In September we wrote that your AI content is now signed, and that the strategic question had stopped being can anyone tell and become did you say so first. At the time there was a technical caveat sitting underneath that, and most people who noticed it never said it out loud.
The caveat was version coverage. Anthropic's marking commitment is tied to the EU AI Act, and the Act gives older systems a transition period. In Anthropic's own words: "Claude models launched in the EU on or after August 2, 2026 will support machine-readable marking at launch," while "the law includes a transition period for AI systems launched before August 2, 2026, and we're working to add marking support for earlier Claude models."
Read as a planning document rather than a compliance notice, that sentence described a window. Newer models marked their output. Older ones, for a while, did not.
That window is closing. Anthropic has notified customers that the text watermark extends to four more models — Claude Mythos 5, Opus 4.7, Opus 4.6, and Sonnet 4.6 — reportedly beginning November 3, 2026. Three others, Opus 4.5, Sonnet 4.5, and Haiku 4.5, still appear in the help-center table with no start date given.
What this actually does, and what it doesn't
Precision matters here, because the alarming version of this story is the wrong one.
The watermark is applied at generation time. It is woven into text as the model produces it. Extending marking to an older model means that model's future output carries the mark. It does not reach back into a document you generated in June and sign it retroactively. Nothing on your site changes on November 3.
So the archive is not newly exposed. What ends is something else: the live workaround. Reaching for an older model to get unmarked text was a real option, and it is being removed model by model, with two or three names left on the list and no reason to think they stay there.
That is the part worth planning around. The instinct to route around provenance now has a shelf life measured in model deprecations.
The absence of a mark became a dated fact
There is a second-order effect that is easier to miss and harder to manage.
Anthropic publishes which models support marking, and increasingly, from when. Its own guidance lists the reasons content might carry no mark at all, and one of them is that "it was generated by a model before marking was supported for that model."
Turn that around. Unmarked text is no longer ambiguous in a useful way — it is consistent with a specific, documented, narrowing set of circumstances. As the covered-model list grows and the coverage dates recede, "no watermark" stops reading as not AI and starts reading as generated before this date, or on one of these models, or edited heavily enough to degrade the mark.
That is not a detector verdict and nobody should describe it as one. It is worse in a quieter way: it is a timeline. The thing a procurement committee or a journalist can now reason about isn't just whether your content is marked. It is when it could have been produced unmarked, and whether what you have said about your process fits that window.
Why the older-model workaround was never the answer
Suppose the workaround had held. Suppose there were a model that never marked anything, indefinitely.
A content operation built on that is a content operation whose core assumption is that a capability curve reverses and a regulatory regime relaxes. Neither has happened. The EU's transparency code was signed by Anthropic alongside roughly 190 other companies. California's AI Transparency Act has been operative since August. Anthropic applies the mark globally rather than regionally. Every force in the picture points the same way, and the only thing version coverage ever bought was time.
Time is useful if you spend it on the artifact. It is worthless if you spend it hoping the deadline moves.
What we do, unchanged
Pixelworkers uses AI throughout our work — research, drafting, analysis, production. We have said so publicly for as long as we have done it. This news changes nothing about how we operate, because nothing in how we operate depended on which model version marks its output.
A human with domain expertise sets the direction, checks the facts, and puts their name on the result before it goes anywhere. Every factual claim traces to a source. Nothing publishes without a person accountable for it. That is not a compliance posture — it is why the work is any good, and it is also, incidentally, the only thing that holds up under a provenance question.
What to do with the remaining window
Three things, and they are the same three we named in September, with the deadline now visible.
1. Write the standard down this week, not this quarter. One page. Where AI is used, where a human reviews, who signs off, what never ships without a person. In September this was an afternoon's work you could schedule. It is still an afternoon's work. The difference is that the voluntary version of this disclosure expires the moment someone asks you for it.
2. Publish it, dated. A plain page on your site, with the date on it. The date is doing most of the work — it is what converts a disclosure into a record. Disclosure made freely reads as confidence; the identical words produced under questioning read as an admission.
3. Ask your vendors which models they are on. If an agency or contractor delivers content to you, their tooling is now your provenance record, and "we use an older version" is an answer with an expiry date attached. The ones doing this properly will tell you without being pressed.
The part worth sitting with
For three years the implicit deal in AI content was that the seams were invisible. The September news was that the seams had become legible going forward. This news is narrower and, in a way, more final: the escape hatch had a version number, and the version numbers are being closed one at a time.
What that rewards is still not the firms that hid it best. It is the firms whose process would survive the disclosure, and who said so while saying it was still a choice. There is less of that choice left today than there was yesterday.
Sources: Anthropic — How Claude marks AI-generated content (accessed November 3, 2026); reporting on Anthropic's customer notification extending text watermarking to Claude Mythos 5, Opus 4.7, Opus 4.6 and Sonnet 4.6 from November 3, 2026 (BigGo Finance, FourWeekMBA); EU AI Act Article 50(2) Code of Practice on Transparency of AI-Generated Content; California AI Transparency Act (SB 942, as amended by AB 853), operative August 2026.
