ChatGPT Watermarks Just Dropped!

Julian Goldie (0:00)

ChatGPT watermarks just dropped, and if you create content with AI, you need to see this. OpenEye is now putting invisible watermarks inside ChatGPT content. And in this video, I want to show you exactly what it means for you, whether your content is safe, and the one simple trick that makes a watermark almost impossible to detect. And there's also one thing you can do to your content that actually drops detection massively. I'll show you that later in this video, so that by the end of this, you'll know exactly how to keep creative content with AI, with total peace of mind, and you'll be way ahead of everyone who's panicking right now. Stick with me to the end, because the last part is the most important. Let's get into it. So, OpenAI are releasing watermarks into their content, similar to what Claude did recently. Google Gemini already do this, and this is down to new EU regulations, as you can see right here.

Now, they do talk about the limitations of current text watermarking technology, but basically, they're saying now that in the EU, they're going to start watermarking illegible text from ChatGPT and Codex over the coming weeks to comply with the EU AI Act. Now, it's pretty interesting. So, customers using the API can turn text watermarking for select models right now, but it essentially means that watermarks are going to be identifiable by AI in the future. Now, what's interesting is that they're more focused on giving people a choice. When I look at this breakdown, so, for example, if you have a look at the announcement from Anthropic, when they were releasing watermarks, there was no real choice involved. It's just like you have to start using watermarks. It's going to be automatically set up inside the new models. With OpenAI, they're switching it around. So, you can basically be clear about whether you want to use watermarking or not.

And a text watermark, if you're wondering what it is, is basically answering one question, which is, was this likely generated by an OpenAI model? Now, this is a bit of an issue, because some people, they don't want their content to be identifiable. Some people don't want to be transparent about using AI. And then other people, they just don't want, for example, like search engines to see that the content is AI generated and then potentially get marked down. Some people worry about that for SEO, which I don't particularly worry about, but that's another point. So what happens here and the way that works is that it has an invisible statistical signal in text as its generator. And then you can see here a detector can look for the signal and it doesn't add visible marks or special characters. You can't actually see if the watermark is added or not to the human eye. It's like a statistical number or a pattern inside the content itself. And then also something important to note here is that it doesn't identify who the person is or what company it is, what account or conversation or prompt it comes from. What's also interesting is that it doesn't slow down the models either, which I guess are probably pretty happy about it. Obviously, there's a lot of limitations as well. So you can see here the open eye say watermarks have limits, they're often undetectable, especially in short passages. This is something the anthropic also mentioned, which is if you have a really short post, it's going to be pretty hard for any AI detection tool to be able to identify that, because there's not enough text to see a clear pattern and signal throughout the text. So if you wanted to have content without watermark, number one, you could use the API, it sounds like you have the option. And then number two, if you focus on shorter stuff, so for example, like you look at posts on X, they're pretty short, those sort of posts wouldn't be identifiable because there's not enough characters or text inside the post to be identifiable as AI generated.

Now also, they said that for now, this is not open for detection. So what that means essentially is not everyone's going to get access to the watermark detector. They said only approved researchers will have access to our detector so they can help us evaluate and improve the technology. This is a common theme, it's just like everything is being gated before it drops, which means that not everyone gets access. They're testing text, walk to marking as an ongoing area of research. So you can see a full breakdown here, actually created a full blog context on it. They've called it Content Provenance, which basically helps understand where content came from, which means like Content Provenance basically, is it detectable or not? And this comes down to the EU text provenance rules.