Here's what's on our plate today:

• 🧪 The 10B AI watermarks that almost nobody has a way to read.

• 📰 US makes allies pick an AI side, Amazon shreds rare books, Smack raises $61M.

• 🧠 Brain Snack: build the thing that reads the label before you ship the label.

• 📊 Poll: what does a mandatory AI watermark actually accomplish?

Let’s dive in. No floaties needed.

The Laboratory

TL;DR

The mark is finally mandatory. The system that would make it mean something still doesn't exist.

• Two announcements, one deadline: Anthropic now watermarks Claude's text and signs its files; Spotify is badging AI artist profiles. Both landed nine days after EU and California transparency rules took effect.

• Everything before this got pulled: OpenAI killed its detector in six months and shelved a working watermark. Meta rewrote its label after photographers revolted. YouTube and Spotify both abandoned creator self-disclosure.

• Law removes the exit: companies used to measure their own labeling systems and switch them off. A statutory watermark can't be quietly retired, which guarantees permanence and nothing else.

• Marking works, reading doesn't: Google's SynthID has stamped over 10B items, but verification sits behind a waitlist and only recognizes Google's own signal.

• Stakes: Deezer tagged AI music for a year while uploads climbed from 10k to 90k a day. Suppression held listening down, not the label.

Why AI watermarks may survive this time & still not work

In 1300, a statute of Edward I ordered that no silver leave a workshop until it carried the mark of a leopard's head. The statute also ordered that guardians of the craft be appointed to visit shops and test the metal first. The mark of a leopard's head, however, could be replicated, so around six decades later, a rule was introduced requiring silverwork to be stamped with the maker's own mark. The idea behind these edicts was to ensure that any piece of silver could be traced back to the person who made it.

But the marks and the large number of workshops in London introduced a new problem. The wardens appointed to check for them were limited in number, and the task of ensuring that every piece of silver had the correct markings was cumbersome and time-consuming. To get around this, the work of checking was shifted to Goldsmiths' Hall instead, where a salaried assayer tested it on the premises. That is where the word "hallmark" comes from, and the arrangement has been in place for more than 700 years.

The reason why the arrangement has lasted so long has little to do with the marking itself; rather, it has sustained because of the system built around it, which included a list of approved makers, people who checked the metal, serious punishment for fake marks, and a stamp that could not easily be removed.

And now, a similar marking system is being developed for AI-generated content. The difference between silver and AI-generated content, though, is that marking a precious metal did not create demand for unmarked pieces, but with AI, that is a very real possibility. And when that happens, private corporations, guided more by profit than by the need to build systems, may decide whether a human invested time and effort in creating something new, or whether it was created using the technology touted as the future of the human race. And in either case, the system relies on the very corporations that both mark the content and sell the tools whose output is being marked.

However, before any of that happens, it is important to understand why companies are trying to mark AI-generated content in the first place, even without a cohesive, holistic system in place.

Two announcements, nine days after the deadline

Last week, Anthropic announced that text generated by its Claude models now carries a watermark woven into the words themselves, invisible to readers and applied at the model level, so it travels wherever the text is pasted. According to TechCrunch reporting, this will be done using provenance metadata signatures under the C2PA open standard, and the signatures will be extended backward to older models.

Hours later, Spotify said it would begin attaching an "AI Persona" badge to artist profiles whose public identity appears to be machine-generated and would, by default, keep badged artists out of editorial and algorithmic recommendations. In a way, both announcements share the underlying promise that AI-generated content will now be distinctly visible to consumers. However, it is interesting to note that the announcements did not happen in a vacuum and were necessitated by the need to comply with upcoming EU legislative requirements.

Under the European Union's transparency obligations for AI-generated content, providers of generative systems must mark synthetic output in a machine-readable format starting August 2. And a similar law took effect the same day in California, in the U.S. Both laws aim to ensure greater transparency; however, achieving this is easier said than done.

What the industry did the last four times

Ensuring that consumers can differentiate between AI-generated and human-generated content is no easy task, considering the pace at which generative AI tech has evolved over the last couple of years. And the easiest way to understand it is to look at how past efforts to do something have fared.

In January 2023, OpenAI launched a tool intended to tell readers whether text was AI-generated and disclosed at launch that it caught 26% of AI writing while wrongly flagging human writing 9% of the time. According to The Register, the tool was withdrawn six months later due to low accuracy. OpenAI then built a text watermark that worked far better and never shipped it, updating a blog post in August 2024 to explain that translation or rewriting made it trivial to circumvent, and that it risked stigmatizing people who use AI to write in a second language.

