Anthropic 从 11 月 3 日起为更多 Claude 模型生成的文本加水印
Claude starts watermarking the text it writes on Tuesday Anthropic announced this back in August, and its newest models have had it since launch. From November 3, it's coming to more Claude models. The text will carry an invisible signal hidden in its word choices, which means your Al writing can be detected, no matter how many times you copy and paste it. You can't see it, and only Anthropic's detector can read it. #claude #anthropic #ai #aiwatermark #chatgpt
Anthropic 从 11 月 3 日起为更多 Claude 模型生成的文本加入不可见水印，该水印藏在选词中，复制粘贴后仍可被其检测器识别，只有 Anthropic 的密钥能读取。该做法基于 Google 2024 年发布的方法，由密钥决定同义选词，主要面向监管机构、研究者、记者和教育机构开放。Anthropic 称因暂无法按地区限制，水印将面向全球用户；ChatGPT 的水印则仅限欧盟用户。
原文讲清了水印藏在选词里的原理和复制粘贴仍可检测的特点，适合做一期原理科普加争议评论。
粉丝 1.8万
- 浏览
- 17万
- 点赞
- 1万
- 评论
- 47
- 分享
- 1963
- 收藏
- 1739
数字来自平台，为最近一次抓取时的数据。
以下是视频口播的自动转写，可能有错字。
Claude is now putting invisible watermarks in all the text that it writes and generates for you, so that it can be picked up by its own ai detection software.
So let's chat about what is actually happening and how any of this is possible. Anthropic announced that from november 4th, it will start watermarking text from even more of its models.
And the catch is, it doesn't matter if you copy and paste that text into a brand new document. It doesn't matter if you type out everything in that paragraph, that ai watermark will still exist, unless you substantially rewrite the content.
As you may already know, ai images already carry that label in its metadata, but now it's going to be applied to written content as well.
And the most interesting part about all of this is actually how you watermark text in the first place. Now, anthropic's way of doing this is based on a google method that it built and published back in 2024.
Basically, when claude writes, it has a few equally good options to choose from. For example, if the sentence is the sky is, it could choose from gray or overcast.
And normally, it just picks one at random. Now, a secret key actually. Makes that choice. You would never notice, but anyone with the key can check if the word choices match.
The watermark can only be read by anthropic's key, and it is a private tool for groups like regulators, researchers, journalists, and educational organizations.
The reason why this is happening right now is because of the eu's new ai rules have made a commitment to flagging ai generated content.
ChatGPT is rolling out its own watermark too, but it's limited to users in the eu, and anthropic has said that because it can't reliably limit by region yet, it's applying to all users worldwide.
So the internet is currently split by people calling this a snitch layer, and others really welcoming the change. So, where do you guys sit?
来源：TikTok · 搜「chatgpt」 · tiktok.com