Did Your Distributor Just Licence Your Music for AI Training? We Read the Contracts

A music lawyer's warning that distributors are quietly getting artists to opt in to AI training has been circulating since February, with LANDR named as the worst offender. It is a good question and almost nobody has answered it from the contracts. So we opened four of them on 10 September 2026 and read the clauses line by line. The answer inverts the claim: the distributor everyone was warned about is the one that asks permission and pays, and the sweeping AI licence nobody is talking about took effect two days ago somewhere else entirely.

Filed 2026-09-10 Read 10 min Method How we work
In short
  • UnitedMasters added an explicit AI clause that took effect 8 September 2026. It grants the right to train models on your masters, to create derivative works, to license your masters to third-party "AI Licensees" for their own training, and to commercialise whatever those models produce. It is royalty-free and it applies by default.
  • That clause is opt-out, not opt-in, and it covers only content uploaded on or after 8 September 2026. The opt-out form excludes your work from future AI use — it does not untrain a model that has already learned from it.
  • LANDR, the distributor named in the viral claim, is the one running an explicit opt-in programme. Fair Trade AI pays 25 percent of net licensing revenue, offered a $5-per-track advance, and you can opt out at any time. Its general terms grant no AI training rights outside that programme.
  • DistroKid and TuneCore have no AI or machine-learning language in their terms at all. What they do have is a very broad general licence grant — "for any purpose on or in connection with the Services" and the right to "prepare derivative works" — which is not an AI clause but is not a prohibition either.
  • We could not verify CD Baby's contract text. Every published terms path we tried returned 403 or 404 on 10 September 2026, so we make no claim about it in either direction.
  • Secondary coverage of this topic is inconsistent and some of it is simply wrong about which distributors have AI clauses. Read the contract yourself, note the date you read it, and re-check it — these documents changed twice in the last six months.
Four plain contract pages fanned flat on a white surface, one edged and underlined in hot pink while the other three stay grey, representing four distributor terms of service of which two carry AI training language
Four contracts, read line by line on 10 September 2026. Two of them have AI clauses.

Since February a music lawyer's video has been circulating with a blunt claim: distributors are getting artists to opt in to AI training in a misleading way, and LANDR in particular. It has 16,600 plays, 984 likes and 96 comments, and a follow-up from the same account predicts the labels will come for consent through distributors specifically. Nobody in either comment section had opened a contract. So we did, reading what four music distributor AI training clauses actually say on 10 September 2026, quoting them rather than summarising them.

The takeaways are not what the video suggests. The distributor it names is the only one of the four that asks permission first and pays for the answer. The genuinely sweeping AI licence in this group belongs to a company nobody is warning anyone about, and it took effect two days before this article.

The claim, and why it is worth testing

The video is not unreasonable. Distribution contracts are long, they are updated without ceremony, and the update notice is an email most people archive. A clause that appears in version 14 of a document you agreed to at version 9 is, functionally, a clause you never read.

It is also the one story in this niche that runs the other way. Almost everything written about AI and distribution asks whether a distributor will accept AI-generated music — our own DistroKid AI music policy and TuneCore AI music policy pages are both that question. This asks what rights the distributor takes in whatever you hand over. Here the artist is not the one deploying AI. The artist is the training data.

What made it worth checking rather than repeating is that the secondary coverage disagrees with itself. One widely-cited roundup reports that DistroKid grants itself broad AI training rights. Another reports the opposite for the same company in the same year. At least one of those is wrong, and a reader has no way to tell which without the source document. That is a bad situation for a decision that is irreversible the moment you upload.

What we read, and what we could not

Table of four distributor contracts read on 10 September 2026: UnitedMasters has a dedicated AI section effective 8 September and is opt-out, LANDR has AI language only inside its Fair Trade AI programme and is opt-in, DistroKid and TuneCore have no AI or machine-learning language at all, and CD Baby could not be retrieved
Two of the four documents carry AI language. The one everyone was warned about is the opt-in one.

We opened each company's public terms, searched for artificial intelligence, machine learning, model training and data mining, and read the general licence grant in full whether or not those terms appeared. Where a clause exists we quote it.

