The system avoids marking code tokens to preserve accuracy but leaves a detectable gap.
The New Stack reports that Anthropic's Fable 5.1 watermarks text output. It skips code tokens that could break accuracy if marked. This creates a blind spot developers cannot ignore. A separate API restriction also targets model distillation at scale. The article states the watermarking system leaves this gap in coverage for certain content types while the restriction seeks to limit large scale extraction of model behavior.
516
1 post, first seen 12h ago
The system avoids marking code tokens to preserve accuracy but leaves a detectable gap.
The New Stack reports that Anthropic's Fable 5.1 watermarks text output. It skips code tokens that could break accuracy if marked. This creates a blind spot developers cannot ignore. A separate API restriction also targets model distillation at scale. The article states the watermarking system leaves this gap in coverage for certain content types while the restriction seeks to limit large scale extraction of model behavior.
The New Stack@thenewstackClaude Fable 5.1 watermark: It has a blind spot developers can’t ignore
The New Stack@thenewstackClaude Fable 5.1 watermark: It has a blind spot developers can’t ignore
