A Comparative Analysis of U.S. Copyright Law and Turkish Law in the Absence of a Turkish AI Act
25 August 2026
On 17 August 2026, Round Hill Music and affiliated rights holders filed two copyright actions before the United States District Court for the Northern District of California. One was brought against Anthropic PBC; the other against Suno, Inc. together with Bright Data entities.
The Anthropic complaint alleges direct copyright infringement, circumvention of technological protection measures, and removal or alteration of copyright management information. The Suno action adds a claim of contributory copyright infringement against Bright Data. Both actions concern the alleged unauthorized use of copyrighted musical compositions and lyrics in the construction or training of artificial intelligence systems. Round Hill initially identified 500 works as representative works, while indicating that the dispute may ultimately concern a substantially larger catalogue. (Justia Dockets & Filings)
At the time of writing, these are allegations contained in newly filed complaints. There has been no judgment on the merits in either Round Hill case. That distinction matters. The interesting question for Turkish law is therefore not whether Round Hill will win in California, but something different:
What would a Turkish court do if substantially the same conduct were brought before it under Turkish copyright law?
The answer reveals an important characteristic of Türkiye’s present AI framework. Türkiye does not lack law applicable to artificial intelligence. What it lacks is legislation telling courts how existing rights should be recalibrated when those rights are exercised against artificial intelligence training.
That difference is fundamental.
First, a Turkish court could not simply apply FSEK to conduct occurring in California
Before considering copyright infringement itself, a Turkish court would have to address applicable law.
Article 23 of Law No. 5718 on Private International Law and International Civil Procedure adopts the lex loci protectionis principle: intellectual property rights are governed by the law of the country for which protection is sought. Accordingly, the fact that an action happens to be heard by a Turkish court does not automatically make Turkish copyright law applicable to every act of copying performed abroad. (Tapu ve Kadastro Genel Müdürlüğü)
For the purpose of this analysis, therefore, the hypothetical must be stated correctly. Assume that equivalent acts of dataset acquisition, reproduction, training, storage or exploitation occurred in circumstances in which protection was sought in Türkiye and Turkish law applied.
Only then does the real comparison begin.
The Turkish case would start with FSEK Article 22, not with an AI statute
The first surprise is that a Turkish claimant would not necessarily have difficulty identifying an existing statutory right capable of reaching AI training.
Article 22 of the Fikir ve Sanat Eserleri Kanunu, Law No. 5846 (“FSEK”), defines the reproduction right broadly. Reproduction includes making copies of a work, completely or partially, directly or indirectly, temporarily or permanently, by any means or method. The Ministry of Culture and Tourism describes the right in the same terms. (Telif Hakları)
That language matters enormously in an AI-training dispute.
An AI developer does not need to publish a conventional pirated copy of a lyric on a website for Article 22 to become relevant. If a protected work is downloaded, stored in a dataset, duplicated during preprocessing, transformed into another machine-readable representation or copied temporarily during a computational process, the starting question under Article 22 is whether a reproduction of protected expression has occurred.
In a Round Hill-type case, therefore, the first stage of the Turkish analysis would be relatively conventional. The court would examine whether the claimants held enforceable rights in the compositions or lyrics, whether the relevant material qualified for protection under FSEK, and whether the defendant had reproduced that protected material without authorization.
This is why saying that “Türkiye has no AI law, therefore AI training is unregulated” is legally inaccurate.
The conduct is not outside the law.
The difficulty begins after reproduction has been established.
The decisive difference: the United States has Section 107; Türkiye does not
Under U.S. law, reproduction is not the end of the inquiry.
Section 106 of Title 17 of the United States Code grants copyright owners exclusive rights including reproduction and preparation of derivative works. But Section 107 then subjects those exclusive rights to the doctrine of fair use. A U.S. court may consider the purpose and character of the use, the nature of the protected work, the amount used, and the effect of that use upon the potential market for or value of the copyrighted work. (Telif Hakkı Ofisi)
That is the doctrinal door through which an AI company can walk.
It does not guarantee that AI training is lawful. It allows the defendant to argue that an otherwise unauthorized reproduction should nevertheless not constitute infringement because of the nature, purpose and economic effects of the use.
The importance of that mechanism became visible in Bartz v. Anthropic.
In June 2025, Judge William Alsup held that Anthropic’s use of books in the training of its large language models was fair use on the record before the court. But the same decision drew a sharp distinction between the training process and Anthropic’s acquisition and retention of books obtained from pirate libraries. The court treated the training reproductions as transformative while refusing to give Anthropic a general fair-use shield for the pirated library copies from which its training materials had been sourced. (Justia Dockets & Filings)
That distinction is directly relevant to Round Hill.
