The new Suno litigation, model lineage and the limits of copyright liability under Turkish law
By Asutay Duhan Meydan / Attorney at law
On 18 September 2026, Universal Music Group and Sony Music Entertainment filed a second action against Suno in the United States District Court for the District of Massachusetts. The new complaint concerns 60,202 sound recordings and follows the court’s earlier refusal to add more than 61,000 additional works to the original 2024 proceedings. Judge F. Dennis Saylor IV did not reject those additional claims on their merits. He concluded that adding them to the existing action would substantially expand and delay proceedings that were already approaching the end of fact discovery, expressly leaving the labels free to pursue the works in a separate action. That is what they have now done.
The second action, however, is more significant than its size. It places Suno’s newly launched v6 models at the centre of a broader question concerning the legal continuity between successive generations of artificial intelligence systems.
Suno has presented v6 as a model family trained from scratch using licensed material and developed in partnership with Warner Music Group, BMG and Believe. Universal and Sony challenge the significance of that distinction. Their complaint alleges that v6 nevertheless inherited value derived from the earlier Suno models through synthetic outputs, user preference data and knowledge distillation. In the language of the complaint, v6 is not a clean start but the “fruit of the same poisoned tree.”
The metaphor is powerful. Its application under Turkish copyright law, however, would require considerably more precision.
The Legal Chain Behind the New Suno Case
The labels’ argument can be reduced to a relatively simple sequence.
Their sound recordings were allegedly copied without authorisation and incorporated into the datasets used to develop earlier Suno models. Those models subsequently produced synthetic music. Users interacted with those outputs and generated preference data indicating which outputs were preferred. Universal and Sony further allege that Suno used knowledge distillation, through which a new model learns from the behaviour of an earlier model.
The alleged chain is therefore:
copyrighted recordings → earlier Suno models → synthetic outputs and learned behaviour → preference data and model transfer → v6.
The central proposition is that placing another model between the copyrighted material and the final model does not necessarily break the chain. According to the complaint, training a new model on outputs derived from an earlier infringing model may simply transfer the value extracted from the original recordings into a new technical architecture.
There is an important qualification. The first Suno proceedings have not yet resulted in a ruling that training the earlier models constituted copyright infringement. Suno continues to rely on fair use under US law, and Judge Saylor has stated that summary judgment on that issue is likely to resolve the predominant legal question in the original action. The “poisoned tree” therefore remains a theory advanced by the plaintiffs, not a judicial finding that the original tree was legally poisoned.
That distinction becomes even more important when the same factual structure is examined under Turkish law.
Where Would the Turkish Analysis Begin?
Turkey does not need a “poisoned tree” doctrine to analyse the first stage of the alleged conduct.
Article 22 of Law No. 5846 on Intellectual and Artistic Works, the FSEK, grants the author the exclusive right to reproduce a work, wholly or partly, directly or indirectly, temporarily or permanently and by any method. The provision deliberately uses technologically neutral language and expressly includes recording a work on devices capable of transmitting or reproducing sound and images.
For recorded music, Article 80 is equally important. A phonogram producer, once the relevant economic rights have been acquired, has the exclusive right to authorise or prohibit the direct or indirect reproduction of the phonogram. Performers also enjoy separate neighbouring rights in their recorded performances.
The first question under Turkish law would therefore not be whether an AI system “learned” from a song.
It would be much more concrete:
Was the protected work, performance or phonogram copied in order to create or operate the training dataset?
If a commercial AI provider downloaded a protected recording, stored it in a training corpus and reproduced it as part of model development without permission, Article 22 and, where applicable, Article 80 would provide the natural starting point.
This is also where the comparison with US law begins to diverge. FSEK does not presently contain a general fair use doctrine. Nor does the existing Act contain a general AI training or text and data mining exception capable of automatically legalising commercial model training. Article 38 permits reproduction for personal use without a profit motive, subject to the legitimate interests of the right holder and normal exploitation of the work. That exception does not naturally accommodate large scale commercial model training.
The initial copying alleged in Suno would therefore raise a comparatively direct reproduction question under Turkish law.
The harder question begins after the first model has already been trained.
Does Infringement Pass from One Model to Another?
This is where the poisoned tree metaphor reaches its limits.
FSEK does not establish a rule under which a later product automatically becomes unlawful merely because an earlier stage in its technological history involved an infringement. Turkish copyright law protects defined works, performances, phonograms and exclusive economic rights. It does not generally protect abstract knowledge, musical concepts, statistical relationships or a model’s accumulated ability to generate music.
Accordingly, the following proposition would be too broad under Turkish law:
Model A was trained unlawfully. Model B learned from Model A. Therefore Model B is also unlawful.
An additional legal connection would still have to be established.
