Volume 2, Issue 7
Labels, Please.
CONTENT LABEL: I tried to use AI to write this, but honestly, I haven’t had much luck training it to use my particular voice, which I happen to like. Even if you don’t, though, the level of AI that I use for this actually falls below the AI-assisted designation, or as I call it, “lowercase ‘ai,’” because I use it to help with formatting, sourcing, and fact-checking, but not for any of the writing. Nothing generative.
If that makes sense to you, then you’ll get the new rating system proposed by an industry coalition led by the IFPI, RIAA, A2IM, and other (mostly) acronyms. This coalition proposes a voluntary labeling scheme that has two classes of music: lowercase ‘ai’ music and uppercase ‘AI’ music. The classes are related to how much generative AI (genAI) is in the recording, specifically those core elements like vocals, lead instruments, and key instrumental performances. If any of those are genAI, then you’re an uppercase. If a few of the non-core elements were made with genAI, then you’re a lowercase. And if you didn’t use any genAI, then you probably don’t need a label at all. Finally, those are my names; they are actually called AI-generated and ai-assisted. You are now as informed as you are going to be, because that’s all we really know at this point. This is just a recommendation, and it leaves a bit of room for interpretation, shall we say?
Also, if you read back that last paragraph, you’ll notice I never said the word “song” because we’re only talking about the sound recording here, not the underlying musical composition. The word “song” can be interpreted in a lot of ways, so this AI labeling scheme is for the sound recording, not the composition.
And another also with a side of news: we’re only talking about genAI here, not things like AI-assisted mixing and mastering. Like, if you rock out with iZotope and all human performances, you are fine. iZotope, ironically, is also doing fine: early this month, they were acquired by BorisFX and thus exited the affectionately named “Native Instruments bankruptcy bundle.” But we are a bit far afield here, so let me get this back on track.
I’m not sure anybody’s really happy with the current situation in understanding the contents of a sound recording, but generative AI has pushed a lot of new issues forward. There are plenty who are resistant to labels and may want to demonize this coalition, but I think this is a case where we have to assume good intent on behalf of the main players here. IFPA, RIAA, and the others are responding to issues that their stakeholder communities surface. Those communities contain artists of all kinds who are struggling to understand how to defend against both threats and actual infringement of their creative works from genAI music.
Also, a lot of people have been calling for labeling to the point where even I have written about it and vlogged about it multiple times. (See these two shorts and one long.) As an advocate and member of the C2PA, I frequently talk about how content credentials can play a role in fulfilling the information needed to make these determinations by exposing the provenance of the ingredients. Regardless, the proposed labeling scheme, if and when it is adopted, will likely be implemented differently according to the biases of each label or distributor, whether or not they consume information in C2PA manifests.
So stay tuned for more content on this. I have a multi-part series coming out soon talking about these labels, how they can be generated, and how C2PA and DDEX can play a role in providing the underlying information.
More Labels!
DANGER, August 2nd Approacheth…
This coming Sunday marks a pretty big day—the confluence of two pieces of legislation from half a world apart. Both acts, the EU AI Act (Article 50) and the California AI Transparency Act (CAITA / AB 853), take effect this Sunday.
- The EU AI Act requires generative AI service providers to mark their video, image, and audio files to indicate if the content is AI-generated.
- The California AI Transparency Act also mandates these marks but additionally requires visible disclosures and prohibits platforms from removing provenance data.
(California intentionally moved its deadline via AB 853 to line up with Europe on the exact same Sunday.)
Now these acts are really targeting the vendors that provide generative AI models and not individual artists, but there is still a cautionary note here. If you use generative AI either as a main part or an accessory part of a sound recording, you need to be able to disclose it.
If you think about your recording in terms of stems, can you disclose where every stem came from? What level of human- or machine-generated content is present? Is it AI-generated or human-created? Perhaps it is programmed through an automation like a drum machine or an arp-synth. Not all of these will require an AI label, but they are relevant for whoever is going to intake your music for making a determination of AI content. If you really want to get ahead of the game, you should start looking at C2PA and DDEX as methods to provide metadata that your upstream consumer will need.