Distributor Document read AI or machine-learning language? Consent model
UnitedMasters Terms of Service, effective 1 May 2026 Yes — a dedicated section effective 8 Sept 2026 Opt-out
LANDR Terms of Service + Fair Trade AI page Yes — but only inside a separate programme Opt-in
DistroKid Terms of Service No — no AI, ML, training or data-mining language Not applicable
TuneCore Terms and Conditions, §6.2 No — no AI, ML, training or data-mining language Not applicable
CD Baby Could not retrieve Unknown Unknown

CD Baby is an honest blank. Every published path to its terms returned 403 or 404 on the day we checked, so we make no claim about its contract in either direction. Our CD Baby AI music policy page covers its position on distributing AI-generated music, which is a different question from what its contract licenses.

UnitedMasters: the clause that took effect on 8 September 2026

Four numbered powers granted by the UnitedMasters AI clause: model training on your masters, creating derivative works to generate AI outputs, licensing your masters to third parties named AI Licensees for their own training, and commercialising any models or outputs produced
Royalty-free, and it applies by default to anything uploaded on or after 8 September 2026.

This is the finding. UnitedMasters' terms carry a section headed "AI and Machine Learning Use of User Content" that opens: "Effective [September 8, 2026], the following applies to User Content and Masters you upload to the platform on or after that date."

UnitedMasters Terms of Service page showing an effective date of May 1, 2026 and a table of contents listing sections including User Content and Feedback and Our Intellectual Property Rights
The UnitedMasters Terms of Service as read on 10 September 2026, effective 1 May 2026.

The square brackets are in the live contract. A bracketed date is drafting scaffolding that normally gets replaced before signature, and it is still sitting in the document readers are agreeing to.

The clause then defines four AI Uses.

# What the clause grants Wording
i Model training "to train, develop, test, validate, and improve machine learning and artificial intelligence models, whether owned, operated, or used by us, our affiliates, or any third party to whom we sublicense such rights"
ii Derivative works "to create derivative works of your User Content, including Masters, for the purpose of generating AI-produced outputs"
iii Third-party licensing "license your User Content, including Masters, to third parties (\"AI Licensees\") for the purpose of training their own machine learning and artificial intelligence models"
iv Commercialisation "deploy, commercialize, license, sublicense, and distribute any models trained pursuant to (i) above, any outputs created pursuant to (ii) above"

Read (iii) again. It is not only that the company may train on your masters. It may hand them to other companies to train on, and under (iv) may commercialise what those models produce. The underlying licence is described elsewhere in the same document as royalty-free, perpetual, irrevocable and sublicensable.

There is a real carve-out. AI uses expressly exclude any output that replicates or is substantially derived from the voice, name or likeness of a specific identifiable individual, and any digital replica as defined under applicable law, each requiring separate prior written consent. So the clause covers the recording, not a synthetic version of the person who made it.

Two limits cut in the artist's favour. It binds only content uploaded on or after 8 September 2026, so a back catalogue delivered before that date is outside it. And you may opt out at any time. We come back to what the opt-out does and does not achieve below.

LANDR: the distributor everyone was warned about is the one that pays

LANDR's Fair Trade AI is the programme the viral claim points at, and it is the clearest opt-in arrangement of the four. Its own page states you must opt in to participate and can opt out at any time.

The economics are published, which by itself distinguishes it. Participating artists receive 25 percent of net licensing revenue, split pro rata — LANDR's example is that a track in a licence of 10,000 tracks earns one ten-thousandth of that 25 percent, net of direct charges, paid monthly. A separate advance offered $5 per eligible track, recoupable, conditional on staying opted in until it recouped or 24 months passed. Eligibility requires holding 100 percent of both recording and publishing rights, which quietly excludes anyone with a co-writer who has not signed off.

LANDR Fair Trade AI Program page headed Investing $1 million into independent artists, describing an opt-in licensing programme that more than 30,000 artists have joined, with an Opt in today button
LANDR's own page: an opt-in programme, with the opt-in framed as the action the artist has to take.

More to the point for the original claim: LANDR's general terms of service grant no AI training rights outside that programme. The relevant general provision is that it may analyse user data to test functionality and to assess and improve its services, which is ordinary service-improvement language rather than a model-training grant.

There is an asymmetry worth naming. LANDR forbids its own users from training models on LANDR samples — "use any Sample for the purpose of training machine learning models, or for any form of Artificial Intelligence (AI) development". The company is stricter about its catalogue being trained on than its default terms are about yours. That is a fair thing to be annoyed about. It is not the thing the video said.