Others fared better, but they too reversed their decision when even work done predominantly by a human and later edited using AI was flagged. The clearest example of this occurred in May 2024, when Meta began tagging images as "Made with AI," and photographers complained that ordinary retouching triggered the tag. Taking note of the criticism, Meta later changed the tag to "AI info".

YouTube and even Spotify tried a different approach: instead of marking content themselves, they asked creators to disclose when they used realistic synthetic footage. However, this did not work, and in May 2026, the video-sharing platform began automatically labeling videos. Spotify’s latest announcement also replaces its existing badge that asked creators to mark their content.

Each of these systems was designed, built, and deployed by the same companies responsible for evaluating its results and deciding what to do with them. That is what makes this time different. The August 2026 rules remove the easy exit: a watermark required by law cannot simply be abandoned when it fails, and a badge adopted under regulatory pressure cannot be quietly withdrawn without inviting further scrutiny.

The one that worked, and the half nobody uses

However, the legislation also changes what the marking will actually stand for. Compliance asks whether the mark was applied, and it does not ask whether anyone downstream was able to read it or act on it. Anthropic's page is unusually direct about the gap: it states that finding a Claude mark doesn't establish who wrote the underlying text, since people use the model to proofread and translate, and that finding no mark proves nothing either, because heavy editing, paraphrasing, translation, short passages, and format conversion can all remove it. This makes it more an exercise in establishing an archive of AI-assisted content than in informing consumers whether a piece of content was crafted by a human or generated by technology.

This reading of the situation is further strengthened by the success of Google’s SynthID, which has been running continuously since 2023 and has marked over 10B pieces of content by May 2025 across images, audio, video, and text. That is a scale that nothing else in this field approaches. Here, the marking itself works, but the problem is that ordinary people have no habit of checking, and no single place to check everything against, and the system only recognizes Google’s own marking. This effectively makes Google the identifier, the labeler, and the archivist.

And that is the shape of the whole problem. Europe legislated the mark and cannot legislate the rest of the medieval apparatus into being. There is no register of makers, because the 82 companies that signed the provider section of the Commission's code include Anthropic, Google, Meta, Microsoft, OpenAI, Mistral, and Cohere, and do not include xAI, Amazon, Apple, ElevenLabs, Stability AI, Midjourney, Suno, or ByteDance. There is no assay office, since nobody tests content before it is published. And the punch doesn't drive into metal, because the C2PA's own documentation concedes that metadata is removable by design and that platforms strip it routinely when they compress and re-encode uploads.

The medieval system also had one advantage that no law can now recover. Content generated before August 2, 2026, needs no label at all, meaning several years of synthetic text, images, and music will remain permanently unmarked, while everything produced after that date carries a stamp.

What actually moved the needle

The lesson of the hallmark was never contained in the mark itself. Its authority came from the system around it: a register of makers, people who tested the metal, rules that could be enforced, and a stamp that stayed with the object as it changed hands.

AI has begun building that system, with companies attaching marks to model output and governments requiring them. The system still lacks reliable verification, a common way to read the marks, a mechanism for dealing with marks that disappear, and clear accountability when they fail. Those gaps will determine whether AI watermarks become something like the hallmark, a mark whose presence carries meaning, or remain what they are today: signals of origin that offer little reason to trust what they tell us.

Brain Snack (for Builders)

Quick Bits, No Fluff

• US tells partners to pick an AI side: A draft State Department letter to the 35 signatories of Washington's AI Opportunity Statement warns that Pax Silica membership cannot be held alongside Beijing's rival coalition, telling countries to "choose deliberately."

• Amazon is shredding rare books for training data: 404 Media tracked a rare book to an Amazon facility in Las Vegas where spines get cut off and pages scanned, because nothing printed before 2022 could have been written by a model.

• Smack raises $61M on Pentagon urgency: The defense AI startup's Series B follows March's designation of Anthropic as a supply-chain risk, which sent the Marine Corps and Navy hunting for vendors who could produce faster.

Wednesday Poll

Meme Of The Day

The Toolkit

• Dust: No-code platform for building custom AI agents that connect to your company's tools and data.

• Framer: Pro website builder with AI agents on the canvas, so teams design, write, and ship production sites fast.

• Lavender: AI sales email coach that scores your drafts live and suggests rewrites to lift reply rates.

Rate This Edition