Sync music is a little different because I have yet to see a sync brief that allows any AI content that would meet the AI-assisted level, so I think it’s a non-issue for now. Anything that meets the brief is likely to not require a label. But it’s still important to start thinking about the provenance of all of the sound assets that you have in a recording and how you’re going to prove what they are and where they came from. Especially if you want to ensure you don’t get an AI label. The best way to prove that you didn’t use AI is to have a method of identifying the source of every track that is consistent, repeatable, and auditable. That builds a huge amount of confidence in your consumers. Whenever you intake something or you output something, you should be thinking about its provenance. Where did it come from? Where is it going? How was it made?
There’s not much of a call to action that most individuals could really follow because there is not a lot of infrastructure for making this declaration. C2PA and DDEX are not super accessible to the average independent musician, so this will probably show up as YAMR (Yet Another Metadata Request).
Policy & Legislation
It was just last month we were talking about Shira Pearlmutter and the US Copyright Office and how that was one really nice beacon of light in a dark landscape where things in the government are working, policy is being made, and artists’ rights are being factored in. But now, this month, I have to report that there are ominous signs on the horizon in the form of H.R.6028. This bill would make Perlmutter’s position, Register of Copyrights, a presidential appointment. The Copyright Office, in case you are wondering, is part of the Library of Congress. Anyway, this is ominous because Trump infamously fired Perlmutter in 2025 right after she issued a policy statement around the use of training content for AI. (The policy generally placing more limits on fair use for AI training than AI companies would like.) Perlmutter then sued the government over her firing, was reinstated through an injunction, and is still there until the case resolves. I think it’s fair to say that there is some fear that if this bill passes, the president could then replace Perlmutter with someone who has an expansive view of fair use, injecting politics into the copyright office and its approach to making policy. However, there’s still a lot we don’t know. For example, we don’t know exactly where the current administration stands on AI training and fair use, except that they generally favor removing regulation and have been very supportive of genAI companies, in particular. In an administration that sees the major stock indices as a primary indicator of economic health, genAI has made some very substantial contributions. It’s also true that the copyright office’s position on fair use could impact the ongoing litigation involving Suno, Anthropic, Udio, and others. So it’s fairly easy to cynically imagine that the administration might appoint a Register of Copyrights who immediately loosens fair-use restrictions for AI training. But the reality is that we have very little insight into what might happen here. It’s not like there was a project 2025 plan for the US copyright office. The bill passed the House and awaits action in the Senate, so we will have to keep a close eye on this one.
SynthID to the Rescue!
You may be aware that a certain T. Swift was married earlier this month. At Madison Square Garden, as one does. The wedding was kept confidential, with guests having to sign NDAs barring, among other things, photos.
So now, as you might expect, there are tons of fake T. Swift wedding photos appearing, and SynthID is turning out to be a strong method for debunking it. Many fake photos were found with a SynthID still recoverable in the image. For the uninitiated, Google’s SynthID is an imperceptible signal injected into the media to identify its origin, but it has been adopted by OpenAI as well. So it looks like lots of folks were using genAI to create fake wedding photos. However, buried in that story just before the end was this tidbit:
Alternatively, OpenAI has a dedicated website that will look for Synth ID or the C2PA Content Credentials protocol, although to be fair, it can only really validate files created by OpenAI tools. It’s not a general-purpose SynthID validator. Which is a shame, but honestly, this kind of validation fragmentation is a growing problem in the provenance field.
TikTok Doubles Down
TikTok has been a part of the C2PA for a few years now and has been implementing C2PA for video. July, however, brings the news that they are now joining the steering committee and continuing to commit to labeling and other methods to bring more transparency to content provenance. So that’s great news for adoption!
SoundPatrol Gets Conformant
Speaking of adoption, congratulations to SoundPatrol for achieving conformance for their validator for music!
Final Thought
It’s 2026 people. We don’t need to put newspapers in proof-of-life photos. You can just use a C2PA-enabled camera (Canon, Sony, Leica, Nikon, and Fujifilm all have them now) and have actual proof, not kabuki-style political theatrics. Just sayin’.
Leave a Reply