DistroKid and TuneCore: no AI clause is not the same as no

Table matching five powers an AI use would need against whether a broad general licence grant carries them: copying and storing is present in all four contracts, derivative works and sublicensing are named explicitly by TuneCore, but training a model and commercialising model outputs are not stated in any general grant
A general grant reaches three of the five powers. The two it does not reach are the two that define AI training.

Neither contract mentions AI. That is the reassuring half. The other half is what the general grants say.

DistroKid Terms of Service page dated as of September 1, 2024, opening with the definitions of the Site, Application and Services and the Terms of Access and Use section
DistroKid's terms, still dated 1 September 2024. No AI, machine learning, training or data-mining language appears anywhere in it.

DistroKid's licence covers submitted content "for any purpose on or in connection with the Services". TuneCore's §6.2 grants a worldwide, non-exclusive, royalty-free, transferable licence, sublicensable, to "use, reproduce, digitize, encode, store, transmit, distribute, modify, adapt, translate, prepare derivative works of, display, perform, communicate to the public and otherwise exploit" submitted content.

Power an AI use needs Present in a general grant? Present as an explicit AI clause?
Copy and store the recording Yes, in all four Yes
Create derivative works Yes — TuneCore names it explicitly Yes
Sublicense to third parties Yes — TuneCore names it explicitly Yes, to named "AI Licensees"
Train a model on it Not stated Stated
Commercialise model outputs Not stated Stated

The two rows that matter are the last two, and the honest reading is that these contracts are silent rather than protective. "Prepare derivative works" and "for any purpose" are broad enough that a company minded to argue the point would start there. A company that intended AI training and wanted it beyond argument would do what UnitedMasters did and write it down.

So the practical answer for DistroKid and TuneCore as of 10 September 2026 is: no AI training clause exists, no permission has been given, and no promise has been made either. Anyone telling you those contracts already license AI training is reading something that is not in the document.

Opt-in, opt-out, and the thing neither one undoes

Side-by-side comparison of LANDR Fair Trade AI and the UnitedMasters AI Uses section: LANDR excludes you until you opt in and pays 25 percent of net licensing revenue with a $5 per track advance, while UnitedMasters includes you until you opt out and pays nothing, and its moderation, fraud detection and content identification uses cannot be opted out of at all
Both are reversible going forward. Neither untrains a model that has already learned from your work.

Consent defaults are where the money and the ethics both sit.

LANDR Fair Trade AI UnitedMasters AI Uses
Default state Excluded until you opt in Included until you opt out
Payment 25% of net licensing revenue Royalty-free
Advance $5 per eligible track (offered) None
Scope Tracks you enrol Uploads on or after 8 Sept 2026
Reversible? Opt out at any time Opt out at any time
Not reversible Moderation, fraud detection, content identification

Now the part both models share. An opt-out is forward-looking. The UnitedMasters terms say that following your opt-out the company will exclude your masters from future AI use. There is no undertaking to remove your work's influence from a model already trained, because that is not a routine operation and no contract here offers it.

That makes the upload the irreversible act and the opt-out a much weaker instrument than it sounds. If you care about this, the moment to act is before delivery, not after a policy email. It is also why the 8 September boundary matters more than it first appears: material delivered before that date sits outside the clause permanently, and material delivered after it is inside from the second it lands.

Five phrases to search for in your own contract

This is generalisable. Open your distributor's terms and search for five strings.

Search for What it tells you Seen in
"artificial intelligence" / "machine learning" Whether an explicit AI clause exists at all UnitedMasters, LANDR
"train" Distinguishes model training from ordinary processing UnitedMasters, LANDR
"derivative work" A general grant broad enough to reach toward AI outputs TuneCore, UnitedMasters
"sublicense" Whether third parties can receive your masters TuneCore, UnitedMasters
"opt out" Whether the default is consent UnitedMasters

Two habits make this worth repeating rather than doing once. Note the date you read it — both contracts with AI language here changed in the last six months, and the UnitedMasters section is nine days old at the time of writing. And read the general grant even when the AI search returns nothing, because that grant is what a company would have to rely on if it ever wanted to argue the point.

If you want the wider commercial picture around these agreements, our music distribution services comparison covers what each one costs and takes, and free music distribution covers the tiers where the terms tend to be broadest.

What this changes if you release AI-generated music

If your tracks come out of Suno or Udio, you now have two separate contract questions rather than one. Whether the distributor will accept an AI-generated release is the familiar one. Whether its terms license your delivered masters for someone else's model training is the new one, and it is answered in a different section of a different document.

