Hiding AI usage is about to get a lot more difficult

More AI stuff? Yes, more AI stuff.  

The EU is requiring “watermarking” to signal content that was made with AI. Legislators are demanding that internet users be informed whether they are seeing content that is AI generated. On Monday 10th August, Claude announced that all future models would support watermarking to comply with the new EU rules.    

The watermarking will appear everywhere, not just in Europe. And it will appear in the models themselves. 

What is “watermarking”

There are two main ways that watermarking will be applied when someone creates something with AI: text and image watermarking, and file labelling. The new legislation does not dictate what method should be used. 

Claude will be watermarking text itself, “Generated text will carry embedded watermarks, and generated files will include digitally signed provenance metadata where supported.” The way AI text is generated is one word at a time. A complicated statistical analysis results in the next word being generated. Sometimes multiple words are equally applicable in the statistical range and when this happens the Claude models will subtly lean towards certain word choices. That doesn’t necessarily mean that we’ll see easily detected trigger words, but that, taken together over many sentences, specific word choice will thread together like a map and can be detected by software that has the key to the map. 

OpenAI has agreed that “people should have better context about the content they see online, including whether it was created or edited by AI.” and AI generated images will have an invisible overlay. Gemini already does this.

What will happen on platform?

Watermarking AI generated content is step one and visualising it to internet users will be step two. AI companies will have to work with platforms to visualise these watermarks to internet users. It’s likely that when someone adds content to a platform, the platform will have software that can detect the watermark and this will then flag the post with a notification to tell users that this is likely to be AI. Platforms are already well aware of the rise of AI, LinkedIn recently adding an option for users to flag a post that “seems like AI slop.” and it’s expected that there will be punitive measures on any post flagged in this way.

Even if the action from the platforms are minimal, the appetite of users for AI generated anything is weak. There is evidence to show that as soon as people know something is AI generated, engagement drops significantly. Users value things that they believe people have put time and effort into but when something is labelled as created with AI, users believe that the creator has not put the time and effort in that they should.

So what?

This does shift the whole conversation away from AI generated content that risks being deceptive, to AI generated content in general. As it stands it is a blunt instrument and there are already users out there who have had posts flagged for things like using Canva’s magic background remover. This seems unnecessarily punitive, especially considering background removers have been around far longer than wholesale AI generated imagery. It may also be that this happens when you use Claude, or something similar, for proofing. 

I suspect that the approach will be refined, both by the platforms and the AI organisations. I could see, for example, a two tier system that overlays a different watermark for “tinkering” vs wholesale creation. This though is still likely to have a negative impact on engagement and authority in the user’s mind.

Of course, there are already people out there looking for workarounds. Like this video shows. But if you are going to go to these lengths, I suggest you just write the thing yourself. It will take less time and be better for your brain. It’s also important to note that the watermarking detection software for written copy does not exist yet, but that any post you write with Claude from here on out will have a watermark and could be subject to AI detection at a later date.

Final thoughts

My advice remains the same (how refreshing), as it appeared in my last two posts, linked here and here. If you can possibly help it, don’t use AI in external facing content creation. Choose the harder route. If you are using AI in your marketing or communications, do it sparingly, specifically and only when you have a fully developed understanding of what you want from it and why. Just because it is a low price option does not mean it is a low cost option.