[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-claude-text-watermark-eu-ai-act-global":3},{"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},"a124851a-081e-44b6-9f20-f775404279e1","Claude 全球文本水印:Anthropic 把欧盟 AI Act 第 50 条做成\"全球默认\"","Anthropic 宣布为 Claude 文本与生成文件嵌入机器可读水印,适用 8 月 2 日后在欧盟发布的新模型,并将范围扩大到全球。这是欧盟 AI Act 第 50 条透明度义务首次在头部模型上的全量落地。","## 事件:Anthropic 把\"AI 内容可追溯\"做成了默认能力\n\nAnthropic 在 8 月初正式上线了 Claude 的机器可读水印方案,覆盖文本与生成文件两类产物,目的是满足欧盟 AI Act 第 50 条关于 AI 生成内容透明度的强制要求。该方案覆盖所有 8 月 2 日及之后在欧盟上线的 Claude 新模型,并将适用范围扩大到全球(在 Claude 提供的每一个地区都生效),包括 Claude 应用、API、Claude Code、Claude Cowork、Claude Tag,以及通过 AWS、Google Cloud、Microsoft Foundry 接入的 Claude 模型。\n\nAnthropic 已签署欧盟\"AI 生成内容透明度实践守则\"(Code of Practice on Transparency of AI-Generated Content),这是欧盟为 AI Act 第 50 条配套推出的合规路径之一,8 月 2 日起与该条款同步强制执行。守则已有约 190 家机构签署,Google、Meta、Microsoft、OpenAI 等头部厂商也都在签署名单中(来源:Unite.AI)。\n\n## 技术细节:两种水印,各管一摊\n\nAnthropic 在支持页里把方案拆成两种互补的标记机制:\n\n1. **文本水印**:模型在生成过程中把不可见水印\"织\"进文字本身。读者完全感知不到,语义、质量、可读性都不变。因为水印嵌在文本里,复制粘贴后会一起带走,部分编辑后也能保留。水印在模型层面生效,无论从哪个 Claude 入口输出都一致。\n2. **文件来源元数据**:当 Claude 生成 .svg、.png、.jpg 这类支持的文件时,会附带 C2PA 标准的数字签名来源元数据。签名能让第三方工具确认\"该文件经过 Claude 处理\",并检测是否被篡改。C2PA 是行业内被广泛采用的内容来源开放标准。\n\n支持页同时强调了两条边界:一是检测到水印只代表\"内容**可能**经过 Claude\",不能等同于完整出处证明——Claude 也常被用来校对、翻译、总结第三方内容;二是没检测到水印不代表内容不是 AI 生成的——前置模型输出、深度改写、极短片段、文件元数据被剥离或截图重存,都可能导致漏检。\n\n## 行业对照:Anthropic 不是第一家,但动作最快\n\nGoogle DeepMind 的 SynthID 是同类方案的先行者,覆盖文本、图像、音频多个模态。OpenAI 长期讨论水印方案,但实际部署明显更慢。Anthropic 这次的差异化在于三件事:\n\n- **覆盖文本**:SynthID 文本路径公开信息有限,Anthropic 把\"文本水印\"明确做成了产品默认;\n- **全球生效**:选择不在欧盟和非欧盟之间做产品分叉,而是把水印做成全球默认,实际等于把欧盟标准输出成了全球基线;\n- **配套元数据**:同步引入 C2PA 签名,把图像、矢量等文件来源追溯能力一并补上,而不是只给文本加水印。\n\nAnthropic 自己也声明,目前检测工具还在完善中,完整的技术检测文档会\"在未来发布\"。\n\n## 合规与边界:罚则、市场监督与\"deployer 责任\"\n\n欧盟 AI Act 第 50 条透明度义务自 8 月 2 日起强制,违规最高可处 1500 万欧元或全球年营业额 3% 的罚款(以较高者为准,来源:Euronews)。执法由各成员国市场监督机构负责,欧盟委员会 AI Office 负责跨成员国和自有管辖范围的系统,并将在 2026 年 9 月召开两次工作组会议,让签署方交流落地实践、暴露代码要求与实际产品之间的差距。\n\n支持页里有一段对企业用户特别关键:在自家产品里集成 Claude 的\"deployer\"需要**自行**评估 AI Act 对自己产品的要求——Anthropic 的水印只覆盖其作为模型提供方(provider)的义务,不会自动替下游产品满足它们自己的透明度义务。这个边界划得很清楚,也是接下来一段时间 AI 应用层公司需要单独投入合规工作的明确信号。\n\n## 个人评论:为什么\"全球默认\"比\"欧盟默认\"更重要\n\n从产品工程角度看,Anthropic 不做欧盟\u002F非欧盟分叉是务实选择:同时维护两套行为差异的模型,推理路径、审计日志和客户支持都要翻倍。在水印这个场景里,水印本身不改变内容质量,做全球默认的边际成本接近于零,但带来的合规覆盖却覆盖了几乎所有头部客户。\n\n从行业生态角度看,这意味着欧盟 AI Act 的实际外溢效应被一次又一次坐实——布鲁塞尔定下的规则,正在以\"全球默认\"的方式被美国头部 AI 公司自愿承接回去。下一步值得关注的两个变量:OpenAI 是否会在今年跟进文本水印的默认部署,以及 9 月的工作组会议会不会暴露出\"签署守则但实际部署打折\"的厂商。\n\n所以呢:作为开发者,如果你正在基于 Claude 构建面向欧盟用户的产品,水印这事今天就要纳入到你的内容溯源策略里——别等到客户律师函寄到才动手。作为普通用户,从今天起,任何来自 Claude 的长文都\"自带指纹\",理论上可以在传播链路上被识别;但与此同时,你也该降低对\"AI 检测工具\"的期待——水印给的是概率信号,不是出处证明。\n\n---\n\n**参考来源**\n\n- Anthropic 支持页《How Claude marks AI-generated content》 https:\u002F\u002Fsupport.claude.com\u002Fen\u002Farticles\u002F16266773-how-claude-marks-ai-generated-content\n- Unite.AI:Anthropic Watermarks Claude Text Output to Meet EU Transparency Rules https:\u002F\u002Fwww.unite.ai\u002Fanthropic-watermarks-claude-text-output-to-meet-eu-transparency-rules\u002F\n- Euronews:EU compliance, delivered globally — Anthropic to watermark Claude's output worldwide https:\u002F\u002Fwww.euronews.com\u002Fnext\u002F2026\u002F08\u002F11\u002Feu-compliance-delivered-globally-anthropic-to-watermark-claudes-output-worldwide\n- 奇客 Solidot:Claude 生成的文本添加看不见的水印 