Anthropic has published a detailed overview explaining how upcoming watermarks will function within text generated by its Claude chatbot. The move is designed to comply with the European Union AI Act’s Transparency Code, sparking considerable debate and mixed reactions among the user base.
According to the company, the system utilizes Google DeepMind’s SynthID-Text approach. It embeds subtle patterns in word choices that remain entirely invisible to human readers while allowing verification via a specific decoding key, with plans for a detection API. Anthropic notes that light edits will not easily remove the watermark, whereas complete rewrites or source code generation will largely bypass or minimize its presence.
This initiative represents a pivotal shift toward accountability and regulatory compliance in generative AI. As other major developers prepare to implement similar standards, transparent identification of AI-generated materials is becoming an industry norm.
- Anthropic adopts Google DeepMind’s SynthID-Text technology for watermarking.
- Implementation ensures compliance with the EU AI Act Transparency Code.
- Watermarks remain invisible to readers and survive minor text edits.
- Source code generation experiences minimal to negligible watermarking effects.
Sources:
