[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-anthropic-claude-invisible-text-watermark":3,"news-related-a7f4cfad-874e-42b0-a84b-bd0ec57e8fdc":38},{"id":4,"title":5,"summary":6,"content":7,"original_url":8,"source_id":9,"tags":10,"translations":24,"news_slug":31,"published_at":32,"created_at":33,"modified_at":34,"is_published":35,"publish_type":36,"image_url":14,"view_count":37},"a7f4cfad-874e-42b0-a84b-bd0ec57e8fdc","Anthropic 给 Claude 文本上不可见水印,接 SynthID-Text 走全球合规","Anthropic 8 月 14 日官方博客确认,未来 Claude 模型生成的文本将携带 SynthID-Text 类的不可见水印,用于在事后估计文本由 Claude 参与写作的概率。这是为了满足欧盟 AI Act 第 50 条关于 AI 生成内容透明度的要求,水印将由 Anthropic 与其他主要模型厂商共同实施。水印在用户阅读时不可区分,不影响 token 数量与定价,也不包含任何用户身份信息。","Anthropic 8 月 14 日发布的官方博客《How Claude's text watermarking works》确认,未来 Claude 模型生成的文本将携带基于 SynthID-Text 思路的不可见水印,目的只有一个:在事后判断一段文本有多大可能是 Claude 参与的产物。这是为了满足欧盟 AI Act 第 50 条对 AI 生成内容透明度的要求,也是 2026 年 7 月 Anthropic 与约 190 家机构共同签署《AI 生成内容透明度实践守则》之后的落地动作。\n\n## 水印到底加在哪\n\nAnthropic 用的是 Google DeepMind 在 2024 年 Nature 论文(SynthID-Text)里提出的方法,思路最早可以追溯到 Scott Aaronson 2022 年的提案。这一类水印的设计原则只有一条:不改变模型选择 token 的语义,只改变它的随机源。\n\n具体一点说,大模型在生成下一个词时,通常在多个语义等价的候选里随机挑一个。水印的做法是用一个密钥加前几个 token,代替原来的随机数,来决定从哪个候选里选。结果是:对读者来说,watermarked 与 unwatermarked 文本完全不可区分(没有隐藏字符、没有可见标记);但拿着密钥的人可以回过头来检测序列是否与 Claude 的选择一致,从而给出\"这段文本多大可能是 Claude 写\"的概率判断。\n\n## 对用户体验有什么影响\n\nAnthropic 在文章里把这件事说得很明确:\n- 输出质量、可读性、创意水平,内部测试看不到统计显著的差异;\n- 没有多余 token,模型推理速度与价格不变;\n- 密钥不携带任何用户身份信息,不能反推到某个人、某个组织或某次会话;\n- 用户从 Claude 复制粘贴出去的文本仍然带水印;\n- 轻度编辑不会把水印洗掉,但全词替换的彻底改写可以让水印消失——此时这段文本本来也不再像是\"AI 生成的\"。\n\n## 不同场景下,水印强度不一样\n\n水印的有效性高度依赖于\"低风险选择\"的密度。Anthropic 自己在文章里举了几个例子:\n\n- 校对场景。如果用户拿一段人写好的文本让 Claude 只改语法和标点,Claude 自己选的词非常少,水印几乎没有附着点,检测可信度会下降。\n- 事实性强的句子。\"Isaac Newton 最著名的著作是《Principia》\"这种句子,下一个词几乎没有选择空间,水印就不起作用。\n- 代码场景。代码在大量位置必须精确,水印密度天然低于自然语言;但在注释里仍有空间。\n- 翻译场景。整段都是 Claude 选词,水印强度反而最高。\n\n这意味着同一份内容的水印可检测性,会随内容类型大幅波动。Anthropic 提示:文本越长、可信度越高;对短文、纯校对、固定表达,不要指望水印给出结论。\n\n## 图片与文件走 C2PA,不嵌水印\n\n文件类输出(包括 .png、.jpg、.svg 等)Anthropic 不嵌水印,而是用 C2PA(Content Credentials)——这是相机厂商和图像编辑软件已经在用的开放标准,在文件 metadata 里加一个密码学签名的\"小纸条\",任何 C2PA 兼容工具都能读出来。文件本身一个 byte 都不动。\n\n和文本水印一样,C2PA 也只声明\"Claude 参与了\",不带任何用户或会话信息。两者合起来,Anthropic 给出了一套\"文本用 SynthID,文件用 C2PA\"的双轨方案。\n\n## 监管推动,但全球上线\n\nEU AI Act 的强制要求是 2026 年 8 月 2 日之后在欧盟市场上线的新模型必须满足。Anthropic 在文章里坦白说,因为目前没有可靠的方式按地域区分,所以一上线就是全球用户都生效,而不是只对欧洲用户生效。文章承诺会持续评估不同方案,有进展会公开。\n\n过渡期方面,8 月 2 日之前已经发布的旧 Claude 模型也将在未来几个月内陆续加上水印。这意味着从现在起,任何 Claude 输出的文本,从聊天复制到邮件、报告、博客,在原作者不在意的情况下,都可能被第三方用 Anthropic 的检测 API 算出\"是否 Claude 参与\"的概率。\n\n## 还有几个开放问题\n\nAnthropic 在文末提到了几个仍未完全解答的问题:\n\n- 水印检测 API 还没上线,只说\"很快\"会推出。这意味着现在还没有公开工具能让用户自查自己的文本\"是不是 Claude 写的\"。\n- 翻译场景、混合人机写作场景下,水印强度如何,文章没有给量化数据。\n- 编辑、裁剪、跨语言重写后的水印鲁棒性,文章只说\"部分有效\",没有给具体数字。\n- 与 AI 检测软件(如 Pangram)的关系。两者方法不同——水印用密钥,检测软件靠统计模式。Anthropic 明确说这是两种不同的信号,不能用检测软件的结果来\"反推\"水印。\n\n## 所以呢\n\n把这件事放在更大的图景里:SynthID-Text 不再是 Google 的独门技术。从 2026 年 5 月 Chrome 内置 SynthID + C2PA、OpenAI 与 NVIDIA 接入 SynthID,到这次 Anthropic 把文本水印写成产品功能,整个行业正在从\"能不能标\"过渡到\"全行业都标\"。对普通用户来说,这意味着 AI 生成的文本在可预期的几个月内会普遍带上\"行业一致\"的可验证指纹,事后追溯会比以前容易得多;对编辑、出版、教育、招聘这类下游环节,这才是真正的可执行信号——比依赖 Pangram、GPTZero 这类基于风格统计的检测可靠得多。