Anthropic watermarks all Claude text, and the disclosure problem lands on employers

Anthropic took the first visible step in operationalizing the code by creating a global opt-out watermark applied at model level across all Claude surfaces.

The Forbes column raises unresolved intellectual property questions, such as what happens when an employee writes a document themselves and then asks AI to edit it.

The EU Code of Practice on Transparency of AI-Generated Content took effect earlier this month, with around 200 companies having signed on in the abstract, including Meta, Microsoft, OpenAI and Anthropic.

The watermarks arrive at a time when many companies have been encouraging employees to use AI, and some people are now worried about being outed and their authorship of work questioned.

Rusic said of watermarks that "these systems can be useful evidence" but that he "would be very cautious about turning them into an automated judgment about authorship or misconduct".

The EU's watermarking laws apply to new models launched in Europe on or after August 2, with older models given a grace period until December.

Anthropic and other AI companies can change watermarking technologies at will, and watermarking is an active research area, so current detectability or removability does not predict how future text will behave.

A Business Insider review of the law, supported by interviews with AI and policy experts, suggests Anthropic's approach may go beyond what the EU requires.

Anthropic's watermark is designed to remain detectable when text is copied and pasted, and can appear in translations, summaries and more extensively edited text.

Caroline De Cock, senior fellow at the Lisbon Council and head of research at the Information Labs, told Business Insider: "Basically, Article 50 called for a scalpel to catch deepfakes and mass synthetic generation, but Anthropic is using a sledgehammer that hits simple proofreading and translation."

Anthropic's announcement triggered backlash: some Claude users said they cancelled subscriptions, and developers created new tools to remove the watermarks.

An Anthropic spokesperson told Business Insider that lighter editing with Claude is less likely to trigger the watermark because "there's very little (if anything) for the watermark to attach to".

Milos Rusic, cofounder and CEO of Deepset, told Business Insider that Anthropic is taking a somewhat broader interpretation of the EU law, because a provenance signal can attach to Claude's word choices even when it did not originate the underlying idea or document.

De Cock said the approach fits Anthropic's brand identity as a safety-first developer and could also reduce its regulatory risks by shifting the burden of filtering and interpreting those labels entirely onto downstream users and external platforms.

An Anthropic spokesperson told Business Insider the company would "continue to evaluate different approaches".

Google said it uses its SynthID technology to invisibly watermark AI-generated text, images, audio and video from its models worldwide.

OpenAI referred Business Insider to an FAQ saying it plans to expand its labels to text, but did not specify a date.

Meta directed Business Insider to a July blog post on interoperable tools to identify AI-generated content; it has so far focused on images and has not provided details of a text-watermarking plan.

xAI did not respond to Business Insider; the company tags some work as AI-generated, including White House press secretary Karoline Leavitt's departure message on X, and that tag later disappeared.

The interval between the August 2 trigger for new models and the December deadline for older models is roughly four months.

Of the four Anthropic rivals Business Insider approached, exactly one, Google, says it currently watermarks text; OpenAI has a plan without a date, Meta has covered images only, and xAI did not answer.

The Forbes column argues that one company has now set the de facto implementation standard that others get measured against, whether they intended to follow it or not.

Governments and other qualified parties may have received a watermark detector, and given the early nature of the space the efficacy of detectors remains to be seen.