AI watermarks create more problems than they solve
Invisible to whom? The trouble with AI watermarks
Joona Nuutinen As a science fiction thinker, I have conflicting thoughts about the new AI watermark enforcement. There are several reasons why adding these “invisible” markers is a bad idea. And “invisible” is nothing but invisible for computers. Everything that describes the content in any way is very much visible, just not for the human eye, and there’s the problem.
Downsides of watermarking code
- Generated code gets extraneous content that could deteriorate code quality
- If the code declares how it was made, it’s easier to hack, so it’s a security risk
- It creates the incentive to add another step in the workflow to remove watermarks, which complicates the process and makes it even harder to track who’s injecting invisible stuff into your code
Downsides of watermarking text
- AI will claim authorship of text that was only grammar-checked, which is misleading
- Watermarking text will make people cautious about using LLMs even if it would make their work better
- Translating content will get a watermark too, regardless of how it was originally made
There is no reliable way to do this
People might still want to generate text and ideas with LLMs, and edit the contents themselves, which also hides the source. They will still be highly influenced by LLMs.
There’s no reliable way of doing this, and I would argue that incorporating the watermarking, detection and removal systems is an unnecessary hassle and is making us focus on the wrong questions.
Whose idea was it?
The obvious upside of watermarking is that people will know how the content is created, but that’s also highly problematic, as we’ll need to come up with a super sophisticated meta description systems that separate the parts of the content that were LLM-generated and which were original ideas. Often the idea is more important than the final form, because the final form can be changed anyway, while maintaining the core purpose of the idea.
The workaround is worse than the problem
We can get around auto-watermarking by adding a non-Anthropic watermark removal agent into the workflow and orchestrating the process with another AI layer. It will cause even more energy consumption and introduce more points of weakness into the generated content, be it code or poetry. It’s better to just skip the watermark in the first place.
Is this the experience we want?
I run an ads automation system, and the EU AI Act’s transparency rules mean that even ads may soon need to disclose that AI was involved. Picture that level of how-this-was-made everywhere: every ad, every product description, every support reply carrying a small label. It is distracting, and it makes the experience of using these services worse. It pulls our attention to how something was made instead of whether it does what we wanted it to do, and whether it is any good.
As a musician I have conflicting thoughts here. I don’t want everything to be auto-generated, because there is little to no artistry in that. But using an LLM is a means to an end, not the defining characteristic of the end result. There are many levels of AI usage, and the “slop” marker is mostly useful at one of them: when AI is used in a super lazy way. A label that cannot tell the difference between that and a person who used the tools well is not telling us anything worth knowing.
Make great things instead
There are all kinds of creative ways of using LLM tools, that do not necessarily result in absolute slop. Having all content that has an LLM in the process be stigmatized as slop may not have the effect that the watermarking system is designed to do.
We should be focusing on how to make great things, and disincentivising the use of the best tools and most advanced thinking, is counterproductive.
Let’s embrace our brains combining with processing tools that are far beyond human capacity.
PS. Since this is an article about labels, here is the honest one. The intro, both lists of downsides, and every paragraph from “People might still want” down to “far beyond human capacity” came out of an unaugmented brain, and not a word of them was changed on the way to this page. The section “Is this the experience we want?” was my argument, in my notes, expanded into paragraphs by Claude. The sub-headings were suggested by Claude and picked by me. The title went through three rounds between us. The release pipeline was Claude: creating the file, checking the frontmatter, building the site, committing. And this paragraph, the meta-cognition of sorting which parts were AI and which were not, was written by Claude and checked by me. Watermark that.