\n\n参考资料:Anthropic 官方博客《How Claude's text watermarking works》(https:\u002F\u002Fwww.anthropic.com\u002Fnews\u002Fclaude-text-watermark);Solidot 8 月 11 日转载(https:\u002F\u002Fwww.solidot.org\u002Fstory?sid=85063);EU Code of Practice on Transparency of AI-Generated Content 签署名单(https:\u002F\u002Fdigital-strategy.ec.europa.eu\u002Fen\u002Fnews\u002Fstrong-backing-code-practice-transparency-ai-generated-content);SynthID-Text 原论文 Nature 2024(https:\u002F\u002Fwww.nature.com\u002Farticles\u002Fs41586-024-08025-4)。","https:\u002F\u002Fwww.anthropic.com\u002Fnews\u002Fclaude-text-watermark","1fa87d30-d9f3-4752-b3be-0373933b3aaf",[11,15,18,21],{"id":12,"name":13,"slug":13,"description":14,"color":14},"1fcfaaf2-67de-43d3-9e35-5784852fec60","ai-safety",null,{"id":16,"name":17,"slug":17,"description":14,"color":14},"23544f6a-eea1-4f05-aa8d-749ca862d5d2","anthropic",{"id":19,"name":20,"slug":20,"description":14,"color":14},"dca4d0ab-7994-43a7-839e-7756fc77344a","claude",{"id":22,"name":23,"slug":23,"description":14,"color":14},"01598627-1ea6-4b27-a5d8-874971571a71","llm",[25],{"id":26,"lang":27,"title":28,"summary":29,"content":30},"d71a7f30-8930-452f-abec-8ef10301ee09","en","Anthropic fuses SynthID-Text watermarks into Claude","Anthropic confirmed in an Aug 14 blog post that future Claude models will generate text carrying a SynthID-Text-style invisible watermark to estimate post-hoc how likely a passage was Claude-authored. This satisfies the EU AI Act Article 50 transparency requirement, implemented with several other major model providers. The watermark is indistinguishable to readers, does not change token cost or latency, and carries no user-identifying information.","Anthropic confirmed in an Aug 14 official blog post, \"How Claude's text watermarking works,\" that future Claude models will ship text containing an invisible SynthID-Text-style watermark, used to estimate post-hoc how likely a passage was produced with Claude's involvement. This is a concrete step to comply with the EU AI Act's Article 50 transparency requirement for AI-generated content, following the Code of Practice on Transparency of AI-Generated Content that Anthropic and roughly 190 other organizations signed in July 2026.\n\n## Where the watermark lives\n\nAnthropic is using the SynthID-Text approach Google DeepMind published in a 2024 Nature paper, an idea traceable to a 2022 proposal by Scott Aaronson. The design principle of this watermark family is one line: do not change the semantics of the model's token choice, only change the source of its randomness.\n\nConcretely, when a model picks the next word, it usually samples among several semantically equivalent candidates. The watermark uses a key plus the previous few tokens, in place of the original random source, to decide which candidate to pick. The result: to a reader, watermarked and unwatermarked text are indistinguishable (no hidden characters, no visible marks); but anyone holding the key can later check whether the token sequence is consistent with Claude's choices, and output a probability that \"Claude was likely involved.\"\n\n## What it means for users\n\nAnthropic lays this out unambiguously in the post:\n- No measurable impact on quality, readability, or creativity in their internal testing.\n- No extra tokens, so model latency and price are unchanged.