This as-told-to essay is based on a conversation with Guillaume Meyer, a Paris-based tech founder and creator of a viral open-source project “Watermarks Remover,” which he made in response to Anthropic’s AI watermark announcement. His words have been edited for length and clarity.
I’m an entrepreneur, and I’ve been in the tech industry for over 20 years.
Last week, Anthropic announced it would add an invisible AI watermark. I knew other vendors, like Gemini, were already doing this, and out of curiosity, I started researching what AI watermarking is and how it works.
I had the idea to build a tool to see if I could remove the invisible watermark. A few hours later, I published the first version of my watermark remover on GitHub.
After my second X post about the project on August 11, it blew up, and I got over 2 million impressions. I later discovered that it went viral on LinkedIn and other social networks, too.
I had to create new accounts just to learn about what people were saying about it.
I was not ready for the attention.
Anthropic’s version of AI watermarks generates more problems than they solve
From a practical standpoint, one of the core issues is that the AI watermark technique that Anthropic is using relies on statistical analysis. This means it can generate false positives and have human victims. Researchers will publish papers labeled with the watermark, even if they use AI to change just one line at the end of a 10-page paper. It could be a nightmare for trust.
The best example I can give of this being problematic is that I’m not a native English speaker. All day, I use Grammarly to proofread what I’m writing, spell-check, look for grammar mistakes, and so on. Grammarly is using AI. If they had to add an invisible watermark, everything I produce will be labeled as AI because I’m using Grammarly.
I get the intention of the regulation, but I think it’s the wrong answer to a real problem.
I built the project in a few hours, but it’s turned into a full-time job
Text watermarks are based on statistical patterns in word choices. So, big picture, the tool checks for the watermark at the beginning, generates slight variations of the text while preserving meaning, checks again, and reiterates until the watermark is disrupted and no longer flagged. We apply this to images in the same way. Instead of rewriting text, we’re rewriting pixels.
I was able to ship the first version of the project in about five hours and have been working on it pretty much full-time since then. I was able to get the project up quickly because I’d done so much research on the techniques.
I had also worked with multiple open-source models in a previous startup, so I had experience assembling various AI tools and piecing the results together. It was for a completely different purpose, but the architecture and the logic were already there for me.
As the project was gaining traction, a lot of people around the world started to contribute to it. So, I’m starting to get some help, but the first two or three days were relatively intense.
I was not ready to go viral
I’m kind of careful with my privacy. I didn’t have an active Facebook or Instagram account before this because that was not my thing. I was really not ready to see influencers talking about my project.
People are promoting it as if it’s perfect. It’s not, and it’s not a one-shot project. It’s going to take months. We’ll need to make adjustments when Anthropic releases detectors. A lot of work is in front of us.
I’m receiving a lot of messages and comments, and 99.9% are very positive. Then there’s 0.1% of the comments that are like, “No, but I want to know when something is AI-generated, and you’re a bad person.”
There are a lot of subtleties in this conversation. People have a hard time getting that I’m not against content attribution; I’m against the technique and the suggested solution. If you go to the repository and look at the project, it’s extremely clear that the intent is not for people to impersonate others or steal content. It’s for their own content and for educational purposes.
I have to reiterate that a lot. There are people who may want to gain visibility from the moment by taking contrarian positions or attacking me. I’m just trying to be nice and explain my position as best I can.
I’m thinking about how to turn this into a business
I’ve started thinking about this a lot. Is it even legally possible to make a commercial product from it? I think so, but I’m not sure, to be honest. The tool is relatively technical and difficult to use correctly. So, making it a simpler solution is one thing to tackle.
But I’m an entrepreneur, so I think it’s natural that, at some point, I would turn this into a business. I don’t know how, and I don’t know when it’ll happen, but it probably will. It makes sense.
This expanded from just a silly open-source project in a corner of the internet to a real topic of conversation. Aside from the tool itself, I think this going viral says something about how people feel about AI watermarking. I think this conversation should’ve happened before the regulation came into effect, but I’m happy it’s happening right now. It’s an important one.
Do you have a story to share about how you’re responding to AI watermarks? If so, please reach out to the reporter at aapplegate@businessinsider.com.
Read the full article here
