Do watermarks spell the end of AI content proliferation?

AI companies are now required to make model outputs detectable as AI-generated under Article 50 of the EU AI Act – or, in other words, watermarked.

The major AI companies have all signed the additional voluntary Code of Practice. OpenAI has still not announced what measures it will be using, despite having built and not released a text-watermarking capability years ago. Google already watermarked AI generated images and videos, which it has since extended to text. Anthropic, however, has triggered the biggest headlines, unveiling text watermarks across all models (including older ones) and signed C2PA metadata in generated files, worldwide – not just in the EU. 

AI’s cultural shift

There are important practical reasons for watermarking AI output, with educational outcomes and professional verifiability highest among them. There is also a cultural factor, as social platforms have begun flagging and downweighting AI content, most notably on LinkedIn. 

AI can be a useful tool, but when over-relied upon it can be to the detriment of the final intended outcome. Watermarking allows for quality control in this respect. 

Responses, however, have been mixed. While Anthropic stated it has seen no notable drop in usage, sentiment has reportedly been shifting, as per Forbes. Some users are concerned that the watermark can appear on text they have only used Claude to edit, making it indistinguishable from wholly generated text and resulting in reputational decline. Others cite concerns that building workflows around a single company’s product can leave them at the whim of that company’s subsequent product decisions. 

The three main reasons consumers use AI is to research / find information (around half of all consumers); to help complete work or study tasks (roughly a quarter); and for ideas / brainstorming (also a quarter), per MIDiA’s Q3 2025 consumer survey. Its use case lends well to completing work more efficiently. However, when watermarked, AI generated or affected text becomes a double-edged sword, potentially signalling a lack of effort or ownership of the output.  In one move, AI has gone from hidden superpower to potential liability for the majority of consumers using it on a casual basis. 

AI versus authenticity 

There are workarounds to the watermarks. Editing can typically disrupt the watermark enough to not get flagged. A free tool to strip Claude’s watermarking appeared on Github within 24 hours. And, of course, ChatGPT is still watermark-free – which is by far the most-used LLM. According to MIDiA’s Q2 2026 consumer survey, 50% of all consumers use ChatGPT, rising to 68% of 20–24-year-olds. Gemini hovers around 40%, while Claude is a distant fifth at roughly 10%, behind other competitors Meta AI and Microsoft Copilot. 

However, ChatGPT will have to comply eventually or risk losing out on the EU’s consumer base. Workarounds take effort on their own, adding friction to what has, until now, been a largely frictionless experience. 

Moreover, even if watermarks are easy to work around on a technical level, there is a broader cultural pull towards authenticity as a differentiator. With social platforms changing their strategies to downweight AI content (most notably through LinkedIn’s “Seems like AI slop” feedback option), it is clear that the marketplace is no longer embracing the easy output of LLMs and is turning back towards a human-centric creative approach.

This raises the bigger question: will AI watermarking become culturally embraced as a part of everyday digital life – or a stamp of partial effort  as audiences increasingly pull away from AI? 

The answer will likely be context-dependent. There are clear use cases for AI in certain professional industries, like coding, for example, where watermarking will merely be confirming what is already accepted and encouraged as a use case. However, as a mass-market proposition, and especially as a content type on social platforms, watermarking will likely spell an era of constriction for AI content creators.