It demonstrates why the statement “AI training is fair use in America” would be much too broad. American law is developing a more granular inquiry: What was copied? How was it obtained? Why was it copied? What happened to the copy? Does the model substitute for the protected work? What market is affected?
Kadrey v. Meta reinforces that caution. Meta prevailed on fair use on the evidentiary record before the court, but the court expressly rejected the idea that copyright law simply immunizes AI training as a category. The result turned significantly on the plaintiffs’ failure to establish sufficient market-harm evidence. A subsequent 2026 order again emphasized that Meta had prevailed because of the evidence presented in that particular case, not because training copyrighted works is invariably fair use. (Justia Law)
A Turkish court would not have an equivalent Section 107.
And that changes the case substantially.
FSEK contains exceptions, but they are not a Turkish version of fair use
Turkish copyright law certainly contains limitations and exceptions. Article 35 regulates iktibas serbestisi, or quotation. Article 38 permits reproduction for personal use without a profit-making purpose, subject to protection of the right holder’s legitimate interests and normal exploitation of the work. Other provisions create specific exceptions for education, news reporting and particular socially justified uses. (Telif Hakları)
But these provisions do not create a free-standing judicial power equivalent to Section 107.
Imagine an AI company copying the complete lyrics of thousands of commercially exploited songs into a training corpus in order to develop a commercial generative model.
It would be difficult to characterize that activity as quotation under Article 35. The object of the reproduction is not ordinarily to quote identifiable passages within an independent scientific, literary or musical work in the manner contemplated by the provision. Nor would Article 38’s personal-use exception provide a natural defence to industrial-scale commercial model training.
This leads to the most important difference between the two systems.
An American judge can accept that reproduction occurred and still ask whether the reproduction is fair.
A Turkish judge can accept that reproduction occurred, but must then find a statutory limitation or another recognized legal basis capable of taking the conduct outside the right holder’s exclusive right.
There is no broadly formulated FSEK provision instructing the Turkish judge to balance transformative purpose, commercial character, quantity copied and market effects in the same way that Section 107 does.
For an AI developer, that is a materially less comfortable position.
If Round Hill proved unauthorized full-work training copies in Türkiye, its prima facie case could therefore be stronger
Assume Round Hill could prove that protected lyrics were intentionally downloaded, reproduced in their entirety and incorporated into a commercial training process without authorization.
Under present Turkish law, I would expect the core infringement question to begin in Round Hill’s favour.
The claimant could point directly to Article 22. The AI company could not simply import American “transformative use” doctrine into FSEK and ask the court to reproduce the Bartz analysis. Turkish courts interpret Turkish statutory limitations; they do not have Section 107.
That does not mean judgment for the claimant would be automatic.
Round Hill would still have to prove its rights, identify the works, establish the relevant reproductions and connect the allegedly infringing acts to the defendant. The evidentiary problem could be considerable where the relevant information about training datasets is held exclusively inside the AI developer’s infrastructure.
There is also an unresolved technical-legal distinction which should not be avoided: the training dataset, training-time copies and the resulting model weights are not necessarily the same legal object.
If copyrighted lyrics were copied into a dataset, Article 22 is comparatively easy to discuss. Whether trained model parameters should themselves be treated as “copies” of the underlying works is a substantially more difficult question. Turkish legislation does not answer it, and it would be premature to pretend that FSEK already provides a settled answer.
This is precisely where the absence of AI-specific legislation begins to show.
The output side is a separate infringement question
Training and output should not be collapsed into one act.
Suppose an AI system was trained on a protected song but never reproduced identifiable protected expression to a user. That presents one legal question.
Suppose instead that a user asks for a song and the system produces substantial protected lyrics, or an output containing protected expression sufficiently dependent on the pre-existing work. That presents another.
Under FSEK, Article 21 concerning the işleme hakkı — the right of adaptation — may become relevant where a new product is derived from a protected work and is not independent of it. Article 22 may again become relevant where protected expression is actually reproduced. Depending on the manner in which the resulting content is made available, other economic rights may also enter the analysis. The statutory rights recognized under FSEK include adaptation, reproduction, distribution, performance and communication to the public. (Telif Hakları)
But a court should not treat similarity of style as equivalent to reproduction of a protected work. Copyright protects the legally protected expression of an author; it does not grant an unlimited monopoly over an artistic idea, genre, method or general aesthetic.
A Turkish Round Hill case should therefore separate three questions that public debate often merges: acquisition of the training material, use of that material during training, and infringing expression appearing in model outputs.
They need not produce the same legal result.
The anti-circumvention allegations reveal another interesting comparison
Round Hill’s complaints do not concern reproduction alone. The Anthropic action also pleads circumvention of protection measures and removal or alteration of copyright management information; the Suno action contains similar theories and additionally alleges contributory infringement against Bright Data. (Justia Dockets & Filings)
U.S. law has a developed statutory architecture for this purpose.
