Watermarks are one of the oldest technologies designed to offer proof of something’s origins. Centuries after their first use on banknotes, stamps, and luxury writing papers, the concept of the watermark has found a valuable place in the digital world, where copying is easy and proof of provenance hard. In the book world, we are most familiar with the watermark concept taking the form of DRM (digital rights management), an additional layer applied to a digital file to prevent copying or manipulation.
ALLi News Editor Dan Holloway
It’s not perfect (there are ways to strip out DRM to get around the protections it offers, though they are often quite laborious, meaning pirates will often look for easier targets), and many choose not to use it, but it provides valuable options when rights holders want to maximize the certainty that all copies in circulation are what they claim to be.
AI Creates a New Provenance Problem
But AI has thrown up another context where certainty of provenance is crucial. When platforms, audiences, and politicians alike want to ensure that AI use is transparent, it can be helpful to look at a file and be sure whether its content has been created using AI or not. And by sure, I don’t mean “by checking for ‘tells.’” I mean sureness baked into the content. Unfortunately, although you can make declarations of human authorship, “not by AI” is not something you can bake in this way for now. But watermarks on AI content very much are possible.
And that brings me to the story, which is that Anthropic is set to watermark content produced by its Claude language model. The move is a response to the requirement for transparency in the European Union’s AI Act.
How Claude’s Watermark Will Work
What will this mean in practice? Well, unlike DRM and other digital watermarks, the “watermark” will be more of a “fingerprint.” That is, the language itself, and not metadata, will be the proof. To my inexpert eye, that sounds a lot like baking in “AI tells” that could lead to false accusations abounding yet again. Anthropic has said that is not the case (while obviously not saying exactly what is the case, for obvious reasons). It also insists that while, to stretch a metaphor, the watermark/fingerprint is not indelible, even relatively rigorous editing of the outputs will not remove it.
Unsurprisingly, many users are claiming to be ditching their Anthropic accounts as a result, but it is one more step (and one sure to be followed as others catch up with EU law) toward making the production of AI-generated books a little harder.
Thoughts or further questions on this post or any self-publishing issue?
If you’re an ALLi member, head over to the SelfPubConnect forum for support from our experienced community of indie authors, advisors, and our own ALLi team. Simply create an account (if you haven’t already) to request to join the forum and get going.
Non-members looking for more information can search our extensive archive of blog posts and podcast episodes packed with tips and advice at ALLi's Self-Publishing Advice Center.
