Read the AI clause before you upload: what the SoundCloud terms fight teaches about platform ToS

Most musicians accept a platform's terms of use once, at sign-up, and never read them again. The terms keep changing anyway. SoundCloud's AI clause is a useful case study because every version is still public: the wording that caused the row, the wording that replaced it, and what is in force today. Read side by side, they show both what a platform can take from an upload and how much a single qualifying phrase can matter.

What the 2024 clause said

SoundCloud's terms dated 7 February 2024 added one sentence to the licence section: "In the absence of a separate agreement that states otherwise, You explicitly agree that your Content may be used to inform, train, develop or serve as input to artificial intelligence or machine intelligence technologies or services as part of and for providing the services."

The row started more than a year later. In May 2025 the tech ethicist Ed Newton-Rex flagged it, and TechCrunch reported it on 9 May. The Verge describes him as the person "who first discovered the change". A SoundCloud spokesperson told TechCrunch the company "has never used artist content to train AI models, nor do we develop AI tools". The spokesperson said the update was meant to clarify how content "may interact with AI technologies within SoundCloud's own platform". TechCrunch added that it couldn't find an explicit opt-out in SoundCloud's web settings.

The company's assurance and the contract's wording were two separate things. "Inform, train, develop" has no limit on the kind of model, and "as part of and for providing the services" is broad enough to cover almost anything SoundCloud chooses to call a service.

What changed in May 2025

On 14 May 2025 SoundCloud CEO Eliah Seton published a letter to artists. He wrote that "the language in the Terms of Use was too broad and wasn't clear enough. It created confusion, and that's on us." He listed what SoundCloud does use AI for: recommendations, search, playlisting, content tagging and fraud prevention. He also pointed to a "no AI" tag that he said signals content can't be used for AI training. Then he set out a new commitment: SoundCloud "will not use your content to train generative AI models that aim to replicate or synthesize your voice, music, or likeness". If a generative AI opportunity comes up, he wrote, it "may" be offered to artists "via an opt-in mechanism". Music Business Worldwide noted the original wording had drawn "fierce criticism" from artists and AI experts.

There is a detail here worth knowing. The "revised Terms of Use" text printed under the letter still contained the old "inform, train, develop" sentence, right after the new promise. The version that actually took effect, dated 21 May 2025, doesn't: it has the new sentence and drops the old one. The current terms, last amended 17 August 2026, keep the same wording: "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, which must be affirmatively provided through an opt-in mechanism."

What the new sentence leaves open

Newton-Rex wasn't satisfied. As quoted by The Verge, he said the new language could still allow "models trained on your work that might not directly replicate your style but that still compete with you in the market". His proposed fix was one short sentence: "We will not use Your Content to train generative AI models without your explicit consent."

In our reading, three qualifiers do the narrowing:

• "Generative". Recommendation, tagging and fraud models aren't covered at all. That is arguably fine, since those systems help listeners find you. But it means the promise is about one class of model, not about training in general.
• "That aim to". The test is the model's purpose, and whoever trains it gets to describe that purpose. A general music generator trained on a large catalogue doesn't obviously "aim" to replicate any one artist.
• "We". The promise binds SoundCloud. The sentence after it, which says neither SoundCloud "nor any third party" may use content for AI training without authorisation, applies to content "delivered to the Platform under separate agreements" and "owned or controlled by third party rights holders". Our reading is that this covers label and distributor deliveries more clearly than it covers a track you upload yourself.

Two other parts of the current terms are worth knowing. Licences in audio "terminate automatically" when you remove the content, though SoundCloud is not obliged to make sure copies already shared to linked sites are deleted. And for "significant changes", SoundCloud promises prominent notice, with the new terms taking effect two weeks later.

What we think

SoundCloud deserves credit. It reversed within days, apologised plainly, and put the promise in the contract instead of leaving it in a blog post. It also keeps every past version of its terms online, which is exactly what made this analysis possible. Many platforms don't.

But the fix is narrower than the headline. "No AI clones of you without opt-in" is a good floor. It isn't the same as "no training on your music without opt-in", and that is the promise artists actually wanted. The market harm from AI music doesn't come mainly from a model that sounds like you. It comes from a model trained on millions of tracks like yours that floods the same playlists. In our view Newton-Rex's one-line version is the right standard, and nothing in SoundCloud's own account of how it uses AI would stop it adopting that wording.

The bigger lesson goes beyond SoundCloud. An upload licence is a contract, and platforms argue about it in court: our Kogon v. Google coverage shows what happens when a company argues that a video platform's terms already licensed AI training. The letter itself is a small warning too. The draft text printed under it and the clause that took effect weren't the same. Read the terms, not the announcement.

Before you upload: AI clauses to look for

• Training verbs. Search the licence for "train", "develop", "inform", "input", "machine learning" and "artificial intelligence". Note which section each one sits in.
• Qualifiers. Watch for "generative", "aim to", "replicate", "for providing the services". Each one narrows a promise or widens a grant.
• Who is bound. Does the promise cover only "we", or also affiliates, partners and licensees?
• Opt-in or opt-out. Which is the default, and where is the setting? If you can't find it, assume there isn't one.
• Carve-outs. Phrases like "in the absence of a separate agreement" or "content delivered under separate agreements" can mean distributor uploads get protections that direct uploads don't.
• Deletion. Does the licence end when you delete a track? What about copies on linked sites, or models already trained?
• Change notice. How are you told when the terms change, and how long before the new terms apply?
• Impersonation rules. Check whether the terms ban others from using AI to copy your voice or likeness. SoundCloud's current terms do.
• Keep a dated copy. Save the terms, with their "last amended" date, when you first upload, and check the date again every few months.