Article 50 Is Live with Anthropic the First to Comply

Article 50 of the EU AI Act became enforceable on 2 August 2026 which requires any generative AI platform to embed a machine-readable mark in AI generated content, so that downstream platforms and regulators can identify said content as machine produced.

The non-compliance carries penalties of up to €15 million or 3% of global annual turnover, whichever is higher.

Anthropic responded by watermarking the text output of every Claude model launched on or after that date, and applied the mark worldwide, not just European users. According to Anthropic, the mark travels with the text even when it’s copied and pasted, and may persist through some editing.

Anthropic isn’t First as Google Already Got There Quietly

Anthropic’s watermark is genuine, operating at the model level, across its API, Claude Code, Claude Cowork, Claude Tag, and the cloud platforms that resell Claude, so it flows automatically into every third-party product built on it.

Google got there first, and quietly when Gemini started watermarking its own output since 2023, extending the technique to text in 2024, two full years before Anthropic’s rollout, and yet it drew none of the reaction Anthropic is getting this week. Google’s text watermark runs inside its own Gemini app and web product. While narrower in scope compared to Anthropic’s rollout, the underlying watermark technique isn’t new.

OpenAI openly admitted in July that watermarking text at scale remains hard to deploy. With Anthropic and Google having both done it, perhaps Anthropic’s rapid response to Article 50 is a marketing flex to signal their tech capabilities?

Social Platforms were already fighting AI slop, not just Anthropic

Generative AI collapsed the cost of producing content to almost nothing, so much so that social media platforms have spent the past year absorbing the consequences in their feeds.

LinkedIn started throttling the reach of posts that read as AI-generated, adding a “Seems like AI Slop” button for users to flag them, and has even replaced its own AI rewrite tool with a plain proofreader.

YouTube terminated 16 channels in January, erasing a combined 35 million subscribers, 4.7 billion lifetime views, and roughly $10 million in annual ad revenue, for mass-produced, low-effort content.

Reddit announced in July that its systems now block 23 million spam views a day, catch 25,000 spam posts and comments daily, and revoke close to 2 million inauthentic votes daily.

Of course none of this was prompted by Article 50 as these platforms built detection and demotion systems on their own. The content flooding their feeds was already costing them the thing that makes them valuable: real people, saying real things. Anthropic’s watermark does at the point of generation what these platforms have been trying to do after the fact, once content already reaches a feed.

Reading as AI-Generated Is Already Costing Businesses Visibility

A Copenhagen Business School study published this year ran two controlled experiments which found that labelling content as AI-generated or AI-enhanced reduced both how much people engaged with it emotionally and how much they acted on it. 

What LinkedIn, YouTube and Reddit are already doing to address AI slop shows content that reads as AI-authored, regardless of watermark checks, is treated differently by both the algorithm and the person reading it.

The lesson for business is undeniably clear if you want to reach target buyers in AI search and social media: Produce content that delivers perspectives that nobody else can generate from AI. That means prose that’s grounded in your own data and written in the language your buyers actually use that’s memorable. 

After all, if you’re not saying something different – why should AI recommend you over the status quo? Why should LinkedIn prioritize your content in your buyers’ feeds?

Signal over Slop is How Our Clients Stay Ahead

Article 50 and the platform crackdowns are both responses to the same problem, which is an internet filled with content that repeats what already exists and adds nothing of value to it.

We’ve built Artios on content that teaches AI something it didn’t already know, written by people and grounded in data science. That’s information gain, and it’s the reason our clients have held their visibility through several rounds of algorithm change across organic and AI search.

It’s also how we can guarantee your business gets recommended by AI. That confidence comes from having proven the approach through changes like this one.

If you want to see how your business benchmarks against your market competitors in AI search  – talk to us.

Yours,

Andreas

Founder & CEO, Artios.io

PS If you know a business who’d like to win the search game, please make an introduction or direct them to artios.io/winning-the-ai-search-game

Andreas Voniatis is founder of Artios.io, author of Data Driven SEO with Python (Springer, 2022), and built his first Large Language Model in 2019 – four years before ChatGPT launched.

Artios helps challenger B2B businesses get chosen by AI when buyers search there.