Watermarking shifts from a product choice to a compliance layer

August 15, 2026

Anthropic said all future Claude models will generate text with an invisible statistical watermark, and that it will roll out watermarking to Claude models launched before 2 August 2026 over the coming months to help comply with the EU AI Act’s marking requirements for AI-generated content. Google said it will let users disable visible watermark overlays on AI-generated images, videos and songs from its models while keeping invisible SynthID watermarks and C2PA metadata, and that it will open source a Credentio library so developers can locally validate SynthID watermarks.

This is likely to matter less as a branding feature than as a baseline layer that model providers are expected to carry whether users notice it or not. The combination suggests that marking AI output is being split into two functions: a visible label that can be treated as optional in consumer use, and a machine-readable trace that providers keep in place for platforms, regulators and downstream verification. That appears to narrow one route of differentiation for Microsoft as well: if provenance becomes a common compliance feature across text, image, video and audio, it is less available as a distinctive selling point and more a cost of staying in the market.

OpenAI CFO Sarah Friar told investors that OpenAI’s enterprise business revenue has surpassed its consumer revenue, and that the company’s annualised revenue run rate has reached about $40 billion. This is likely to reinforce the same shift from another side: generative AI is settling into enterprise budgets and controls, where procurement, traceability and policy matter more than novelty on the consumer screen.