Anthropic watermarks older Claude models, but leaves bypass on two versions
EU AI Act compliance drive expands to four older models starting November 3, but API users can dodge the mark on Opus 4.6 and Sonnet 4.6 with a single parameter tweak.
What to know
- Anthropic expands watermarks to four older Claude models starting November 3 to comply with the EU AI Act's December deadline.
- A loophole exists on Claude Opus 4.6 and Sonnet 4.6: developers can bypass the watermark by adjusting any one of three sampling parameters (temperature, top_p, top_k).
- Anthropic warns users to keep default settings but does not prevent the workaround, raising questions about the robustness of compliance.
- The watermark is meant to help distinguish AI-generated text from human-written content, but effectiveness remains disputed by experts who question whether it degrades text quality.
“The watermark contains no information about the user, their organization, or their conversations with Claude.”
Anthropic, AI company · Mashable ↗
Anthropic AI company deploying watermarksJohn Gruber Blogger
How it unfolded 2 developments, newest first · click a bar or a number to jump articlesposts
-
2
Mashable reports watermark rollout begins November 3
Mashable confirmed Anthropic's Friday announcement that older Claude versions will receive watermarks, noting this is the second phase of compliance after the initial August rollout.
-
first by Mashable, 10d ago
-
-
1
Anthropic announces watermark expansion with loophole on older models
Anthropic notified API customers it would expand watermarks to Claude Mythos 5, Opus 4.7, Opus 4.6, and Sonnet 4.6, starting November 3. On Opus 4.6 and Sonnet 4.6, users can disable the watermark by tweaking temperature, top_p, or top_k sampling parameters.
“If you rely on the watermark for your own compliance, keep these settings at their defaults.”
— Anthropic -
1 outlet first by The Decoder, 11d ago · read ↗
-
-
background
Anthropic begins watermarking new Claude models — Anthropic started watermarking all Claude models released after August 2, 2026, embedding an imperceptible pattern in word choice to comply with EU AI Act requirements.