The mechanism relies on the SynthID-Text approach, which manipulates the statistical probabilities of word selection. When a model generates text, it often faces multiple equally valid options for a sentence, such as choosing between "overcast" or "grey" to describe weather. Under standard operations, the model uses a random number generator to pick; with the watermark active, it uses a specific key and preceding words to influence the selection.
Anthropic Adopts Google’s SynthID to Watermark AI Text
To comply with the European Union’s AI Act, Anthropic is rolling out invisible watermarks for Claude-generated text using technology developed by Google DeepMind. The system embeds machine-readable patterns directly into responses by subtly influencing word choices that do not alter the overall meaning or quality of the output.
This process leaves a distinctive, machine-detectable pattern within the text that remains invisible to human readers. Anthropic confirmed that this implementation will not increase costs for users or degrade the performance of the model. Alongside this text-marking system, the company is integrating C2PA support for images to align with broader transparency mandates. While Google has utilized SynthID for Gemini since 2024, other industry players like OpenAI face similar regulatory pressure to implement detection methods as the EU AI Act takes full effect.
Comments (0)
No comments yet. Be the first!