O conteúdo está em inglês; a navegação e os filtros mantêm o idioma selecionado.

Claim of universal invisible Claude text watermark lacks primary confirmation

The supplied material conflates an alleged Claude Code request-marking experiment with claims that every Claude output now carries a text watermark; it does not establish the latter.

Publicado em 11/08/2026, 08:56:41Última atualização 11/08/2026, 08:56:41Publicação editorial

The report says Anthropic began embedding an invisible, machine-readable watermark in all Claude-generated text on August 2, alongside C2PA provenance metadata for files. However, the supplied support consists largely of social posts and secondary coverage, without a linked Anthropic announcement, help page, or technical specification.

A separate set of excerpts describes invisible Unicode markers in Claude Code request prompts, reportedly used to identify suspected unauthorized resale or model-distillation infrastructure. Request-side markers are technically and functionally distinct from a detectable watermark embedded in model output, so one claim cannot validate the other.

Accordingly, claims of worldwide coverage across every product, universal text marking, resilience after copying or editing, and third-party detection tools remain unverified. They should not be presented as established fact without primary Anthropic documentation.

Evidências das fontes 6

A verificação de fontes não garante exatidão permanente. Considere a fonte original, a data e as correções posteriores.

Mediummedium.com · Complementar

The result: Anthropic could read, from the server side, whether your Claude Code requests were being routed through flagged infrastructure — and you would have no idea this classification was happening, because nothing in your output would look different. The marker lived in invisible characters in a part of the prompt that doesn’t render visually. ## Why Anthropic Did This — And Why the Reason Doesn’t Fully Excuse It Anthropic’s explanation, delivered by engineer Thariq Shihipar within a day of the Reddit post, was that this was a March experiment aimed at detecting unauthorized API resellers and distillation attack pipelines. Anthropic’s Thariq Shihipar called it a March experiment against unauthorized resellers and distillation, and said the rollback was already planned. [...] ## What the Code Actually Did The mechanism worked in two stages, and both stages were designed to be invisible to the human eye. First, Claude Code checked your environment at runtime. Specifically, it lo