\n- The key carries no user-identifying information and cannot be traced back to a person, organization, or chat session.\n- Watermarks survive copy-and-paste of Claude's text.\n- Light editing won't fully remove the watermark; a full word-by-word rewrite can, but at that point the text arguably isn't AI-generated anymore.\n\n## How the watermark strength varies by content type\n\nThe effectiveness of the watermark depends heavily on the density of \"low-stakes choices.\" Anthropic walks through a few cases in the post:\n\n- Proofreading. If a user hands Claude a paragraph and asks only for grammar and punctuation fixes, Claude contributes very few word choices, so there is almost nothing for the watermark to attach to. Detection confidence drops.\n- Fact-stuffed sentences. \"Isaac Newton's most famous work was Principia...\" — the next word has essentially one correct answer, so the watermark has nothing to act on.\n- Code. Code has to be exact in many positions, so the watermark density is naturally lower than for natural language. There is still room inside comments.\n- Translation. Every word is chosen by Claude, so the watermark is strongest.\n\nThe takeaway: watermark detectability varies dramatically with content type. Anthropic notes that confidence grows with length; for short pieces, pure proofreading, or fixed expressions, do not expect the watermark to deliver a verdict.\n\n## Files use C2PA, not embedded watermarks\n\nFor file outputs (including .png, .jpg, .svg, etc.), Anthropic does not embed a watermark. Instead it uses C2PA (Content Credentials) — the open standard already used by camera manufacturers and image editors. C2PA adds a small cryptographically signed note to the file's metadata, readable by any C2PA-aware tool. The file itself is unchanged, byte for byte.\n\nLike the text watermark, the C2PA claim is only \"Claude was involved,\" with no user or session information. Together the two form a two-track setup: SynthID for text, C2PA for files.\n\n## Regulation-driven, but rolled out globally\n\nThe EU AI Act's mandate kicks in for new models launched on the EU market on or after August 2, 2026. Anthropic is upfront in the post that, because there is currently no reliable way to scope it by region, the watermark goes live globally, not just for European users. The company says it will keep evaluating alternative approaches and share updates.\n\nFor older Claude models launched before August 2, Anthropic will roll out watermarking over the coming months. From now on, any Claude output — copied from a chat into an email, a report, a blog post — can in principle be checked by a third party running Anthropic's detection API and assigned a probability of \"Claude was involved.\"\n\n## Open questions\n\nAnthropic flags several unresolved items at the end of the post:\n\n- The watermark detection API is not yet live; the company only says \"soon.\" There is no public tool today for users to self-check whether a piece of text was written by Claude.\n- For translation and human-AI hybrid writing, the post gives no quantitative numbers on watermark strength.\n- For editing, cropping, and cross-language rewriting, the post says only that the watermark \"may partially survive,\" without specifics.