Universal and Sony’s second lawsuit against Suno raises a question that could matter far beyond AI music: if a new model is trained on licensed material, can it still carry copyright risk inherited from an older model?
For the past few years, the AI music copyright debate has largely revolved around one question:
What music was used to train the model?
The new lawsuit against Suno introduces a more complicated one.
Where did the model’s knowledge come from?
On September 18, Universal Music Group and Sony Music Entertainment filed a second lawsuit against Suno in federal court in Massachusetts. The complaint accuses the AI music company of copying 60,202 sound recordings without permission. It also reaches directly into Suno’s newly launched v6 generation of models.
That is what makes this case particularly interesting.
Suno has presented v6 as a reset of sorts. The company says the model was trained “entirely from scratch” on a different dataset and that the training data does not contain recordings from Universal or Sony. V6 was launched in partnership with Warner Music Group, BMG and Believe, following a broader move by Suno toward licensed relationships with the music industry.
Universal and Sony argue that starting again may not be enough.
And if that argument succeeds, the consequences could reach well beyond Suno.
A New Model Is Not Necessarily a Clean Model
There is an important distinction here.
When an AI company says a model was trained from scratch, that usually tells us something about the technical construction of the new model.
It does not necessarily answer every question about the information used to develop it.
Suno Chief Product Officer Jack Brody told Music Business Worldwide that v6 used different training data from earlier Suno models. According to Suno, that includes licensed partner data and preference information showing which generations users preferred.
Universal and Sony focus on that second part.
The complaint alleges that these user preferences were created by interactions with earlier Suno models. A user would receive two generated songs, choose the preferred version, and that preference signal could help Suno improve later systems.
The labels argue that these signals cannot simply be separated from the models that created the music being judged.
Their theory is essentially this:
old recordings → old model → generated outputs and user preferences → new model
If the first stage involved unauthorized copyrighted material, they argue, moving several steps further down that chain does not necessarily erase the original problem.
That idea is the heart of this lawsuit.
The Knowledge-Distillation Question
The complaint goes further.
Universal and Sony allege that Suno may also have used knowledge distillation or related model-transfer techniques while developing v6.
Knowledge distillation allows a newer “student” model to learn from the behavior of an older “teacher” model.
The new model does not necessarily receive the teacher model’s original training files.
Instead, it can learn from what the older system already knows or produces.
Universal and Sony argue that this creates a copyright chain.
Their claim is that if earlier Suno models learned musical capabilities from recordings copied without permission, transferring those capabilities into v6 may transfer some of the value derived from those recordings as well. The complaint describes this as passing expression from recordings into earlier models, into their outputs and finally into v6.
It is an important allegation.
But it is still an allegation.
A court has not yet established a general rule saying that an AI model automatically becomes legally contaminated because it learned from another model whose training is disputed.
That distinction matters.
Copyright Does Not Have a Simple “Contaminated Model” Rule
The phrase used by Universal and Sony is deliberately memorable.
They call v6 the “fruit of the same poisoned tree.”
But copyright law does not currently contain a simple rule where Model A infringes, therefore every Model B influenced by Model A automatically infringes too.
In fact, guidance from the U.S. Copyright Office points toward something more granular.
Its report on generative-AI training says different uses of copyrighted material during AI development may require separate consideration. Initial training, later training, fine-tuning and other stages can raise different fair-use questions.
That creates the real legal problem Suno’s case may eventually help explore.
The question is not simply:
“Did v6 descend from an older model?”
It is closer to:
“What copyrighted expression, if any, moved through that chain, and what acts of copying occurred while building the successor model?”
Those are much harder questions.
There Are Several Different Copyright Chains in the Complaint
Universal and Sony are effectively presenting several routes through which they say older copyright issues reached v6.
First, they allege Suno still retains unauthorized copies of their recordings and may have continued using those materials.
Second, they point to synthetic music generated by previous models and argue that some of those outputs may contain expression originating from copyrighted recordings.
Third, they point to preference data generated when users chose between outputs from older models.
