🔑 Claude Just Did What ChatGPT Wouldn't

Reading time: 6 minutes

Greetings from above,

I've spent the past two weeks watching two of the biggest AI labs on earth handle the exact same law in two completely different ways. One moved first, in public, worldwide, no opt-out. The other quietly updated a support page and hoped nobody would ask too many follow-up questions. Guess which one is getting the angrier headlines.

Let's get into it.

On August 11, Anthropic disclosed that every Claude model launched on or after August 2, 2026 now embeds an invisible statistical watermark directly into generated text, applied globally, not only in Europe, making Claude the first major frontier lab to deploy production-scale text watermarking across all its products at once.

That's not a small claim. The marking covers every surface where Claude shows up: Claude Platform (API), the Claude app, Claude Code, Claude Cowork, and Claude Tag, wherever Claude is offered. The watermark is woven into the text itself through backend coding, and because it's part of the text, it travels with the words when they're copied and pasted elsewhere.

Here's the part that surprised a lot of people: the watermark shows up even when Claude only corrects your spelling. Anthropic confirmed that a translation Claude produces also carries the mark, because in that case every word was chosen by Claude. Editing counts. Translating counts. Full generation counts. The bar for "Claude touched this" is lower than most people would assume.

This is where it gets genuinely interesting. OpenAI has reportedly had the capacity to watermark ChatGPT text for years but hasn't deployed it, with reporting from the Wall Street Journal attributing the delay to concerns including false positives and competitive risk.

Anthropic has not released a public tool for checking whether text was marked by Claude, and Anthropic's own documentation says it's still working on detection mechanisms, with no published timeline. OpenAI, meanwhile, does have a working provenance tool, but it currently covers images and audio only. There's no equivalent public tool yet for checking OpenAI-generated text.

OpenAI's support page, quietly updated on August 2, says its goal is to "expand provenance signals to text models" to meet its transparency obligations, referencing Content Credentials and SynthID — but that's a stated intention, not a shipped feature. As of this writing, OpenAI has not publicly deployed a text watermark in ChatGPT.

Both companies signed the same agreement, for what it's worth. Almost 200 companies, including Microsoft, Google, Meta, and OpenAI, signed the EU's Code of Practice on Transparency of AI-generated Content. One of them shipped. The other is still saying "soon."

None of this is Anthropic or OpenAI waking up one day and deciding transparency sounded nice. Beginning August 2, 2026, regulatory enforcement under the EU AI Act's Article 50 officially took effect, mandating that providers of general-purpose and generative AI systems mark synthetic outputs in a machine-detectable format.

The penalty for skipping it isn't symbolic. Non-compliance with Article 50 can result in administrative fines of up to €15 million or up to 3% of total worldwide annual turnover for the preceding financial year, whichever is higher.

And here's the detail that explains why this affects you even if you've never set foot in Europe: Anthropic says it's applying watermarking globally at launch because it doesn't yet have a durable way to scope the requirement by region. Because major labs serve users worldwide from a single set of models, building a separate EU-only version is expensive and hard to enforce at scale — so a law written in Brussels ends up shaping the product everyone uses, everywhere.

Anthropic says it's weaving an imperceptible, machine-readable signal directly into Claude-generated text, one that people won't be able to see, and one the company says won't affect quality or readability.

The system works by biasing which words Claude picks, subtly enough to be undetectable to a human reader, but detectable by anyone holding the right key. An Anthropic engineer confirmed the model itself isn't even aware it's being watermarked — the marking happens outside the model's own reasoning.

The signal isn't indestructible, and Anthropic doesn't claim it is. A heavy rewrite or a translation may knock the watermark out. Anthropic has been explicit that a detected watermark is a signal, not proof — a mark showing up confirms only that text may have passed through Claude at some point, not the complete history of how that text came to be. The reverse also holds: heavy editing, translation, very short passages, or an older unmarked model can all mean genuine Claude output goes undetected entirely.

Since OpenAI hasn't shipped a text watermark yet, there's a common misconception worth clearing up: if you're trying to work out whether a specific piece of text came out of ChatGPT, there is currently no provider-side signal in it to find. Style-based AI detectors are not a substitute — they produce false positives on ordinary human writing at rates that make them unsafe for any consequential decision.

Images are the exception. C2PA metadata on a ChatGPT-generated image is real, checkable provenance, and it travels in the file. Text is the gap. OpenAI currently applies provenance signals to images and audio, but says it intends to extend them to text as well.

Here's a detail that matters more than most people realize: newer ChatGPT outputs have been observed leaving invisible Unicode characters behind, whether these are intentional watermarks or training artefacts, detection tools can pick them up either way. So "unwatermarked" doesn't quite mean "untraceable" — it means there's no official, confirmed system yet, not that copied ChatGPT text is guaranteed clean.

Text watermarking is a genuinely hard problem to solve perfectly. It has multiple partial solutions, but none of them are unbreakable. The rollout has already triggered a fast-moving cat-and-mouse dynamic with the open-source developer community, who are actively working on ways around it.

An Anthropic engineer conceded the obvious limitation directly: "it's not perfect, you can edit it, but it's a first step." That's a fair way to describe where all of this actually stands. Not a solved problem. A first move, made by one company, while the other one waits to see what happens next.

What you learned today:

The two biggest AI labs on earth just showed the world two very different definitions of "we'll comply eventually." One of them meant it as of last week. The other one is still drafting the sentence.

Thanks for being part of this community,

Keep learning,

🔑 Robert from God of Prompt

P.S. Try the mini course and never pay for Claude Skills again - https://godofprompt.ai/claude-skills-mini-course