What do you think about AI tools leaving their own watermarks on AI generated content?
I know many countries (especially in the European Union) are moving toward requiring AI-generated content to be distinguishable from human-generated content. Gemini already does this with images, and now Claude does it with text.
Social platforms are also creating their own identifiers that can detect and label AI-generated content, such as AI labels on X and LinkedIn.
I understand all sides of this:
→ Governments want to prevent misinformation.
→ AI companies can use it as a way to promote their tools.
→ Companies and SoMe platforms can distinguish AI-generated content from human-created content.
→ And AI tools can essentially create their own “trademark” on the content they generate.
To me, there’s something slightly paradoxical about AI models being able to put a kind of “trademark” on something that was itself created by learning from works made by other people tho.
What do you think about labelling AI content in such a way?
Replies
It makes sense from a "stop misinformation" angle, less sure about the "trademark on something trained from other people's work" part, that framing feels a bit off to me too. Curious if it actually reduces misuse or just makes it easier to spot the lazy attempts.
I'm for labelling in principle. I just don't think this version does much.
Text watermarks are way easier to get around than image ones. With images it survives editing. With text it's just a pattern in word choice, so paraphrasing or a quick run through Google Translate wipes it. So it mostly catches people who weren't hiding anything in the first place.
On the trademark bit, marking something isn't really claiming you own it, just saying where it came from. The training question is a real one, it's just a separate fight.
I think it's most useful for AI companies actually. A lot (most?) of today's internet content is generated by LLMs. Using that data for training new generation models creates self-feeding loops and can lead to model collapse.
I find this rule unfair. The AI is a tool. Whatever the AI created was under the user's directions and instructions. Why won't they also label what "so called human-created" with the tools he used to create them. It's ridicules to me.
ClawTeams
The failure mode I worry about isn't the labels that appear, it's what people start reading into their absence.
Once labelling feels normal, "no label" quietly stops meaning "unknown" and starts meaning "human." That's a claim the system never actually made. And since text marks are the easy ones to strip, as Nelson said, the people who get labelled are the ones who weren't hiding anything, while anyone with a reason to hide runs it through a paraphrase and comes out looking more human than the honest user. The burden lands exactly backwards.
The other gap is that a label reports process, but what people actually want to know is whether the thing is true and whether someone stands behind it. Those aren't the same question. Plenty of AI-assisted writing is checked line by line by a person who'll answer for it, and plenty of fully human writing is nonsense nobody verified.
I'd rather see provenance attached to the publisher than to the tool. Who is accountable for this holds up better than what typed it.
The provenance on the publisher point above is the one I would build around. I ship an AI assisted tool, and the trust question users actually ask is never who typed this. It is do you stand behind it. A watermark answers a different question.
Mateusz's self feeding loop problem is the more urgent one for anyone training on scraped data right now. Labels only help there if they survive editing and paraphrase, and text watermarks mostly do not.
What I would actually want is simpler than a watermarking standard: whoever publishes something is on the hook for it being correct, tool used or not. That does not need new technology, just enforcement of a liability that should already exist.
@busmark_w_nika I’d separate disclosure from forensic provenance. A visible “AI-assisted” label is useful context, but it should not be treated as proof: text marks can be removed, false positives can stigmatize human work, and an absent label must remain “unknown,” not “human.” For images or video, a signed creation-and-edit history can provide stronger evidence. For text, I’d prefer a publisher-level disclosure plus revision history and a named person accountable for factual claims. That tells readers both how the content was produced and who stands behind it without pretending a fragile watermark can settle authorship.
I like that we're moving towards AI transparency, but I think there's already room for improvement. One example I saw last week was an instagram post getting flagged because the creator used Canva's background remover tool. Something like that feels quite different than all of the ChatGPT-designed flyers and images floating around. I'm not sure what the right solution is though 🤷🏻♀️