Section 1201 of the Digital Millennium Copyright Act prohibits circumvention of technological measures controlling access to protected works and also regulates circumvention technologies and services. Section 1202 separately prohibits specified forms of intentional removal or alteration of copyright management information, including information identifying a work, its author, its copyright owner and certain licensing information. (Telif Hakkı Ofisi)
Turkish law is not completely silent here either.
FSEK Article 72, in its current form following the 2021 amendment, concerns technological measures used to protect works, performances, phonograms, productions and broadcasts. It criminalizes specified commercial activities involving products, tools and services directed at disabling effective technological measures such as access controls, encryption and reproduction-control mechanisms. (Türkiye Büyük Millet Meclisi)
FSEK Additional Article 4 is even more directly comparable to the American CMI provisions. It prohibits the unauthorized removal or alteration of information concerning the work, the author, rights holders and conditions of use, as well as certain subsequent dealings with copies from which such information has been removed or altered. (Türkiye Büyük Millet Meclisi)
The statutes are not identical. In particular, the architecture of DMCA Section 1201 addresses the act of circumvention and trafficking rules in a more explicit and developed manner. The current wording of FSEK Article 72 is focused significantly on circumvention products, tools and associated commercial services. A Round Hill allegation that an AI company itself bypassed technical website barriers would therefore require a more careful Turkish characterization rather than a mechanical transplantation of the American Section 1201 claim.
But again, this demonstrates that the Turkish problem is not the complete absence of legal rules. It is the absence of rules designed around the AI data supply chain.
Bright Data exposes a liability question that Türkiye has not specifically solved for AI
The Suno case is particularly instructive because Round Hill did not sue only the developer of the model. It also sued Bright Data, alleging contributory infringement in connection with the acquisition of data used by Suno. (Justia Dockets & Filings)
American copyright law has developed doctrines of secondary liability through which an actor who did not personally commit every act of infringement may nevertheless face liability for materially contributing to another’s infringement under the required legal conditions.
FSEK does not provide an AI-specific allocation of responsibility among model developer, dataset supplier, scraping provider and downstream deployer.
That does not make supporting actors immune under Turkish law. Depending on the facts, general principles of liability remain relevant. Article 61 of the Turkish Code of Obligations provides for joint and several liability where several persons together cause the same damage or are responsible for it on different grounds. (Türkiye Büyük Millet Meclisi)
But the absence of an AI-specific chain-of-responsibility rule matters. A future case may require a Turkish court to decide, using rules written for another technological era, how responsibility should be divided between the company that requested the data, the company that scraped it, the company that prepared the dataset and the company that trained or deployed the model.
That is not an abstract problem. The Bright Data allegations put the issue directly on the table.
The remedies are also fundamentally different
The financial architecture of the two legal systems should not be overlooked.
Under 17 U.S.C. §504, a copyright owner may, subject to the statutory requirements, elect statutory damages instead of actual damages and profits. The amount can rise to $150,000 per work where willful infringement is established. U.S. copyright law additionally provides injunctive and impoundment remedies, while Sections 1201 and 1202 have their own civil-remedy structure. (Telif Hakkı Ofisi)
That per-work statutory-damages structure is one reason litigation involving hundreds or thousands of works can create enormous headline exposure in the United States.
FSEK follows a different model.
Turkish law provides actions aimed at stopping and removing infringement and permits monetary claims under its civil-remedy provisions. Of particular importance, Article 68 allows the right holder, where protected economic rights have been used without authorization, to claim up to three times the amount that could have been requested had a contract been concluded or the current market value determined under the statutory conditions.
A Turkish court would therefore not simply take “500 works × USD 150,000” and reproduce the American calculation. The valuation exercise would be conceptually different and could put the hypothetical licensing market for AI training at the centre of the Turkish damages analysis.
This difference may become commercially significant if collective AI-training licences develop into a measurable market.
U.S. and Turkish law would therefore approach the same facts differently
| Issue | United States | Türkiye |
| Initial copying | §106 reproduction right | FSEK Art. 22 reproduction right |
| General AI-training defence | §107 fair use permits case-specific balancing | No equivalent general fair-use clause |
| Commercial full-work training | Potentially defensible depending on the four-factor analysis and facts | Difficult to place within quotation or personal-use exceptions |
| Lawful versus pirated source | Highly relevant; Bartz separated transformative training from pirate-library acquisition | Source and authorization would be highly relevant, but there is no AI-specific lawful-source training regime |
| Technological circumvention | DMCA §1201 | FSEK Art. 72, but differently structured |
| Copyright management information | 17 U.S.C. §1202 | FSEK Additional Art. 4 provides a meaningful analogue |
| Data supplier / scraper liability | Secondary copyright liability doctrines | FSEK plus general liability principles, including TBK Art. 61; no AI-specific allocation |
| Damages | Actual damages/profits or statutory damages, potentially up to $150,000 per work for willfulness | FSEK civil remedies, including the Art. 68 mechanism of up to three times the relevant licence/market amount |
| AI-specific copyright framework | No comprehensive federal AI copyright statute, but extensive copyright case law and fair-use doctrine | No enacted horizontal AI Act and no specific AI-training/TDM copyright regime |
What Türkiye actually lacks
This is where I would locate the legislative deficit.
