I don’t like the Claude watermark announcement. For images, yes, watermarks are important. “A picture is worth a thousand words” and there’s too much of a chance for misinformation. For text, it’s more complicated, as essays will often be 95% human, with AI-assisted edits.

@manton totally agree. I, for example, write in Italian, which is my language. And then ask Claude to translate for me. That would be marked as AI generated. Which actually is not in terms of content and idea.

@teofranchi That’s a great example.

@manton @teofranchi this feels like Microsoft Word or Google Translate spying on you, claiming credit for touching your words.

@manton I’m interested to see how this is implemented. In higher ed, there is legitimate concern about researchers using AI to generate journal articles and overwhelming the peer review system. I could see watermarking as a deterrent to having an AI generate an entire literature review.

@bobwertz I believe it’s implemented with minor (but not random) changes in which words are generated, creating a pattern over a long run of text. So it probably doesn’t work for short essays, but would for a long journal article. I would feel better about it if it was documented to only be applied to text over X thousand words.

@manton And the text that is all Claude-generated doesn’t need a watermark… the writing style gives it away immediately 😂

@manton need to see how its implemented, but it'll be an accessibility pain to de-strganographize text to then be able to process it as regular UTF-8. That being said, I am a proponent, while also recognizing the difficulties.

@manton It’s becoming more challenging to detect AI-written assignments that my students submit. Often, they are aware enough to add manual instructions to avoid most of the online AI-checkers. I would welcome a watermark if it could be implemented well. I’m glad at least that they’re trying.

@KyleEssary I can understand that perspective too. It’s a hard problem.