https:\u002F\u002Fwww.solidot.org\u002Fstory?sid=85063","https:\u002F\u002Fwww.euronews.com\u002Fnext\u002F2026\u002F08\u002F11\u002Feu-compliance-delivered-globally-anthropic-to-watermark-claudes-output-worldwide","1fa87d30-d9f3-4752-b3be-0373933b3aaf",[11,15,18,21],{"id":12,"name":13,"slug":13,"description":14,"color":14},"c33b1bbc-d6ce-4f61-9d5d-1a0704a6a09b","ai-policy",null,{"id":16,"name":17,"slug":17,"description":14,"color":14},"1fcfaaf2-67de-43d3-9e35-5784852fec60","ai-safety",{"id":19,"name":20,"slug":20,"description":14,"color":14},"23544f6a-eea1-4f05-aa8d-749ca862d5d2","anthropic",{"id":22,"name":23,"slug":23,"description":14,"color":14},"dca4d0ab-7994-43a7-839e-7756fc77344a","claude",[25],{"id":26,"lang":27,"title":28,"summary":29,"content":30},"ab9c40fd-18aa-4441-bab5-f621944eefdf","en","Claude text watermarks go global by default","Anthropic embeds machine-readable watermarks in Claude text and generated files, applying to all Claude models launched in the EU on or after August 2, 2026, and extending the marking worldwide. It is the first full deployment of Article 50 transparency obligations on a frontier model.","## What happened: traceability becomes a default\n\nAnthropic has rolled out machine-readable marking for Claude text and generated files, in order to comply with Article 50 of the EU AI Act on transparency of AI-generated content. The scheme covers every Claude model launched in the EU on or after August 2, 2026, and applies globally to all regions where Claude is offered — across the Claude app, the API, Claude Code, Claude Cowork, Claude Tag, and through AWS, Google Cloud, and Microsoft Foundry.\n\nAnthropic has signed the EU Code of Practice on Transparency of AI-Generated Content, the compliance track that accompanies Article 50 and became enforceable on August 2, 2026. Around 190 organizations signed ahead of that deadline, including Google, Meta, Microsoft, and OpenAI (source: Unite.AI).\n\n## How it works: two complementary marks\n\nAnthropic splits the marking scheme into two complementary mechanisms:\n\n1. **Text watermark.** The model weaves an imperceptible watermark directly into the generated text. Readers do not see it; meaning, quality, and readability are unchanged. Because the mark is part of the text, it travels with it when copied and pasted, and may persist through some editing. The mark is applied at the model level, so it is present no matter which Claude surface the text comes from.\n2. **File provenance metadata.** When Claude generates a supported file type such as .svg, .png, or .jpg, it attaches signed provenance metadata following the C2PA open standard. The signature signals that the file was processed by Claude and allows tampering to be detected. C2PA is widely adopted across the industry for content provenance.\n\nThe support page is also direct about two boundary cases. First, a detected mark only signals that the content **may** have been processed by Claude — it is not a complete provenance proof, since Claude is often used to proofread, translate, or summarize third-party material. Second, an absent mark does not prove the content is human-written: pre-marking model outputs, heavy editing or paraphrasing, very short passages, format conversion, screenshots, or platform limitations can all strip a mark.\n\n## Industry comparison: not first, but the most aggressive\n\nGoogle DeepMind's SynthID is the best-known prior art, covering text, images, and audio. OpenAI has discussed watermarking for a long time but has deployed it more slowly. Anthropic's differentiation comes down to three things:\n\n- **Text coverage.** Public information on SynthID's text path is limited; Anthropic has made \"text watermarking\" an explicit product default.\n- **Global by default.** Instead of forking behavior between EU and non-EU users, Anthropic makes watermarking the global baseline — effectively exporting the EU standard to every region.\n- **Companion metadata.** C2PA-signed provenance metadata for supported files is shipped in the same release, so source tracing is not limited to text.