Watermarking AI

Anthropic says new Claude models will weave an imperceptible watermark directly into generated text—not images, but the text itself—to comply with the EU AI Act, while simultaneously courting investors for what could be the largest IPO of all time. The company is also adding C2PA signed provenance metadata to generated files. Meanwhile, Apple's glass iPhone plans remain on track despite rumors of cancellation, YouTube doubled its monetization thresholds for creators, FlightAware is suing Kalshi over flight cancellation betting markets, and Meta appears to be pivoting back toward open-weight AI.

Anthropic's dual track: watermarking and IPO

Anthropic published a help document explaining that going forward, Claude models will embed watermarks inside generated text and attach digitally signed provenance metadata to files. "You won't see it, and it doesn't change the meaning, quality, or reliability or readability of Claude's response," the company said. "Because the watermark is part of the text, it will travel with the text when it's copy and pasted elsewhere and may persist through some editing." The marking will apply to output from supported models wherever Claude is offered worldwide, including through third-party providers like AWS, Google Cloud, and Microsoft Foundry.