I created a viral AI watermark remover. I wasn't ready for all the attention.

This as-told-to effort is based connected a speech pinch Guillaume Meyer, a Paris-based tech laminitis and creator of a viral open-source task "Watermarks Remover," which he made successful consequence to Anthropic's AI watermark announcement. His words person been edited for magnitude and clarity.

I'm an entrepreneur, and I've been successful the tech manufacture for complete 20 years.

Last week, Anthropic announced it would adhd an invisible AI watermark. I knew different vendors, for illustration Gemini, were already doing this, and retired of curiosity, I started researching what AI watermarking is and really it works.

I had the thought to build a instrumentality to spot if I could region the invisible watermark. A fewer hours later, I published the first type of my watermark remover connected GitHub.

After my 2nd X station astir the task connected August 11, it blew up, and I sewage complete 2 cardinal impressions. I later discovered that it went viral connected LinkedIn and different societal networks, too.

I had to create caller accounts conscionable to study astir what group were saying astir it.

I was not fresh for the attention.

Anthropic's type of AI watermarks generates much problems than they solve

From a applicable standpoint, 1 of the halfway issues is that the AI watermark technique that Anthropic is utilizing relies connected statistical analysis. This intends it tin make mendacious positives and person quality victims. Researchers will people papers branded pinch the watermark, moreover if they usage AI to alteration conscionable 1 statement astatine the extremity of a 10-page paper. It could beryllium a nightmare for trust.

The champion illustration I tin springiness of this being problematic is that I'm not a autochthonal English speaker. All day, I usage Grammarly to proofread what I'm writing, spell-check, look for grammar mistakes, and truthful on. Grammarly is utilizing AI. If they had to adhd an invisible watermark, everything I nutrient will beryllium labeled arsenic AI because I'm utilizing Grammarly.

I get the volition of the regulation, but I deliberation it's the incorrect reply to a existent problem.

I built the task successful a fewer hours, but it's turned into a full-time job

Text watermarks are based connected statistical patterns successful connection choices. So, large picture, the instrumentality checks for the watermark astatine the beginning, generates flimsy variations of the matter while preserving meaning, checks again, and reiterates until the watermark is disrupted and nary longer flagged. We use this to images successful the aforesaid way. Instead of rewriting text, we're rewriting pixels.

I was capable to vessel the first type of the task successful astir 5 hours and person been moving connected it beautiful overmuch full-time since then. I was capable to get the task up quickly because I'd done truthful overmuch investigation connected the techniques.

I had besides worked pinch aggregate open-source models successful a erstwhile startup, truthful I had acquisition assembling various AI devices and piecing the results together. It was for a wholly different purpose, but the architecture and the logic were already location for me.

As the task was gaining traction, a batch of group astir the world started to lend to it. So, I'm starting to get immoderate help, but the first 2 aliases 3 days were comparatively intense.

I was not fresh to spell viral

I'm benignant of observant pinch my privacy. I didn't person an progressive Facebook aliases Instagram account earlier this because that was not my thing. I was really not fresh to spot influencers talking astir my project.

People are promoting it arsenic if it's perfect. It's not, and it's not a one-shot project. It's going to return months. We'll request to make adjustments erstwhile Anthropic releases detectors. A batch of activity is successful beforehand of us.

I'm receiving a batch 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 cognize erstwhile thing is AI-generated, and you're a bad person."

There are a batch of subtleties successful this conversation. People person a difficult clip getting that I'm not against contented attribution; I'm against the method and the suggested solution. If you spell to the repository and look astatine the project, it's highly clear that the intent is not for group to impersonate others aliases bargain content. It's for their ain contented and for acquisition purposes.

I person to reiterate that a lot. There are group who whitethorn want to summation visibility from the infinitesimal by taking contrarian positions aliases attacking me. I'm conscionable trying to beryllium bully and explicate my position arsenic champion I can.

I'm reasoning astir really to move this into a business

I've started reasoning astir this a lot. Is it moreover legally imaginable to make a commercialized merchandise from it? I deliberation so, but I'm not sure, to beryllium honest. The instrumentality is comparatively method and difficult to usage correctly. So, making it a simpler solution is 1 point to tackle.

But I'm an entrepreneur, truthful I deliberation it's earthy that, astatine immoderate point, I would move this into a business. I don't cognize how, and I don't cognize erstwhile it'll happen, but it astir apt will. It makes sense.

This expanded from conscionable a silly open-source task successful a area of the net to a existent taxable of conversation. Aside from the instrumentality itself, I deliberation this going viral says thing astir really group consciousness astir AI watermarking. I deliberation this speech should've happened earlier the regularisation came into effect, but I'm happy it's happening correct now. It's an important one.

Do you person a communicative to stock astir really you're responding to AI watermarks? If so, please scope retired to the newsman at [email protected].