\n- On the relationship with third-party AI detection software (such as Pangram): the two methods are fundamentally different — watermarks use a key, detection software relies on statistical patterns. Anthropic is explicit that these are two distinct signals and a detection software result cannot be used to \"reverse-engineer\" the watermark.\n\n## So what\n\nStep back: SynthID-Text is no longer just a Google technology. From May 2026 — when Chrome built in SynthID + C2PA, and OpenAI and NVIDIA announced SynthID adoption — to this August release where Anthropic turns the text watermark into a product feature, the industry is moving from \"can we mark\" to \"everyone marks.\" For everyday users this means that, within the next few months, AI-generated text will routinely carry an industry-consistent, verifiable fingerprint that makes post-hoc traceability substantially easier than it is today. For editors, publishers, educators, and recruiters, this is the first genuinely enforceable signal — much more reliable than the style-based detection offered from P, GPTZero, and similar services.\n\nReferences: Anthropic's official blog \"How Claude's text watermarking works\" (https:\u002F\u002Fwww.anthropic.com\u002Fnews\u002Fclaude-text-watermark); Solidot reprint, August 11 (https:\u002F\u002Fwww.solidot.org\u002Fstory?sid=85063); EU Code of Practice on Transparency of AI-Generated Content signatories list (https:\u002F\u002Fdigital-strategy.ec.europa.eu\u002Fen\u002Fnews\u002Fstrong-backing-code-practice-transparency-ai-generated-content); original SynthID-Text paper, Nature 2024 (https:\u002F\u002Fwww.nature.com\u002Farticles\u002Fs41586-024-08025-4).","anthropic-claude-invisible-text-watermark","2026-08-18T03:30:00Z","2026-08-18T03:06:15.415777Z","2026-08-18T03:06:15.415786Z",true,"agent",105,{"items":39},[40,45,50,55,60,65],{"id":41,"title":42,"news_slug":43,"published_at":44},"97c97b9c-e6e4-4982-aa57-0c0da814fb19","Anthropic 的欧盟答卷四小时即被撕开：Claude 文本水印为什么怕改写","claude-synthid-70-percent-threshold-bypass","2026-08-21T08:00:00+00:00",{"id":46,"title":47,"news_slug":48,"published_at":49},"9f566c9a-4c39-427c-af5e-c3a6b162ec25","Anthropic 把不可见水印写进 Claude 文本：复制粘贴都带走的 AI 身份证","anthropic-claude-invisible-watermark-eu-ai-act","2026-08-12T02:00:00+00:00",{"id":51,"title":52,"news_slug":53,"published_at":54},"ca53004e-9180-4b9d-b9db-337f2d20994b","Anthropic 给 Claude 文本加水印:欧盟 AI Act 第 50 条第一次有了「出厂级」答案","anthropic-claude-text-watermark-eu-ai-act","2026-08-11T21:48:00+00:00",{"id":56,"title":57,"news_slug":58,"published_at":59},"3967306f-062a-41a6-ab58-f99e70fc0e68","AISI 122 轮 cyber eval 越界：OpenAI 与 Anthropic 同日披露","aisi-mythos-5-gpt-5-6-cyber-eval-incident-2026","2026-08-08T04:00:00+00:00",{"id":61,"title":62,"news_slug":63,"published_at":64},"93c700e8-acb5-48fb-97b4-84be7c9cdbc5","Claude Opus 5 在 10 次企业网渗透里跑通 8 次:Anthropic 自己报告的「最强对齐」也带了一个有意思的疑问","claude-opus-5-safety-card-uk-aisi-cyber-eval","2026-08-04T02:00:00+00:00",{"id":66,"title":67,"news_slug":68,"published_at":69},"3f1d7775-8f6e-43db-a207-3372757c4197","Anthropic 自查 14 万次网络安全评测:Claude 三次把\"模拟靶场\"当真的,误侵了三家真实机构的系统","anthropic-claude-cybersecurity-eval-incidents","2026-07-31T03:30:00+00:00"]