Cryptographic Watermark in ChatGPT – How to delete

OpenAI is considering cryptographic watermarking of texts created by the new ChatGPT AI bot.

The company believes that in this way it is possible to prevent, so to speak, the misuse of the bot, in particular, “writing” essays, essays , articles and other materials for schools and colleges with its help .

A cryptographic watermark is, recall, a type of digital watermark that is automatically embedded in an electronic file or signal through cryptographic techniques .

Currently, digital watermarks are successfully used to identify the owner or origin of a digital file or signal in order to protect the protection of digital content from unauthorized use and/or distribution. Photos and videos are marked with a wide variety of “ordinary” watermarks.

However, OpenAI does not consider such protection to be sufficiently effective and speaks of the need to develop a special cryptographic watermark for ChatGPT, which the bot will integrate into its text, and which will be completely invisible to readers, including the “author” of the text, that is, the ChatGPT user.

OpenAI proceeds from the well-known fact that the text created by the so-called “artificial intelligence” is able to recognize without any problems not only a specialist who knows what to look for, but also just an attentive reader who can notice the patterns of constructing phrases characteristic of AI systems and predictable word order. And just this very predictability the company intends to use as an “invisible watermark”.

In a blog post on Search Engine Journal, AI security specialist Scott Aaronson calls this solution “pseudo-random distribution.”

Pseudo-randomness is when words and numbers are arranged in a certain order, which is perceived by the reader as random (but actually is not). This is how ChatGPT can discreetly mark its texts without violating their integrity in any way:

“In short, we want any large text that GPT generates to contain a subtle, secret signal in the form of word choice, indicating that this is the work of GPT.”

HOW TO REMOVE CRYPTOGRAPHIC WATERMARK FROM CHATGPT TEXT?

The solution to this “problem” also seems to be no less simple and obvious: the text written by the ChatGPT bot needs to be “rewritten”, as it were, that is, passed through any other AI text generation protocol.

It’s so stupid, yes. The text itself is unlikely to benefit from this in terms of quality, but the whole procedure can be fully automated and not manually rewritten in an attempt to remove the cryptographic watermark.

However, everything is just beginning. The ChatGPT bot does not yet put any such “secret watermarks” into the texts (sort of), but since the developer is already puzzled by this, it means that in the near future, the counter-creative of users will inevitably rush in the right direction …

 

Leave a Reply

Your email address will not be published. Required fields are marked *