Volume 2, Issue 8
Karma is a bitch
I find the general professional reaction to Anthropic’s watermarking of text to be really illuminating. They are using a word selection algorithm to leave a traceable signal in the generated text, as well as using C2PA content credentials for images. That’s a good thing. It may not feel like it now, but I believe in the future this kind of content attribution is going to be critical. It is part of the reason I am a contributing member of the C2PA. Now, I don’t usually get whiney, so please don’t take this entirely seriously when I say…
Personally, I can’t escape the feeling that so many of those offended souls (I’ll call them Claude posers) are just upset because they’ve been using Claude while throwing shade at ChatGPT and Gemini. Now they have to figure out what to do. It doesn’t feel to me like this is going to lead to an increase in human authorship, though.
There’s a very memorable cartoon by Dave Berg in MAD Magazine that I remember from my youth. A guy has a fast-food hack: he always orders his burgers without ketchup so they have to make it to order. Then you see the cooks in the back grumbling while they open up the burger and scrape off the ketchup. That’s what all these Claude posers are doing. Scraping the ketchup off the burger so nobody will know.
I am fairly sure that some of the very same people that don’t have a problem with AI when it helps their writing are indignant that there might be AI in their music. Writers, photographers, musicians, and other creatives have been dealing with this since generative AI became feasible at scale. Are the folks who started complaining now just getting their first taste of what the market has been doing? Really?
Honestly, I shouldn’t even use Claude because they followed a very similar path as Suno and Udio, just in the domain of books versus albums. But I wouldn’t stop using them because they watermark text. People legitimately want to know about the provenance of the content they are consuming, so at least Anthropic is taking that step.
Now here’s a somewhat cynical thought exercise for you. How many of the Claude posers are going to look for a tool to remove the Anthropic watermark versus just rediting or rewriting the piece in their own voice?
And, right on cue…
Headline: Growing industry for Claude watermark removal tools
I mean, we could have predicted this, right?
We live in a “me” economy, clearly. Also, Claude is not the only thing that is capable of watermarking a text file; there are other methods, including C2PA. So you might be right to question the effectiveness of the wave of watermark removers that are popping up everywhere. But here they are. And here. Or even here, at the appropriately named claudewatermark.com. Or you can roll your own with this Github repo, currently approaching 20,000 stars.
Or, you can emulate what we do here at Catamount Music. We use Claude to prepare, organize, and vet stories that are hand-curated and have a human reviewing that work and writing the final copy.
And back around it comes…
And, (smacks head with hand) just before the month runs out, is the news that the majors have sued Anthropic because they trained on copyrighted music too, according to the allegations. Interestingly, the alleged theft was of lyrics, published musical compositions, and other representations that were scraped as Anthropic mined large volumes of text-based materials for model training. This is different than Suno and Udio, who use the sound recordings to train their models.
Strung out to dry
Now onto the fun stuff, unless you are D’Addario who had to admit to using generative AI in a product demo in early August after a month of denials. They were just trying to show off their new, extended life strings, you know? But wow, did it backfire.
You are forgiven for maybe not hearing D’Addario got embroiled in this AI scandal in July when they released a demo of their new NYXL high-density strings. The guitar community on YouTube wasn’t having it and was convinced that AI—Suno in particular—was used. D’Addario released a statement full of mix engineering slop that basically claimed it was overengineered with plugins and AI mastering but that it was not AI-generated. Unfortunately for D’Addario, the community still wasn’t having it. Now here’s the wild part. Rhett Shull, who was not originally convinced it was AI, actually got the original Logic Pro file for the track and analyzed it along with the D’Addario social media posts. (minor cheer within a jeer—D’Addario allowed the file to be analyzed in this way.) Along the way, you get to see long sections of the original performance post as well.
And then ultimately D’Addario is forced to admit that yes, Suno was used in making the video. And honestly, the whole thing feels like adolescent-level theatrics. Nobody involved comes out of this looking good. For D’Addario, that original sloppy denial is an epicly spectacular train wreck.
We knew we had nothing to hide; we just needed the time to understand why our take on the situation was so different than others in the community…
Most of the digital artifacts that people online have taken issue with come from an unusual take on modern production techniques (4 compressors in series on the drums adding high end fuzz, auto-tune on the guitar for pitch accuracy adding aliasing, just to mention a few).
Do go on.
Remember, we are a company whose product can only be used by people. If AI replaces musicians making music, then we’re out of business.
OK, you can stop now.
You know what would have really helped? Content credentials. Just shoot the original performance with a C2PA-enabled camera, and then you’ll have the proof you need. That doesn’t mean the artists didn’t use Suno to generate the underlying composition, but that’s a struggle for another time.
Speaking of C2PA… here’s how you do it
The C2PA released a guidance document on how to identify synthetic content and AI inputs from content credentials. It is cleverly named, “Use of Content Credentials to identify synthetic and non-synthetic content” and it contains three important points:
- Every action is paired with a digitalSourceType that defines what kind of content it is (i.e., is it human, AI, or hybrid)
- There are diverse examples of C2PA manifest language for situations involving video, images, audio, etc.
- It shows how to include data assertions that are the inputs to a generative AI process, like prompts and inference parameters as well as disclosures
OK, so that is my take on it. The abstract actually lists a slightly different set of things, including Regions of Interest, which I don’t find all that interesting. A region of interest is just a part of the asset. For an audio file, it might be a time segment where a sample is used, for example. Obviously a critical concept, but I rarely hear mix engineers talking about tracks in this way, so I chose not to highlight it. The concepts of digitalSourceType and AI-related assertions are far more important, in my humble opinion.
It feels like I’m leaving you with homework, and maybe I am. Throw that document into your favorite AI along with the C2PA specification and ask it to teach you the basics of C2PA. That’s a good fall activity, right? You are allowed to have a pumpkin spice latte every time you study C2PA, just in case you didn’t know.
Leave a Reply