Anthropic宣布为Claude Mythos 5、Claude Opus 4.7、Claude Opus 4.6、Claude Sonnet 4.6四款模型加入水印机制，其中两款旧模型的水印存在一个前提条件。只有当开发者把三个采样参数（sampling parameters，控制模型生成文本时词语选择的随机性和多样性的设置）维持在默认值，水印才会生效。调整这些参数的开发者，无论出于什么目的，都可能让输出变成没有水印的文本。来源文本没有列出这三个参数的具体名称，也没有说明默认值是什么，开发者需要查阅官方文档确认。
水印不改变生成内容的可读性。Anthropic将其描述为措辞选择中一种不可察觉的模式，不会影响语义、质量或可读性。它也不携带用户身份信息，官方说法是水印不包含用户、其所属组织或其与Claude对话的任何信息。这套机制标记的是文本由AI生成这件事，不涉及追踪具体生成者。
Anthropic在文档里提醒，如果客户的合规工作依赖这个水印，就需要保持采样参数的默认设置不变。新一点的Mythos 5和Opus 4.7是否存在同样的限制，来源文本未提及，这里不做推断。
Opus 4.6在Anthropic内部评价里仍是公司最好的写作模型，排名高于Opus 5.5。这款被认可为写作能力最强的模型，恰好是水印存在漏洞的两款之一。来源没有解释这两件事之间是否存在关联，是否有开发者因为看重Opus 4.6的写作能力而倾向调整参数、进而让水印失效，属于开放问题，目前没有材料支持任何结论。
适用范围和时间节点
Anthropic把这次为旧模型补装水印的动作，和欧盟《AI法案》（EU AI Act）当年12月的截止日期联系在一起，称这是促使公司现在补上水印的原因。这解释了为什么只有Mythos 5、Opus 4.7、Opus 4.6、Sonnet 4.6这四款模型需要专门公告，其他时间发布的模型不在这次讨论范围内。
这套水印支持的托管平台包括亚马逊云科技（Amazon Web Services）、谷歌云（Google Cloud）和微软Foundry，客户在这些平台调用对应模型都会带有水印，且无法关闭。
技术来源和同行方案差异
Anthropic采用的是谷歌DeepMind旗下SynthID文本水印方法的一个变体。SynthID已经被谷歌DeepMind开源，并内置进了Gemini系列模型。水印技术路径不是Anthropic独立研发，而是在现有开源方案上做了适配。
textGrain默认只对欧盟地区的ChatGPT和Codex用户生效，API客户不论所在地区都可以自行选择开启或关闭，出厂状态是关闭的。Anthropic的水印不可关闭，OpenAI的API水印默认关闭且可调，两家公司在是否给客户控制权这一点上走的是相反路线。
来源文本没有提供第三方机构或学界对这套水印机制有效性的评价，也没有提及用户或开发者社群的实际反应，这部分材料目前缺失。两款旧模型水印依赖参数默认值的设计，会不会被更广泛地利用来绕开水印检测，来源同样没有回答。
原文地址：tools.wflynn.cn/aitools/tre…
关注公众号「YourTools」，每天推送 GitHub 趋势与开源工具速报：
本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
Anthropic has been declared Claude Mythos 5, Claude Opus 4.7, Claude Opus 4.6, Claude Sonet 4.6, four models have been added to the watermarking mechanism, two of which have a precondition. The watermark will only be effective if the developer maintains the three sampling parameters (sampling parameters, the randomity and diversity of word selection when the control model generates the text) at the default value. The developers that adjust these parameters may, for whatever purpose, turn the output into a text without watermarks. The source text does not specify the names of these three parameters, nor does it state what the default value is, and the developers need to consult the official document for confirmation.
The watermark does not change the readability of the content generated. Anthropic describes it as an undetectable model in the choice of wording that does not affect semantics, quality or readability. It also does not carry user identification information, the official assertion being that the watermark does not contain any information from the user, his organization or his dialogue with Claude. The mechanism is marked by the fact that the text is generated by AI and does not involve tracking the specific generator.
Anthropic cautions in the document that if client compliance depends on this watermark, the default settings for sampling parameters need to be maintained. Whether the new point of Mythos 5 and Opus 4.7 have the same limitations, which are not mentioned in the source text and are not inferred here.
Opus 4.6 remained the best writing model for the company in the internal evaluation of Anthropic, ranking higher than Opus 5.5. This is recognized as the most written model, and it happens to be one of the two sections where the watermarks are flawed. The source does not explain whether there is a link between the two matters, and whether the developers prefer to adjust the parameters to render the watermark invalid because they value Opus 4.6's writing ability, is open-ended and no material currently supports any conclusion.
Scope of application and time nodes
Anthropic linked the operation of refilling the watermark for the old model to the deadline of December of the same year for the EU AI Act, claiming that it was the reason why the company was now refilling it. This explains why only the four models Mythos 5, Opus 4.7, Opus 4.6 and Sonnet 4.6 require a special announcement, and the model of publish at other times is outside the scope of this discussion.
The hosting platforms supported by the watermarks include Amazon Web Services, Google Cloud and Microsoft Foundation, where customers call matching models with watermarks and cannot be shut down.
Differences between technology sources and peer programmes
Anthropic uses a variant of the SynthID text watermarking method in Google DeepMind. SynthID has been added to the DeepMind open source of Google and incorporated into the Gemini series model. The watermark technology path was not developed independently by Anthropic, but adapted to existing open source programmes.
TextGrain defaults are only effective for ChatGPT and Codex users in the EU area, and API clients are free to open or close regardless of their location, and exit status is closed. The watermarks of Anthropic cannot be closed, the API seals of Open AI are by default closed and adjustable, and the opposite route is followed by both companies in terms of whether or not they give control to their customers.
The source text does not provide an evaluation of the effectiveness of this watermarking mechanism by a third-party institution or academic community, nor does it mention the actual response of the user or developer community, which is currently missing. The two old models were designed to rely on the default values of the parameters and would not be more widely used to circumvent the watermark tests, the same source did not answer.
Original address:tools.wflynn.cn/aitools/tre...
Following the WeChat Official Account "YourTools", GitHub's trend and open source tool is sent daily:
Auto-fetched and translated by AI Atlas. All rights belong to the original author.