Google unveils Pixel 11 phones, Watch 5 and Gemini AI upgrades

Anthropic announced this week that its Claude models will begin embedding invisible watermarks in text outputs to meet requirements under the European Union's AI Act. The policy applies to all new models released globally, not just those sold in Europe, and will mark any content the system processes, including human-written material that receives only light editing.

The company described the watermark as imperceptible to readers and designed to survive copying, pasting, and some modifications. It will also attach digitally signed metadata to generated files where possible. Anthropic tied the move directly to the EU law, which mandates watermarking for AI-generated or manipulated text, audio, images, and video. Models released after August 2 must comply, with a grace period running until December 2026 for older versions.

Tech investors and developers have raised immediate questions about the rollout. Bill Gurley, a prominent venture capitalist, noted that details on how the watermark functions remain limited, leaving uncertainty over who can detect it and under what conditions. Others worry the system could affect output quality or create uneven standards for lightly assisted writing versus fully human work. Because the mark applies even when Claude merely edits existing text, critics say it risks labeling ordinary documents as machine-influenced without clear justification.

The approach marks a broad expansion of content tracking. Rather than limiting the feature to overtly generated material, Anthropic has chosen to tag everything the model touches. This method ensures compliance across borders but raises practical issues for users who expect their own words to remain unmarked. No public technical specifications have been released yet, which has fueled speculation about potential side effects on speed, accuracy, or future restrictions.

The EU regulation frames these requirements as tools for transparency and accountability. Yet the decision by one leading AI firm to implement them worldwide suggests the rules carry weight beyond European markets. Companies face pressure to standardize practices that could influence how content is identified and evaluated in newsrooms, schools, and online platforms. Observers note that once such markers exist, they could be used to sort or deprioritize material in ways that extend past the original regulatory intent.

Anthropic has not detailed enforcement mechanisms or partnerships with detection services. The absence of that information leaves open questions about long-term control over what counts as authentic writing. As more firms follow similar paths, the line between human and machine contributions may grow harder to navigate without relying on proprietary tools held by the companies themselves.

The policy arrives at a time when public trust in both regulators and large technology developers remains low. Mandates that originate in Brussels and then apply to users everywhere illustrate how rules in one jurisdiction can shape global standards with limited input from those affected. Whether the watermarks deliver meaningful benefits or simply add another layer of hidden oversight will depend on future disclosures and real-world testing.