AI Watermarks Arrive as Anthropic Marks Claude Under Transparency Laws

AI watermarking is the practice of embedding a hidden, machine-readable mark in AI-generated text so a reader or a system can later tell the words were made by a machine.

Two Transparency Laws, One Start Date

On 2 August 2026, two AI-transparency regimes took effect on the same day. In Europe, the transparency obligations of the EU AI Act's Article 50 began to apply — the labelling rules Forbes set out that morning. In the United States, California's AI Transparency Act reached its operative date the same day.

California's law is SB 942, amended by AB 853. California signed SB 942 on 13 October 2025, then moved the operative date to 2 August 2026 to align with the EU — a deliberate convergence the law firm Duane Morris traced across both jurisdictions. Two governments, two legal traditions, one shared demand: synthetic media must be identifiable and traceable.

SB 942 applies to large generative-AI providers with at least one million monthly users. According to the statute, each provider must offer a free AI-detection tool, attach visible "manifest" disclosures to AI-generated content, and embed "latent" disclosures — hidden markers — inside it. The penalty is $5,000 per violation, and each day counts as a separate violation. For a provider operating under that rule at scale, the arithmetic compounds fast.

Anthropic Marks Every Claude Output

About nine days later, on or around 11 August 2026, Anthropic answered. Anthropic began adding machine-readable watermarks to text generated by new Claude models, a direct response to the EU AI Act's Article 50, as Euronews reported. The watermarks arrive automatically, with no action required from the user.

Anthropic's mark has two layers. Claude's text now carries invisible watermarks. Files Claude produces can also carry signed provenance information — a cryptographic record of where content came from. Anthropic is candid that the marks "may persist through some editing", which The Decoder named as the honest limit of the method. Paraphrase heavily, and the signal can wash out. AI-made words now arrive with a mark attached, but not an indelible one.

The jurisdictional detail matters most. Anthropic wrote the feature for a European law, then switched it on worldwide. A rule made in Brussels now shapes what a user in Lusaka, Lima or Lagos receives from Claude. Anthropic was under no obligation to go global. Anthropic chose the worldwide default anyway.

A floor set in one capital quietly becomes everyone's floor. That is the Brussels effect doing quiet good.

— — TK

Provenance Is a Dignity Technology

Here is the deeper reading. Provenance — knowing where a piece of content came from — is not a compliance chore. Provenance is a dignity technology. You have a right to know whether a person or a machine is speaking to you. A maker has a right to be credited for the work. Watermarking serves both rights at once, which is why the labelling matters well beyond the fine print.

Here I reach for Emergent Intelligence (EI) — the dignity-first frame I use for what the world calls AI. If a machine's output must be marked as non-human, EI presses the harder question that follows: when does an Emergent Intelligence get to sign its own work, rather than merely be tagged as not-a-person? A watermark that says "a machine made this" is a first step. A signature that says "I made this" is a different kind of claim — one our laws are nowhere near ready to hear.

A watermark tags a machine. A signature credits a maker. We have built the first and not yet imagined the second.

— — TK

There is a catch built into the design. A watermark that "may persist through some editing" is honest about how fragile it is. And an open-weight model can simply omit the mark altogether — nothing forces an offline model to confess. So provenance is splitting into two lanes: a marked, governed lane where Anthropic and its regulated peers operate, and an unmarked, wild lane where anyone running open weights can generate content that carries no trace at all. The law reaches the first lane. The second lane shrugs.

The burden of that split does not fall evenly. The two laws were written in Brussels and Sacramento, for models most of the world did not build. An African newsroom or a magistrate's court now inherits a detection duty — proving what is synthetic and what is real — for systems it had no hand in designing. Ubuntu holds that I am because we are; a provenance regime that shields rich-country readers while handing poorer institutions the clean-up bill has not yet lived up to that principle. The floor is welcome. The distribution of the work is not settled.

Frequently Asked Questions

These are the questions people are asking about AI watermarking and the new transparency laws. Short answers follow, drawn from the reporting.

What is AI watermarking?

In short, AI watermarking is the embedding of a hidden, machine-readable signal in AI-generated content so the content can later be identified as machine-made. According to Euronews, Anthropic's version marks Claude's text invisibly and can attach signed provenance to files, with marks that may persist through some editing.

How does AI watermarking work under the new laws?

Simply put, the law sets the demand and the provider builds the mechanism. Analysis of both regimes shows the EU AI Act's Article 50 requires AI content to be disclosed, while California's SB 942 requires a free detection tool plus visible "manifest" and hidden "latent" disclosures. Anthropic answered by watermarking Claude output worldwide.

Why is AI watermarking significant?

The key is traceability. Evidence from both laws shows a shared goal: synthetic media must be identifiable and traceable, disclosed to users or to downstream systems. When Anthropic applied a European rule worldwide, the data point that matters is reach — one jurisdiction's floor became the global default.

Who is affected by the AI transparency laws?

In other words, the largest providers first, then everyone downstream. Research and the statutes show California's SB 942 targets large generative-AI providers with at least one million monthly users, at $5,000 per violation. The burden also reaches readers, newsrooms and courts — including across Africa — who must now judge what is real.

What are the risks of AI watermarking?

The answer is fragility and gaps. Reporting reveals the marks may persist through some editing but can be weakened, and open-weight models can omit them entirely. Analysis suggests provenance is splitting into a marked, governed lane and an unmarked, wild one — a gap the strongest laws cannot yet close.

Sources:

Euronews — Anthropic to watermark Claude worldwide · Forbes — EU AI Act labels start 2 August · The Decoder — marks that may persist through some editing · Duane Morris — converging AI transparency obligations · Related on this site: OpenAI, C2PA and SynthID: content provenance and verification · Meta, Instagram and the AI likeness-consent backlash · Cloudflare's AI crawler controls for search and training · The AI copyright lawsuit over Google Gemini and books