Neither has anything to do with the third gate, which is automated. Distributor screening systems read the audio, not the paperwork. Tracks do get flagged and rejected by those systems, and no clause in any of these contracts changes what a spectral analysis sees — that is what Undetectr addresses, and it is a genuinely unrelated problem from the one this article is about. Our AI music copyright page covers the ownership question that sits underneath both.

It is worth being clear-eyed about the wider situation, because contract terms are not most people's actual bottleneck. The recurring complaint in the AI-music communities is not that distribution is impossible; it is that nobody listens once a release is live. Distribution is largely solved and discovery is not.

That is the honest case for not routing everything through platforms whose terms you do not control. Paid sync placement — getting a track into a show, a game or an ad — is the part of this market where the money is discussed in concrete numbers, and it does not depend on algorithmic reach at all; played.fm runs that pitch route alongside a direct storefront where you keep 100 percent and own the customer relationship. Our sync licensing guide covers how that side of the business works regardless of which platform you use.

The wider point is smaller than the video that prompted it, and more useful. One distributor now takes AI training rights by default. One asks and pays. Two are silent. The difference between those positions is about 400 words of contract text, and it is on the public internet in every case. Read it before you upload, note the date, and check it again in six months.

Frequently asked

Questions readers ask.

It depends entirely on which one, and the answer changed on 8 September 2026. UnitedMasters now has an explicit clause granting itself the right to train machine learning and AI models on masters uploaded on or after that date, and to license those masters to third parties for their own training, unless you opt out. LANDR only does it if you actively opt in to its Fair Trade AI programme, which pays a revenue share. DistroKid and TuneCore have no AI or machine-learning language in their terms at all as of 10 September 2026.

There is no artificial intelligence, machine learning, model training or data mining language anywhere in DistroKid's terms of service as of 10 September 2026. Some secondary articles claim otherwise and quote a clause we could not find in the document. What DistroKid does have is a broad general licence to use submitted content "for any purpose on or in connection with the Services", which is not an AI training grant but is also not a bar to one. If that distinction matters to you, the honest position is that the contract is silent, not that it says no.

Opt-in. LANDR's own page states that you must opt in to participate and can opt out at any time. Participating artists receive 25 percent of net licensing revenue, divided pro rata across the tracks in a given licence, paid monthly through the LANDR account. There was also a $5-per-eligible-track advance, recoupable against future revenue, which required staying opted in until it recouped or 24 months passed. Eligibility requires holding 100 percent of both the recording and publishing rights.

UnitedMasters provides an opt-out form linked from the AI section of its terms of service, and the terms say you may opt out of any AI use at any time. Two limits are worth knowing. First, moderation, fraud detection and content identification uses are described as integral to platform operation and are not subject to opt-out. Second, the terms say that following your opt-out the company will exclude your work from future AI use, which does not reverse training that has already happened.

No, and no distributor contract we read promises that it does. UnitedMasters commits specifically to excluding your masters from future AI use. Removing a work's influence from a model that has already been trained is not a routine operation, and none of these agreements offer it. This is the strongest practical argument for reading the terms before you upload rather than after, because the opt-out is a forward-looking switch and the upload is the irreversible part.

Search the document for six strings: "artificial intelligence", "machine learning", "train", "derivative work", "sublicense" and "opt out". The first three tell you whether an AI clause exists. "Derivative work" and "sublicense" tell you how far a general grant reaches even when no AI clause exists, because those are the two powers an AI use would need. "Opt out" tells you whether the default is consent. Note the date you read it, because two of the four contracts here changed in the last six months.

No. Every agreement we read is a licence, not a transfer of ownership, and each states that you retain your rights and warrant that you hold them. The practical question is not ownership but scope: a royalty-free, perpetual, irrevocable, sublicensable licence to train models and commercialise the outputs is a long way from a licence to put a track in shops, even though you still own the copyright in both cases.

Yes, and it is the inversion worth noticing. Most coverage in this niche asks whether a distributor will accept AI-generated music. These clauses ask the opposite question — what rights the distributor takes in the recordings you deliver, however they were made. Both matter, and they are unrelated gates. A contract that says nothing about AI training has no bearing on whether an automated screening system flags your upload, and clearing screening has no bearing on what the contract licenses.

The verdict, in one sentence: Undetectr.

Contract language and screening systems are unrelated gates, and reading one tells you nothing about the other. Undetectr is the tool in our benchmark that addresses the layer distributors actually screen for. $39 one-time, roughly 90 seconds per track.