Anthropic, a leading AI model provider, has announced that it will begin watermarking content processed by its models to comply with the European Union's AI Act. The law, which applies to all AI system providers, requires watermarks to be added to AI-generated or manipulated audio, image, text, and video outputs. This move is set to affect all new models released globally, not just in the EU, starting immediately.
The EU's AI Act, which came into effect on August 2, provides a grace period until December 2026 for providers to update previously released models. Anthropic has confirmed that all new models will carry embedded watermarks from the outset, with text outputs featuring invisible watermarks and other generated files including digitally signed provenance metadata where supported.
The company is taking a comprehensive approach to watermarking, applying it to all processed content where supported, regardless of whether the AI system is performing an assistive function or substantially altering the user's text. This means that even minor edits, such as grammar corrections, may be watermarked. However, the effectiveness of this approach will not be fully known until Anthropic releases a detection tool that can be tested.
Anthropic's decision to watermark all processed content is more extensive than required by the EU's AI Act. The law does not require watermarks for cases where an AI system performs an assistive function, such as grammar correction, or where it does not substantially alter the user's text or its meaning. Despite this, Anthropic is choosing to apply watermarks across the board, which may result in watermarked content that the law was not intended to target.
The company has stated that it plans to share details on how to detect watermarks in order to offer technical support, as required by the EU's law. This will provide more clarity on how the watermarking system works and how it can be identified. For now, the watermarks will remain invisible to users, but the impact of this new approach on AI-generated content will be closely watched in the coming months.
What's Next
As Anthropic rolls out its watermarking system, the company will likely face scrutiny over its implementation and effectiveness. The release of a detection tool will be a key milestone in understanding how the system works and how it can be used to identify watermarked content. With the EU's AI Act setting a new standard for AI regulation, other providers will be watching Anthropic's approach closely, and it may set a precedent for the industry as a whole.
<small>Source: Ars Technica — read the original story there.</small>
