Claude Cowork launched as a research preview on January 12. The product: describe the outcome, hand over the execution. By August 2, Anthropic had begun marking output from newly launched Claude models. Older models are still in a transition period. The new models watermark generated text and add C2PA metadata to supported image files: SVG, PNG, and JPG. For models that support marking, the watermark follows Claude across the API, Claude, Claude Code, Claude Cowork, Claude Tag, and deployments through AWS, Google Cloud, and Microsoft Foundry. There is no off switch. Anthropic files this under transparency and EU compliance. The four largest US hyperscalers guided to roughly $725 billion in combined 2026 capital spending, much of it for AI infrastructure. Each of them now ships an agent whose pitch comes down to one verb: delegate.
Write a paragraph in a second language. Ask Claude to fix the grammar. The returned text can carry the same mark as a page Claude generated from scratch. Anthropic says the watermark means Claude “processed” the text. It does not say Claude wrote it. The company also says proofreading, translation, and summarization can leave a mark even when the underlying ideas and text are yours. I traced the platform version of this when Substack shipped a detector that guessed. Anthropic moved the signal to the source. The mark is stamped there and cannot be switched off on supported models.
A watermark that cannot say what happened
The question is what Claude did. The watermark cannot answer it. Anthropic says a detected mark “is not fully conclusive.” Its absence does not establish human authorship either. People use Claude to proofread, translate, summarize, and convert files, so a detected mark does not establish who originally wrote the work. These are not edge cases.
Full generation and a fixed comma are treated as the same event: metered the same, tagged the same. So can a clean translation and a hallucinated fabrication. Every distinction that matters to authorship collapses into one bit: Claude touched this. The mark cannot say whether the work belongs to a person, a model, or anything in between.
There is no reader yet. Anthropic says detection tooling is “forthcoming.” Until then, the signal sits inside text nobody outside Anthropic can currently decode. You cannot inspect the category, check the assignment, or contest it. Google validated its SynthID text watermark in an October 2024 study spanning nearly twenty million live Gemini responses. Anthropic released the label and promised the reader later. Filed under transparency.
A spell-checker leaves no provenance when it corrects a word. Nobody needs a permanent record that spell-check touched the sentence. Claude’s presence produces the thinnest possible record: a yes that cannot say what it is a yes to.
They spent two years telling you to hand it over
The watermark lands on the exact behavior the industry has spent two years trying to increase. In October 2024, Amodei coined “a country of geniuses in a datacenter.” In March 2025, at a Council on Foreign Relations event, he predicted AI would be writing ninety percent of code within three to six months. By May 2025, he was telling Axios that half of entry-level white-collar jobs could disappear within one to five years.
By September 2025, Benioff said Salesforce had “rebalanced” its support headcount from about nine thousand to about five thousand, crediting AI agents. Salesforce says many of those workers were redeployed rather than laid off. In January 2026, Amodei expanded the country-of-geniuses image to 50 million agents, each with the knowledge of a Nobel Prize winner. Nadella opened Build this year describing agents that act on users’ behalf, reasoning and executing autonomously. Cowork tells you to describe the outcome and let the model execute. The product line wants more of your work moving through the model.
I have argued separately that the country of geniuses does not exist in the systems most companies run. That argument stands on its own.
Delegation is the product, not the pitch
Every handoff increases usage. These companies do not sell a tool once and walk away: API access is metered, subscriptions are tiered, and the bill grows with the amount of work pushed through the system.
Anthropic’s disclosed average cost per developer per active day rose from six dollars to thirteen between February 2025 and April 2026. Anthropic says heavier usage and a shift toward newer models drove the change, not a price increase. That is still the pressure on the meter. More use means more tokens consumed. Across Anthropic’s metered products, more tokens mean more revenue. Amodei has said that being off by a year on growth can bankrupt the company.
On supported models, the watermark attaches to the same contact. Anthropic monetizes the model’s contact with your work, then records that same contact in the output you carry away. The revenue and the label still accrue to different people.
The industry counted lines generated. Tokens burned. More than eighty percent of Anthropic’s merged code authored by Claude. For two years, the public numbers measured volume through the machines. None answered whether the output was good. Now the system records something about the output itself, and the record says only that a machine was there. Quality still gets no mark.
Anthropic chose the broad version. Google did too.
OpenAI had the option. In 2024, the Wall Street Journal reported that it had built a text watermark internally rated about 99.9 percent effective when enough new text was generated. The internal debate over releasing it was already two years old then. By May 2026, OpenAI’s public watermark rollout covered images; its last text-specific update still described text watermarking as a method under consideration. Its April 2023 survey of loyal ChatGPT users found nearly a third would be turned off by the tool. Google already marks text with SynthID. Nothing about shipping a model required this particular design.
Anthropic points to EU compliance. Its implementation goes wider than an EU-only deployment. For supported models, the watermark follows the model across Claude products and surfaces. Anthropic also signed the voluntary EU transparency code alongside Google, Meta, Microsoft, Mistral, OpenAI, and others.
The mark is also removable. Anthropic says very short text, or text that has been heavily edited, paraphrased, translated, or mixed into other writing, may lose the mark. Run an open model on your own hardware and there was never an Anthropic mark to remove.
Paraphrase can beat it. Changing models can avoid it entirely.
The uncomfortable part
There is a transparency case for provenance. A feed drowning in synthetic text needs signals that tell people where material came from. This instrument treats authorship and spell-checking alike. The people it is supposed to inform cannot yet decode it. Anyone motivated enough can try to reword the mark away.
Anthropic keeps the revenue and leaves you the ambiguity. Then it files the arrangement under transparency. That clown show is not funny. It is pathetic.
This article is overstating the issue in a way that it s very misleading. Text watermarking works by changing the source of randomness that is an inherent aspect of token generation. A succession of random choices that match the models own randomness have a higher probability of having been generated by that model.
By definition, then, you need a large number of such choices to have high confidence that AI made them. Anthropic themselves admit that "detecting a watermark.. doesn’t work well on small samples, where there are fewer word choices and thus less information to go on"
In other words, neither a single fixed comma nor a corrected spelling mistake is rich enough to be watermarked.
A translation of a substantial piece of text would be susceptible to watermarking, but that's entirely valid. As any translator will tell you, translation is a form of interpretation. Offloading that interpretation to AI is a significant decision that is worth flagging.
More here:
https://www.anthropic.com/news/claude-text-watermark