Recent Posts

Anthropic just kept a promise it made three years ago. The promise was never the hard part.

Anthropic's Claude models now embed an invisible watermark in every response, plus signed C2PA metadata on every file, with no way to turn either off. The company is framing this as leadership. It is closer to follow-through. Anthropic pledged to build exactly this at the White House on July 21, 2023, alongside six other companies making the identical commitment the same afternoon. It shipped 1,108 days later. TechCrunch and Forbes covered the announcement without checking either date.

One pledge, six clocks

Google had a working beta in 39 days. Meta and OpenAI followed at 200. Google's own text watermark, the direct ancestor of Claude's, landed at 298. OpenAI abandoned its separate approach and adopted Google's method outright at 1,033. Microsoft's watermark slipped twice past its original February target before its own release notes confirm a late-June launch, at just over 1,060. Anthropic shipped last, five weeks after Microsoft. The pledge was identical. The clock was not.

The mechanism inside Claude's mark, Google's SynthID-Text, was open-sourced through Hugging Face in October 2024, tested against twenty million live Gemini responses first. C2PA, the provenance standard underneath all of it, traces to Adobe's Content Authenticity Initiative, founded in November 2019, four years before the pledge that supposedly started this.

What the rule says it's for

Recital 133 of the AI Act lists logging alongside watermarks, metadata, and cryptographic fingerprints: one detection method among several, not a separate record of judgment. The Act's actual accountability machinery, human oversight and record-keeping under Articles 8 through 15, applies only to systems classified high-risk under Article 6: hiring, credit scoring, biometric identification. A chatbot sits outside that classification entirely. The one place Article 50 gestures at human responsibility at all is a narrow carve-out: a deployer publishing AI-generated text on a matter of public interest is exempt if a named person holds editorial responsibility for it. The rule applies that distinction once, to one publishing scenario, and stops. It never asks what the named person actually reviewed, or extends the question to the providers building the systems in the first place. The European Commission's own guidance says as much: Article 50 compliance does not establish that a system is lawful under the Act.

Testing the asserted side

Anthropic's own documentation concedes the mark's limits: light editing survives, a full rewrite does not, and the tool cannot tell a passage Claude wrote from one it heavily revised. Paraphrase attacks and a dedicated framework for stripping n-gram watermarks both defeat the signal outright, in published research, not theory. The signal is real. It is also easy to remove.

Three systems, three approaches

The divergence becomes clearer at the regulatory level. China's AI content-labeling rules took effect on September 1, 2025, through a single administrative framework covering text, image, audio, and video. WeChat and Douyin moved to comply immediately. The EU arrived at transparency through a different route. Article 50 sits inside the AI Act's broader risk-based structure, with the rules for AI-generated content moving through codes, guidance, and implementation processes.

The US has taken neither path. The AI Labeling Act of 2026 remains in the Senate, while California has moved separately. Its disclosure rules became operative on a timeline tied to the EU's transparency requirements, although the state's framework takes a narrower approach to the types of content it covers, following amendments signed into law in 2025.

The difference is not simply speed. China can impose a technical requirement through administrative authority. The EU embeds transparency obligations within a comprehensive, rights-based and risk-tiered framework. The US remains divided between federal proposals and state-level legislation. Three legal systems. Three different routes to the same question: how should people know when AI was involved?

Provenance is not accountability

The industry has now watermarked over 100 billion pieces of content. Every one of those marks answers the same question: was a machine involved. None of them answers what the machine decided, on what basis, or who checked it before it went out. That second question doesn't have a pledge behind it, a coalition working on it, or a deadline attached to it. It's the one that actually determines whether an organization can stand behind AI-assisted work when someone asks it to.

For a business using AI at scale, the harder question begins after the output exists. What was the system asked to do? What did it produce? What was changed? Who reviewed it? Who approved it? Those are the questions a watermark cannot answer. They require a record around the use of AI itself: the instructions, the review, the changes, the decision, and the person who ultimately stood behind the result. That is the gap Occams' AI Integration practice is built to address. We help businesses put AI into real workflows with the review, controls, and documentation needed to know not only that AI was used, but what happened between generation and release.