The SoundCloud terms U-turn doesn't go far enough
They still haven't ruled out training generative AI models on users' music without consent.
Following the revelation last week that SoundCloud had changed their terms to allow them to train AI models on their users’ music, SoundCloud have issued three statements.
First, they said they haven’t trained generative AI models on users’ music to date. Cue people pointing out this said nothing about their future plans.
Next, they said users would be able to opt out of any future gen AI training. Cue people pointing out this was incredibly unfair.
Now, they have said they will change their terms. This is being presented as a win for artists. But, like their previous statements, it still reflects a position that is fundamentally terrible for artists when you read between the lines.
What does their terms update say?
Their terms update removes the clarifying language allowing them to train on people’s music, and adds language that says they won’t train generative AI models that aim to replicate or synthesize your voice, music, or likeness without explicit user consent (emphasis mine).
It would have been very easy to simply say, “We will not use Your Content to train generative AI models without your explicit consent.” Why didn’t they? Why specifically prohibit only models that replicate artists’ voice & likeness?
The difference between voice & likeness models and general-purpose models
In AI music, there are, broadly, two types of generative AI model. (I’m simplifying, but I think it’s useful to illustrate what’s at stake here.)
Voice models. These are generally designed to replicate an individual artist's voice.
Music models. These are more general-purpose, taking in a text prompt and generating music in a variety of styles in response. While they often do end up replicating the likeness of individual musicians and groups, their developers usually argue that they’re not designed to do this.
The language in SoundCloud’s terms seems to prohibit the first, but not the second.
That is, they rule out training things like voice models to replicate individual artists’ voices, but they leave open the possibility of training general AI music generation models on people’s music - and doing so without those musicians’ permission.
Why is this an issue?
This is incredibly problematic, because many artists don’t just care about stopping AI models replicating their likeness. They care about their music being used, without their permission, to train generative AI models that will compete with them in the market.
The US Copyright Office commented on the market dilution effects of generative AI in the report on AI training they released last week, saying it presents a major issue for the fair use defense of generative AI training. Generative AI competes with the work it’s trained on, and the people behind that work - whether or not it is specifically replicating their voice or their style.
How can SoundCloud solve this?
Eliah Seton, SoundCloud’s CEO, says in his post:
Our position is simple: AI should support artists, not replace them.
If he and SoundCloud genuinely believe this, the solution is simple.
Remove this: “We will not use Your Content to train generative AI models that aim to replicate or synthesize your voice, music, or likeness without your explicit consent.”
And replace it with this: “We will not use Your Content to train generative AI models without your explicit consent.”
With that change, musicians will be able to trust that SoundCloud will not train generative AI models on their music without their permission, and I suspect people would return to the service. I certainly would, and I’d be the first to thank them for listening to artists’ concerns.
But without that change, the terms update has no teeth. Whether by design or not, it is more marketing speak than it is meaningful for artists. SoundCloud continues to refuse to rule out training generative AI models on users’ music without asking permission first.