AI Watermarks Go Mainstream as EU Labeling Rules Reach Enterprise Content

Anthropic said this week that it will watermark text generated by its models, TechCrunch reported on August 11, and on August 14 it published a technical explainer on how the mark works. The same day, Google gave Gemini users a switch to remove the visible logo from AI images, while keeping its invisible SynthID mark in place. Both moves trace back to the same date: August 2, when the EU AI Act's rules on marking AI-generated content began to apply. For enterprise leaders, provenance is becoming part of how documents and customer messages are produced, with limits buyers should understand.

What changed for Claude, Gemini and other AI tools this week?

According to TechCrunch, Anthropic said models released after August 2 will automatically watermark the text and files they generate, and it plans to extend the practice to older models. Because the mark is applied at the model level, it covers the API, the Claude apps and Claude Code, and it travels with text that is copied and pasted.

Anthropic's explainer added that it is applying the watermark worldwide, not just in Europe, because it does not yet have a durable way to limit it by region. Customers outside the EU therefore get the mark whether or not local rules require it. Files such as .png, .jpg and .svg instead get a signed C2PA content credential in their metadata saying Claude was involved.

Google went the other way on visible labels. PCWorld reported that a new Media watermark setting in Gemini lets users turn off the visible logo on images, videos and music, except where visible AI labels are legally required. Google's Josh Woodward described it as "striking a balance here between creative control and safety," and said the invisible SynthID mark and metadata stay embedded.

Anthropic said it signed the EU's new transparency code in July, and TechCrunch noted that Google, Meta, Microsoft, OpenAI and others have also committed to it.

How does AI text watermarking work, and what are its limits?

Anthropic said its watermark is a version of SynthID-Text, described by Google DeepMind in a 2024 Nature paper. Where a model would pick at random among equally good words, the watermark substitutes choices derived from a secret key, so whoever holds the key can test whether a passage statistically matches. No hidden characters are added, and Anthropic said the mark carries nothing identifying users, organizations or chats.

As Unite.ai summarized, Anthropic said detection works poorly on short passages, on factual text where accuracy leaves few word choices, and on code, which must be exact. A light edit probably will not remove the mark, but a full rewrite will. Detection only estimates the likelihood that Claude was involved; it cannot prove human authorship, identify other AI systems, or separate text Claude wrote from text it heavily edited. As of this week, a detection API was planned but its details were still being worked out.

How much should enterprises trust Anthropic's claims?

Independent scrutiny is thin. Unite.ai observed that Anthropic's quality claims rest mainly on its own internal testing and on Google DeepMind's earlier study of the technique, with no outside evaluation cited. Because detection depends on a key Anthropic holds, and no detection service was yet available, outsiders could not yet test the mark's reliability themselves.

Fortune reported that determined users can weaken statistical text watermarks and that heavy rewriting or translation may strip this one, and its reporter concluded that "watermarking alone isn't going to clean up the internet." Axios pointed out that human copy merely polished or translated with Claude could still test as AI-generated, and that OpenAI's marking has focused on images and audio rather than text, leaving coverage uneven. Google deserves the same caution: optional visible labels shift more of the burden onto invisible SynthID signals readers cannot see.

What does the EU AI Act require on AI-generated content?

Marking duties fall on AI providers. The European Commission said on July 31 that its marking obligations take effect August 2, and that about 190 organizations had signed the voluntary Code of Practice on Transparency of AI-Generated Content, which offers a recognized way to show compliance. About half the signatories are small or recently founded companies, the Commission said, and the code has separate sections for providers and deployers.

Deployers have their own duties under the Act's Article 50. A June analysis by ComplexDiscovery laid them out: disclose deepfakes, and disclose AI-generated text published to inform the public on matters of public interest unless it has gone through human review and editorial responsibility. The Act also requires that people be told when they are interacting with an AI system such as a chatbot. Penalties can reach €15 million or 3% of worldwide turnover. The analysis also said, citing Commission guidance, that systems on the market before August 2 may have until December 2, and that signatories must support interoperable detection by February 2, 2027.

Why should enterprises care if they are not AI providers?

First, contracts, marketing copy and support replies drafted with Claude may carry a statistical mark, so teams that publish in the EU should know which assets came from which tools.

Second, detection is probabilistic and should not be used as a lie detector, especially since it may flag text a person wrote and only edited with AI. HR and legal teams should not treat a score as proof about an employee or applicant.

Third, provenance becomes a records question. ComplexDiscovery said vendors should expect customers to ask whether AI-generated work carries machine-readable provenance markings that survive export and processing, and it advised discovery teams to preserve that metadata rather than strip it.

Girish Bhat, SVP at Revefi, told a AI Connect webinar on moving from human in the loop to agent ops that data, security and AI governance will converge because "fundamentally it's all about provenance." At an AI Demo Day webinar on August 5, SAS Principal Data Scientist Robert Blanchard, who demonstrated forgery detection, argued that "trust can no longer be passive."

What should AI and data leaders do now?

Start with an inventory of generative AI tools whose output reaches customers, regulators or the public, and whether your organization acts as a provider or deployer for each. That also surfaces unsanctioned tools, the first step in getting shadow AI under control. Then add marking, metadata preservation and independent detection testing to vendor questionnaires, review publishing workflows, and set a policy on how detection results may be used.

Watermarks will not settle debates about AI authorship, and vendors' reliability claims still need independent testing. But with EU rules in force and major providers marking output by default, provenance is becoming a standard property of enterprise content, and teams that plan for it now will spend less time reacting later.

Frequently asked questions

Does Claude watermark its text?

Anthropic said models released after August 2, 2026 watermark generated text using a version of Google DeepMind's SynthID-Text technique, and it plans to extend this to older models.

Can AI text watermarks prove who wrote something?

No. Anthropic said detection only estimates the likelihood that Claude was involved, works poorly on short or factual passages and code, and can be defeated by a full rewrite. Press coverage noted text only polished with Claude may also be flagged.

What does the EU AI Act require of companies using AI?

Deployers must disclose deepfakes and AI-generated text published to inform the public on matters of public interest unless it is under human editorial review, and people must be told when they interact with an AI system such as a chatbot. Marking obligations fall on providers.

Disclosure: AI Enterprise Magazine uses AI tools in its editorial process, including Anthropic's Claude models. Anthropic had no role in this article.

Filed under: AI Governance & Trust, Security & Fraud, AI Strategy & ROI