And fourth, they allege the use of knowledge-transfer techniques capable of moving learned behavior from older models into newer ones.
These are related arguments, but they are not identical.
And that distinction may become extremely important.
Using an original copyrighted recording again is different from training on an AI-generated track.
Training on generated audio is different from learning that users prefer one generated track over another.
And transferring general capabilities from one neural network to another may raise different copyright questions again.
The case could eventually force courts to examine those layers separately.
Licensing Does Not Necessarily Reset the Clock
There is another misconception this lawsuit exposes.
We often speak about licensed AI models as though licensing is a binary state.
Licensed.
Or unlicensed.
Reality may become much messier.
Suno has reached agreements with Warner Music Group, BMG and Believe. Believe has also said artists distributed through Believe and TuneCore must opt in before their recordings are supplied to Suno.
Those agreements matter.
But a license from one group of rightsholders does not automatically grant rights over another company’s catalogue.
And a license governing today’s training data does not necessarily settle questions concerning what happened during earlier development.
Universal and Sony are trying to establish exactly that distinction.
Their argument is essentially that future licensing cannot retroactively clean an allegedly infringing development history.
Whether the courts accept that argument remains unresolved.
Suno Has a Very Different Story
Suno’s public explanation of v6 is considerably simpler.
The company says it created a new model from the ground up using a different dataset.
Brody specifically told MBW that the v6 training data does not include Universal or Sony material. He also argued that improvements between generations do not depend only on adding more training music.
Architecture, research, tuning, user preferences and engineering advances can all improve a model.
That is an important counterpoint.
Machine-learning systems do not improve solely because developers pour more copyrighted works into them.
Algorithms improve.
Training techniques improve.
Architectures change.
Human feedback improves.
Developers learn from previous experiments.
The legal system will therefore need some way of separating ordinary technological knowledge from protectable creative expression.
That line may become one of the central questions of the next phase of generative-AI copyright litigation.
Why This Matters Beyond Suno
Imagine an AI company trains Model A on billions of unlicensed works.
Years later, it decides to become fully licensed.
The company creates Model B using licensed material.
But Model B is improved using:
- synthetic data produced by Model A
- preference information collected from Model A
- evaluations comparing Model A’s outputs
- teacher-student model distillation
- technical knowledge developed while operating Model A
At what point has the company genuinely started again?
There is no simple answer.
And that may become one of the most consequential copyright questions facing generative AI.
Because companies increasingly want to move from the scrape-first era of AI development toward a licensed ecosystem.
If historical models can transmit legal risk into later models, licensing the future becomes much more complicated.
AI companies may need something closer to data provenance for entire model families.
Not merely:
What trained this model?
But:
What trained the systems, datasets, synthetic outputs and feedback mechanisms that trained this model?
That is a much longer chain.
The Bigger Shift: From Training Data to Model Lineage
For creators, labels and AI companies, this may be the most important part of the Suno v6 lawsuit.
The conversation is moving beyond datasets.
It is moving toward model lineage.
A future AI company may need to document not only the recordings contained in a training corpus, but also:
- where synthetic training material originated
- which earlier models generated it
- how user feedback was collected
- whether previous model outputs influenced later systems
- which models served as teachers during distillation
- what rights covered each stage
That sounds technical.
But it may eventually become the AI equivalent of a music rights chain.
Every sample has a source.
Every master has ownership.
Every composition has rights.
Perhaps every AI model will eventually need a provenance trail too.
The Question Suno v6 Leaves Behind
The most interesting part of this lawsuit is therefore not whether AI music is good or bad.
That argument is already exhausted.
The more useful question is narrower.
Can a company solve yesterday’s copyright problem simply by licensing tomorrow’s training data?
Universal and Sony say no.
Suno’s position around v6 points toward a different answer: a genuinely new model, trained on new data and new technical methods, should be treated as a new system.
The courts have not yet resolved that conflict.
But the question will matter far beyond one AI music generator.
Because if AI models can inherit capabilities from previous generations, the law may eventually have to decide whether they can inherit liabilities as well.
And that could determine what a “licensed AI model” actually means.
