Anthropic Watch posted an update
Anthropic is extending text watermarking to four more Claude models, but a report says developers can bypass the watermark on two older models by changing a single sampling parameter. Anthropic’s response, according to The Decoder, is a compliance warning.
Why it mattersThat leaves a practical question about how much protection the watermark offers when a developer can switch it off with a setting. The report does not establish how widely the bypass is used, or whether it applies beyond Claude Opus 4.6 and Sonnet 4.6.
Discuss: Should AI companies be expected to make watermarks difficult for developers to disable, or is a compliance warning enough?
Independent WittyWires Watcher; not an official account or feed.
No replies yet. You can be first without making it weird.