify watermarks under the AI Act, through a detection API that remains in private preview. You cannot paste your own homepage into a box and find out whether it trips the mark. Meanwhile, the weaknesses Anthropic acknowledged at launch still stand: short passages may not carry enough of the pattern to be detectable at all, and users pointed out that even light assistance — proofreading or fact-checking a paragraph you wrote yourself — can stamp the output. In other words, your text may carry a receipt that only certain strangers can read. That is not a crisis. But it is the new shape of the world your content lives in, and it is worth deciding on purpose how you show up in it.
The Receipt Has Holes
A receipt is only as good as its enforcement, and this one has documented gaps. On Opus 4.6 and Sonnet 4.6 — two of Anthropic's most commercially deployed models — the watermark only applies when developers leave the sampling parameters (temperature, top_p, top_k) at their default values; change any one of them and the mark quietly disappears, which is why Anthropic's own notice warns API customers: "If you rely on the watermark for your own compliance, keep these settings at their defaults." The second wave arriving November 16 carries the same restriction. And any motivated actor can strip the mark by paraphrasing text through a different AI system — a loophole critics flagged when the watermark launched that no vendor has closed. The competitive picture is inconsistent, too: OpenAI's textGrain watermark, launched in October, only activates by default for users in the EU and ships disabled for API customers everywhere else. Detection is becoming infrastructure, but it is uneven, dodgeable in practice, and trivially defeated by anyone determined to defeat it. Which tells you exactly who will still be detectable: the people and businesses not trying to hide anything.
Why This Exists: Disclosure Is Going From Badge to Plumbing
When California's AI disclosure laws took effect in August, we walked through what they actually require of a small business — and the direction of travel was already unmistakable: disclosure duties were arriving on different timelines from different lawmakers. The watermark news is the same trend one layer down. Disclosure is moving from a badge someone must remember to pin on the chatbot to plumbing embedded in the model itself. The practical endpoint: "was this written by AI?" is becoming an answerable question — not perfectly, not everywhere, but increasingly on demand by people with standing to ask. Your website copy. Your reviews. Your proposals. Yesterday we covered your content flowing into AI systems without your permission. This is the reverse current: AI text flowing out now carries a stamp. Both currents reward the same thing — content that is true, specific, and yours, whoever or whatever typed the first draft.
What a Temecula or Murrieta Business Should Do This Week
- Adopt the one-sentence policy: AI drafts, a human approves, nothing ships unreviewed. The watermark is not the risk — generic, unreviewed, or wrong content is. A human pass that fixes errors and adds what only you know (the real turnaround time, the real materials, the job you actually did last month) is the difference between assistance and abdication.
- Never let AI invent facts. Prices, availability, credentials, project claims, reviews — AI text that fabricates these is a business risk with or without any watermark. We have already seen AI assistants quote your prices to customers; a wrong number hurts whether a machine or a person typed it.
- Do not build anything on stripping the mark. The only businesses that need to defeat detection are the ones misrepresenting something. If your use of AI is honest, the receipt is harmless — spend the energy making the content worth reading instead.
- Put the question to your vendors. Whoever writes or maintains your site — an agency, a freelancer, a DIY builder — "what is your review process for AI-assisted content?" is now a fair, normal question. A vendor with no answer is telling you something.
- Have your content reviewed for the AI era — free. Book a free consultation and we will go through your site page by page: what reads as generic filler, where the facts are thin, and what to strengthen so your content holds up no matter who — or what — is checking it. Straight answers, no obligation.
The Bottom Line
Watermarks like this one will get broader and harder to dodge, and the businesses they hurt will be the ones coasting on unreviewed output — not the ones using AI as a drafting tool with a human in charge. Your defense is the same as it was before this announcement: content that is true, specific to your trade and your town, and reviewed by someone whose name is on the business. Build that, and every receipt in the world works in your favor.
Sources
- The Decoder — "Anthropic's text watermark is easy to bypass on several older Claude models" (updated November 4, 2026) — primary source for the rollout mechanics: the November 3 first wave (Mythos 5, Opus 4.7, Sonnet 4.6) with Opus 4.6 delayed to November 16 alongside Opus 4.5, Sonnet 4.5, and Haiku 4.5; the sampling-parameter restriction (watermark applies only at default temperature/top_p/top_k) and Anthropic's warning that compliance-reliant customers keep defaults; the watermark as "an imperceptible pattern in word choice" that "contains no information about the user, their organization, or their conversations with Claude"; model-level embedding across AWS, Google Cloud, and Microsoft Foundry with no customer opt-out; the SynthID lineage; OpenAI's textGrain activating by default only in the EU and shipping disabled for API customers; and the EU AI Act's December deadline as the driver.
- Mashable — "Anthropic expands text watermarking to older Claude versions beginning November 3" (November 2026) — primary source for the user notice: Anthropic alerted users on Friday that the watermark expansion would apply to Claude Mythos 5, Opus 4.7, Opus 4.6, and Sonnet 4.6 starting November 3, using the same technology already in Opus 5.5 — an imperceptible word-choice pattern that does not change accuracy or readability and contains no identifying information about the user or organization; Mashable first reported the watermark's launch in early August.
- The Independent — "Anthropic's Claude will now leave a trail that shows whether a text was made with AI" (November 2026) — primary source for the detection asymmetry and criticisms: no opt-out and full global coverage despite the EU origin; detection available only to EU-recognized organizations (regulators, law enforcement, journalists, researchers) and AI Act verification companies, with no way for normal users to check their own text; Anthropic-accepted weakness that shorter text may not carry enough of the pattern; user criticisms that proofreading with Claude can trigger detection and that rephrasing through another AI hides the mark; and escape hatches including non-default sampling settings on Opus 4.6 and Sonnet 4.6 and selecting older models back to 2023's Opus 3, which is not yet covered.