Türkiye does not primarily need an AI copyright provision because FSEK is incapable of recognizing copying. Article 22 already does that.
What it lacks is a legislative answer to the question when machine reproduction for AI development should be permitted, prohibited or licensed.
Current law does not specifically tell a developer whether training on lawfully accessible copyrighted material is permissible. It does not create a dedicated text-and-data-mining exception for general commercial AI training. It does not establish an opt-out mechanism through which right holders can reserve their works from AI training. It does not set a transparency standard requiring model providers to identify or meaningfully disclose copyright-relevant training sources. Nor does it specifically determine the legal consequence of using unlawfully acquired material even where the eventual machine-learning purpose might be characterized as innovative or transformative.
There is also no dedicated statutory answer for memorization, model weights, dataset retention, post-training deletion, provenance records or the division of responsibility between a scraper, dataset producer, model developer and deployer.
These are not minor procedural details. They determine whether the market develops around permission, collective licensing, opt-out, statutory exception or litigation.
As of 25 August 2026, Türkiye still has no enacted horizontal Artificial Intelligence Act. A “Yapay Zeka Kanun Teklifi” numbered 2/2234, submitted to the Turkish Grand National Assembly on 24 June 2024, remains at committee stage. The official summary describes an intended framework concerning safe, ethical and fair AI use, protection of personal data and privacy, and regulation of AI development and use. It is not currently enacted law. (Türkiye Büyük Millet Meclisi)
Its existence therefore does not solve a Round Hill-type dispute today.
If the case were before a Turkish court today
If substantially the same facts were proven in a case governed by Turkish law, the central question would probably be simpler at the beginning and more difficult at the end than in California.
It would be simpler because FSEK Article 22 gives the claimant a direct statutory route: demonstrate protected works, demonstrate unauthorized reproduction and establish the defendant’s responsibility for that reproduction.
It would be more difficult because Turkish law then offers neither the defendant nor the court a purpose-built doctrine for AI training.
Anthropic or Suno could not simply say: “Our use was transformative, therefore apply fair use.” There is no Section 107 in FSEK.
At the same time, the court should resist the opposite shortcut. It should not conclude that every instance in which a machine computationally processes a protected work necessarily establishes every form of copyright infringement claimed by the right holder. Dataset copying, training-time reproduction, model weights, memorized content and user-facing outputs are legally distinct stages and should be analysed separately.
That is precisely why legislation is required.
The absence of an AI Act does not create a legal vacuum. It creates legal asymmetry.
The right holder has an old but powerful rule: unauthorized reproduction is within the exclusive right.
The AI developer has old and narrowly designed exceptions that were never written with machine learning in mind.
And the judge is left between them.
Conclusion
Round Hill Music v. Anthropic and Round Hill Music v. Suno may ultimately become important American copyright cases, but their significance for Türkiye lies somewhere else.
They expose what happens when copyright legislation written around human acts of copying encounters industrial-scale machine learning.
The United States addresses that collision, imperfectly but flexibly, through fair use. Bartz already demonstrates the consequence: a court can find AI training transformative while simultaneously treating the acquisition of pirated source material as a separate infringement. Kadrey demonstrates the other side of the same doctrine: market harm can determine the result, and an AI company does not receive immunity merely because the disputed copying occurred during “training.”
Turkish law currently has no equivalent mechanism.
FSEK is strong enough to regulate the copy.
It is not yet precise enough to regulate the training relationship surrounding that copy.
If Türkiye legislates in this field, the central policy decision should therefore not be framed as “copyright versus artificial intelligence.” That would miss the real issue. The legislature must decide where permission is required, where machine analysis should be permitted, whether lawful access should matter, how right holders may reserve their works, what transparency AI providers owe them, and how liability should move through the data supply chain.
Until those questions are answered legislatively, a Round Hill-type case in Türkiye would be decided principally through FSEK provisions written long before generative AI existed.
Those provisions are capable of producing a judgment.
Whether they are capable of producing the right regulatory balance for an AI economy is a different question entirely.
Av. Asutay Duhan Meydan
Meydan AI Law
This article is a comparative legal analysis based on legislation and publicly available court records as of 25 August 2026. The Round Hill proceedings are pending, and the allegations discussed above have not yet been adjudicated on their merits.