OpenAI 在欧盟为 ChatGPT 和 Codex 生成文本加入 textgrain 水印

ChatGPT is now adding watermarks to all text from ChatGPT and Codex in the EU as part of a Europe-wide drive for transparency in the usage of generative artificial intelligence. Here's how this new 'textgrain' system from OpenAI works. It is similar to other systems that labs such as Anthropic have added to Claude, in that it bakes statistical patterns into the choice of words. Here's some info on how well that pattern survives editing and a planned AI detector tool soon to be released. #chatgpt #openai #ai #aitools #artificialintelligence #Tech #technology #learnontiktok #claude #anthropic #genai

AI 导读

OpenAI 在欧盟为 ChatGPT 和 Codex 生成的所有文本加入名为 textgrain 的不可见水印，通过影响选词概率嵌入统计模式，粘贴到其他应用或重排文本都无法去除。OpenAI 测试显示，替换 10% 的词后检测率从 92% 降到 66%，替换四分之一后降到 17%。配套的 AI 检测工具只向获批研究者和专家机构开放申请，不面向公众。

打开原文tiktok.com

选题角度

视频讲清了 textgrain 水印的选词机制和替换词后的检测率变化，适合做原理拆解加编辑对抗演示。

互动数据

浏览: 55万
点赞: 2.5万
评论: 449
分享: 3282
收藏: 7679

约为这个来源平时的 5.8 倍

数字来自平台，为最近一次抓取时的数据。

正文 · 原文

以下是视频口播的自动转写，可能有错字。

ChatGPT is now watermarking all the text that it writes in ChatGPT and codex for all users in the eu. And you can't see this watermark.

And pasting the text into other apps, rearranging the text does not remove it. So let me explain how it works and how to get around it.

So, this is not like previous watermarks or signs of ai that we've talked about. This is a new system from OpenAI called text grain, and the watermark is built into ChatGPT's choice of words.

So, you might know that when ai writes, it's choosing the next word from a few different options. So, maybe great, brilliant, and excellent would all work in a sentence.

But this watermarking system has this secret key in the background that influences the probability that certain kinds of words get chosen.

And only the ai detector tool issued by OpenAI has this key and can see these patterns. Now in a very short piece of text that pattern is not detectable but in writing of any length those repeated patterns of word choice are something that their tool will be able to spot with almost 百分之 1 0 0 certainty so rewriting and rearranging does not really remove those patterns but it can weaken it in OpenAI is test replacing 百分之 1 0 of the words with other words with the same meaning.

Dropped detection rates from 92% to 66%. If you replaced a quarter of the words in a piece of writing, the detection rate drops to just 17%.

So no, there's not a one-click way around this. You are just going to have to rewrite large parts of whatever ai writes for you.

And by the way, this watermark is in all the other ais now. Trust me, in the long run, that is a good thing for you. Writing is a form of thinking.

People who write more are smarter because it activates one of the most useful parts of your brain. Now, of course, it's absolutely fine for ai to give you a head start, to help you with the research, help you get off the blank page.

I do that all the time. I don't think this watermark makes any of that a problem. And anyway, what about this detector tool?

Well, ai are opening applications for people who want to use it, but it will just be approved researchers and expert organizations.

So it's not gonna be a public tool. If you want a tool to detect ai writing, you just use something like the free copy leaks one, or maybe pangram, something like that.