Ask HN: Watermarked ChatGPT? | Hacker News Reader

Ask HN: Watermarked ChatGPT?

4 pointsdakiol7 comments

I've read that OpenAI is working on watermarking ChatGPT outputs so that it's easier to recognize them on the wild. Obviously, this would be a great feature for companies/organizations that want to avoid AI content on their platforms (e.g., Twitter, Medium, etc.) or "plagiarism" (e.g., universities). I, as a consumer would definitely not pay for ChatGPT if it provides watermarked output. It's so obvious:

- If I work on social media and I use ChatGPT to improve my writing/be more creative, I would definitely not use ChatGPT if Twitter suddenly starts to label my tweets as "ai-generated"

- If I'm working on my Phd thesis and I use ChatGPT to better organize my writing (e.g., because perhaps I'm not a native English speaker), but the university I'm submitting my Phd to is paying ChatGPT to analyze my text and see if there is any watermark, well, then I won't use ChatGPT to begin with

- And many more examples where ChatGPT is used to publish content

So, I don't really understand the "feature": if ChatGPT introduces watermarking without the possibility to disable it, wouldn't they be shooting themselves in the foot? Who's going to use watermarked ChatGPT? Very few people.

Obviously nobody thought to use a translation to a language like, I don't know, french, translate it again to maybe German and the translate it again to English?

I bet the translation destroys whatever watermark is introduced in the original message.

But that's a hassle. It may work, but you end up with a version in the target language that may not be as good as the original and needs correction. More often than not, most of the text structure remains intact (so the watermark is not gone).

This does not make any sense. If you're going to release, release it with all. If you're going to change, change all!

With chatGPT, you are getting rid of so much bs writing so why keep the bs writing?

I would just have it generate and then put it side by side with a blank document and type by hand, and edit as I see fit without copy-pasting the text.

That's not how watermark works for ChatGPT. The watermark is the sequence of words. It has nothing to do with the string representation (e.g., invisible characters or whatever). Even if you modify the sequence of words slightly, the watermark is still recognizable. You would have to modify the sequence extensively, but then it would be like writing the thing on your own... so there would be no point on using ChatGPT.

What's your source on this? The watermarking that you're talking about sounds highly speculative. While it may be possible to encode an identifier in a long enough string, I don't think the average (relevent) ChatGPT output has enough entropy to meaningfully identify it from billions of other potential outputs.

I was skeptical as well, but it seems well explained here https://techcrunch.com/2022/12/10/openais-attempts-to-waterm...

Reply on news.ycombinator.com