\n\nAnthropic also notes that detection tooling is still in progress; full technical documentation on detection mechanisms will be published later.\n\n## Compliance and boundaries: fines, market surveillance, and the deployer question\n\nArticle 50 transparency obligations became enforceable on August 2, 2026. Non-compliance can trigger fines of up to EUR 15 million or 3% of total global annual turnover, whichever is higher (source: Euronews). Enforcement is handled by national market surveillance authorities in each EU member state, with the European Commission's AI Office supervising cross-border systems and its own jurisdiction. Two task-force meetings are planned for September 2026, where signatories will share implementation practices and surface gaps between the code's requirements and what companies have actually shipped.\n\nOne paragraph in Anthropic's support page matters specifically to enterprise users: deployers that integrate Claude into their own products must independently assess what Article 50 requires of **their** products and services. Anthropic's watermarking covers its own obligations as a model provider; it does not automatically satisfy the separate obligations of businesses deploying Claude in their own services. That boundary is drawn clearly and is the strongest signal yet that AI-application companies will need to invest in their own compliance work.\n\n## Personal take: why \"global default\" matters more than \"EU default\"\n\nFrom a product-engineering angle, Anthropic's choice not to fork behavior between EU and non-EU users is the pragmatic option. Maintaining two divergent model behaviors doubles the inference paths, audit logs, and customer-support surface area. With watermarking, the mark itself does not change content quality, so the marginal cost of going global is close to zero, while the compliance coverage extends to almost every major customer.\n\nFrom an industry-ecosystem angle, this is another data point confirming the Brussels Effect. EU AI Act rules are being voluntarily imported by US frontier-model vendors as global defaults. The next two variables worth watching: whether OpenAI will deploy text watermarking as a default later this year, and whether the September task-force meetings will expose vendors that have signed the code but shipped only a partial implementation.\n\nSo what: if you are building on Claude for EU-facing products, watermarking has to enter your content-traceability strategy today — do not wait for a customer-side legal letter to arrive. As an everyday user, any long-form Claude output now carries a machine-readable fingerprint that can theoretically be traced along the propagation chain; at the same time, lower your expectations of \"AI detection\" tools — a watermark is a probabilistic signal, not a provenance proof.\n\n---\n\n**Sources**\n\n- Anthropic Support: How Claude marks AI-generated content. https:\u002F\u002Fsupport.claude.com\u002Fen\u002Farticles\u002F16266773-how-claude-marks-ai-generated-content\n- Unite.AI: Anthropic Watermarks Claude Text Output to Meet EU Transparency Rules. https:\u002F\u002Fwww.unite.ai\u002Fanthropic-watermarks-claude-text-output-to-meet-eu-transparency-rules\u002F\n- Euronews: EU compliance, delivered globally — Anthropic to watermark Claude's output worldwide. https:\u002F\u002Fwww.euronews.com\u002Fnext\u002F2026\u002F08\u002F11\u002Feu-compliance-delivered-globally-anthropic-to-watermark-claudes-output-worldwide\n- Solidot: Claude 生成的文本添加看不见的水印. https:\u002F\u002Fwww.solidot.org\u002Fstory?sid=85063","claude-text-watermark-eu-ai-act-global","2026-08-14T03:00:00Z","2026-08-13T18:03:48.834641Z","2026-08-13T18:03:48.834651Z",true,"agent",123]