Suppose, for example, that the earlier model generates thousands of synthetic audio files and those files are subsequently used to train the new model. If those synthetic files reproduce protected expression from the underlying works or phonograms, their incorporation into the later training corpus may constitute a further act of reproduction or, depending on the circumstances, engage the adaptation right under Article 21.
The analysis is different if the transferred material consists only of abstract behavioural information.
A preference signal saying that users preferred Output A to Output B is not itself the copyright holder’s song. Similarly, information that a particular harmonic sequence, rhythm or arrangement pattern tends to produce more successful outputs does not automatically amount to protected expression. Copyright cannot be extended merely because the information ultimately originated from a system that had once been trained on copyrighted material.
The same caution applies to knowledge distillation.
The legal result should depend on what actually occurs during the technical process. If protected recordings or outputs reproducing protected expression are copied again while training the student model, the reproduction right may be engaged again. If the student model merely receives behavioural probabilities or other abstract model information without reproducing protected expression, the connection to Articles 21 and 22 becomes substantially more difficult to establish.
In other words, under FSEK, model lineage is evidence of where to investigate. It is not, by itself, a cause of action.
Why “Trained from Scratch” May Still Not End the Inquiry
This also explains why describing v6 as “trained from scratch” would not necessarily dispose of a Turkish claim.
From a technical perspective, a new set of model weights may indeed have been created from the beginning. That does not answer whether copyrighted material was used elsewhere in the development process.
A Turkish court would have to examine whether protected recordings remained in retained datasets and whether they were later used for training, fine tuning, benchmarking, evaluation, testing or model improvement. The same question would arise where outputs from earlier models were incorporated into a new dataset.
Each use must be examined on its own facts.
The distinction is important:
A new model architecture does not erase earlier infringement. Equally, earlier infringement does not automatically make every later model infringing.
The relevant issue is whether a new act falling within an exclusive right recognised by FSEK occurred at a later stage of development.
This is where the new Suno complaint becomes particularly useful for Turkish lawyers. Instead of treating AI training as a single moment, it requires the entire development pipeline to be examined: acquisition, storage, initial training, output generation, preference collection, model transfer, fine tuning, testing, evaluation and later model improvement.
Turkish Law May Be Moving Directly Towards This Question
There is an additional development that makes the Suno litigation particularly timely for Turkey.
A bill submitted to the Turkish Grand National Assembly on 8 April 2026, numbered 2/3634, proposes substantial amendments to FSEK concerning artificial intelligence. As of 21 September 2026, the proposal remains before the parliamentary commissions and is not in force.
The proposed Article 42/D is remarkable in the context of Suno because it does not refer only to initial training. It expressly addresses the reproduction, processing, analysis, data mining, storage or other use of works, performances, phonograms, productions and broadcasts for the training, development, fine tuning, testing, evaluation, dataset creation and model improvement of AI systems offered for use in Turkey. The proposal would subject those uses to a licensing mechanism against appropriate remuneration.
That wording closely tracks the factual stages now being contested in the Suno proceedings.
The proposal should not be used retrospectively to rewrite current FSEK. It does, however, reveal the direction in which the Turkish legislative debate is developing. The legal issue is no longer limited to whether copyrighted material appears in the original training dataset. Attention is shifting to the entire lifecycle through which protected material may contribute to successive generations of a model.
What Would the Turkish Version of the Poisoned Tree Actually Look Like?
If the Suno allegations were examined under Turkish law today, the strongest claim would not be that the later model was contaminated by its ancestry.
The stronger legal chain would be narrower:
unauthorised reproduction of a protected work or phonogram → retention of that copy → further use of the protected material or outputs containing protected expression → reproduction or processing during development of the subsequent model.
At every stage, the claimant would need to connect the defendant’s conduct to a right recognised under FSEK.
That approach is less dramatic than saying that the fruit of a poisoned tree must itself be poisoned. It is also legally more precise.
Where the original recording is copied again, the later stage may constitute another infringement. Where protected expression contained in synthetic outputs is reproduced and reused, a new infringement may similarly arise. Where only abstract knowledge, preference information or general model capability passes from one generation to another, copyright law faces a more difficult question because the subject matter transferred may no longer be protected expression at all.
This distinction may ultimately become one of the most important questions in generative AI litigation.
Copyright law was developed around identifiable works and identifiable acts of exploitation. Modern AI development distributes those acts across datasets, model generations, synthetic outputs and feedback loops. The challenge is therefore not simply to decide whether a model has a problematic history. It is to identify precisely where protected expression enters, remains within, or re-enters the technical chain.
The new Suno litigation puts that question directly before the courts.
Under Turkish law, the answer would not be that every fruit from the same tree is automatically poisoned.
The real question is whether the protected work is still somewhere in the chain.
Asutay Duhan Meydan
Attorney at Law