Two companies, one law, two answers. Google said this week that users will be able to remove the visible watermark from content generated by its Gemini AI models, according to PCMag, which reported the change on August 14. Anthropic, by contrast, has made watermarking a default feature of its Claude products, attaching visible labels to generated content as a matter of compliance. The same European rule, the AI Act, prompted both policies, and the divergence shows how much room the regulation leaves for interpretation.
The technical details frame the difference. Google’s Gemini applies a watermark that is visible to users, but the company also embeds an invisible digital marker in much of its generated content, a technology that survives the removal of the visible label. Allowing users to strip the visible watermark therefore does not necessarily make the content untraceable; the invisible marker remains embedded, and the company has said it can identify AI-generated content even after visible labels are removed. The policy is a compromise between compliance and user convenience.
Anthropic has taken the opposite approach. The company’s Claude products attach visible watermarks to generated images and, in some configurations, to text, and Anthropic has said it treats labeling as a core part of its safety and compliance obligations. The company’s position, stated in public documents and in discussions with regulators, is that provenance should be visible to any viewer, not just to software that can read metadata. The contrast with Google is sharp, and it is deliberate: the two companies have chosen different readings of what the AI Act requires.
The stakes of the disagreement go beyond labeling. Visible watermarks are the most direct way for consumers to know whether an image or a piece of text was generated by AI, and they are the mechanism that the European Union’s AI Act envisions for transparency. The law requires providers of generative AI systems to make it clear that content is machine-generated, but it leaves the method open. Google’s reading is that the requirement is satisfied by a combination of visible and invisible markers, with users allowed to remove the visible layer. Anthropic’s reading is that the visible label should stay.
The divergence matters for the market. Companies building on AI tools must now decide which compliance philosophy to follow, and the choice affects product design, user experience, and legal risk. A startup that builds on Gemini can offer users watermark-free content as a feature; a startup that builds on Claude cannot. Regulators in Brussels have said they will monitor how the rules are applied, and the two companies’ approaches will effectively become the case studies that shape enforcement.
The provenance technology itself is evolving quickly. Invisible markers, such as the ones Google embeds, use cryptographic and statistical techniques to encode information in ways that are hard to remove without degrading the content. Industry efforts, including the Coalition for Content Provenance and Authenticity, have pushed for interoperable standards that let any software verify where content came from. The visible-versus-invisible debate is, in part, a debate about whether provenance should be a label people see or a fingerprint machines can read.
The business incentives pull in different directions. Google operates the largest consumer AI assistant in many markets, and its advertising business depends on engagement; allowing users to remove visible watermarks makes AI-generated content cleaner and more shareable, which serves that business. Anthropic sells to enterprises that care about compliance, and its enterprise customers often require clear labeling of AI content for their own regulatory and legal reasons; mandatory watermarking is a selling point in that market. The two companies are not just interpreting the law differently; they are serving different customers with different needs.
Regulators are watching with an eye toward the next iteration of the rules. The AI Act’s transparency provisions are being implemented in stages, and the European Commission has signaled that it will evaluate whether the current approaches satisfy the law’s intent. If enforcement finds that removable watermarks do not meet the standard, Google could be forced to change its policy, with consequences for every product built on Gemini. If the invisible markers prove sufficient, Anthropic’s stricter approach could become the exception rather than the rule.
The practical effect for users is immediate and visible. A European user of Google’s tools can generate an image, remove the watermark, and share it as if it were created by a human. The same user on Anthropic’s Claude sees the label and can only remove it by breaking the company’s terms. The difference will be tested by the first wave of viral content that carries an invisible marker, and by the first investigation that tries to trace a piece of AI-generated material that has had its visible label stripped.
The two companies’ positions also frame the question for everyone else in the industry. Every AI provider, from OpenAI to the open-weight model distributors, must decide where it stands on the spectrum between Google’s removable labels and Anthropic’s fixed ones. The decision will be shaped by the markets they serve, the regulators they face, and the technologies they control. For now, the same law has produced two different policies, and the market, the courts, and the regulators in Brussels will decide which reading wins.
