WHERE DOES “FAIR USE” STAND IN TURKEY?

By Asutay Duhan Meydan
Attorney at Law | Meydan AI & Tech Law

An AI provider defending its training practices in Turkey cannot assume that an argument based on American fair use will survive translation into Turkish law. The same copying operation may encounter a different statutory structure, different limitations and a different route to liability.

Turkish copyright law contains no general equivalent of Section 107 of the United States Copyright Act. It nevertheless recognises specific limitations on exclusive rights and requires courts to distinguish protected expression from facts, ideas, methods and other material that cannot be monopolised through copyright.

The absence of a general fair use defence therefore strengthens certain infringement claims. It does not, however, relieve a claimant of the obligation to identify the protected work, prove entitlement to the relevant rights and demonstrate the acts allegedly constituting infringement.

The central difficulty is the imbalance between a broadly drafted reproduction right and exceptions designed for specific and traditionally defined forms of use. Turkish copyright law can reach the copying involved in AI training, but it does not yet provide a clear and comprehensive rule determining when that copying should be permitted.

The statutory quotations below are English translations of the relevant Turkish provisions. The Turkish legislative text remains authoritative.

This analysis should be read together with our earlier examination of how Turkish copyright law applies to AI training disputes.

I. The Statutory Starting Point: Reproduction

Law No. 5846 on Intellectual and Artistic Works, commonly referred to as FSEK, establishes a broad reproduction right. Article 22(1) provides:

“The right to reproduce the original or copies of a work, in any form or by any method, wholly or partly, directly or indirectly, temporarily or permanently, belongs exclusively to the author.”

The provision expressly covers temporary and permanent reproduction. An AI provider cannot establish lawfulness merely by explaining that protected material was processed in computer memory, converted into machine-readable form or deleted after training.

The first question is whether protected expression was reproduced. If it was, the court must determine whether the act was authorised or covered by a statutory limitation.

The training process must therefore be divided into legally identifiable stages: obtaining the source material, assembling the dataset, producing technical copies, conducting the training operation, retaining the resulting model and supplying outputs to users. These stages may engage different rights and should not be treated as a single indivisible act.

A model may produce no infringing output even though unauthorised reproductions occurred during training. Conversely, an unlawful training copy does not mean that every subsequent output infringes copyright.

Whether model weights themselves contain reproductions of particular works is a technical and evidential question. It should not be presumed merely from the fact that a work was included in the training process.

This distinction also prevents the indiscriminate accumulation of claims. Article 21 regulates the adaptation right, while Article 25 regulates communication to the public. Neither is automatically infringed whenever Article 22 is engaged. Each allegation requires its own factual and legal basis.

II. Where Turkish Law Places Permitted Use

The closest functional counterpart to fair use is found in the statutory limitations contained in FSEK. These include quotation, personal use, educational uses and certain uses serving the public interest. However, they do not operate as an open-ended balancing clause.

Under Section 107 of the United States Copyright Act, courts consider the purpose and character of the use, the nature of the work, the amount and substantiality taken and the effect upon the potential market.

Turkish law follows a different structure. The user must ordinarily identify an applicable statutory permission and satisfy its conditions. Describing a use as transformative, innovative or socially beneficial is not sufficient by itself.

Quotation Under Article 35

Article 35 provides:

“Quotations from a work are permissible in the following cases:

1. The inclusion of certain sentences and paragraphs from a work made available to the public in an independent scientific and literary work;
2. The inclusion, in an independent musical work, of portions of a published musical composition in the nature of themes, motifs, passages and ideas;
3. The inclusion of works of fine art made available to the public and other published works in a scientific work, to an extent justified by the purpose and for the purpose of elucidating its contents;
4. The display, by projection or similar means, of works of fine art made available to the public in scientific conferences or lectures, for the purpose of elucidating the subject.

The quotation must be made in an identifiable manner. In scientific works, in addition to the name of the work and its author, the place from which the quoted portion was taken must be indicated.”

These requirements concern the nature of the new work, the purpose of incorporation, the extent of the material taken and the manner in which the source is identified.

The provision cannot be reduced to a fixed numerical threshold. A short quotation is not necessarily lawful, and a longer use is not automatically unlawful. The relationship between the purpose and the amount taken must be assessed in the circumstances of each case.

For general-purpose AI training, the difficulty is structural. A training dataset does not automatically qualify as an independent scientific or literary work. Statistical learning does not necessarily establish the explanatory relationship required by Article 35. Recording the source internally also does not, by itself, satisfy the conditions governing quotation.

A provider relying on Article 35 must therefore explain how its particular use falls within the language and purpose of the provision. The general proposition that artificial intelligence development involves scientific activity will not be enough.

Personal Use Under Article 38

Article 38(1) states:

“All intellectual and artistic works may be reproduced exclusively for personal use without pursuing a profit-making purpose. However, such reproduction may not, without a justified reason, prejudice the legitimate interests of the right holder or conflict with the normal exploitation of the work.”

The fact that copies are not distributed to the public does not bring the use automatically within Article 38. A private corporate server used to develop a commercial model does not become a sphere of personal use merely because the training corpus is inaccessible to customers.

Nor does the use of the word “training” bring machine learning within every educational exception. Article 33 addresses performance for face-to-face education under specific conditions. Article 34 regulates selections and compilations intended for education. The technical language of AI cannot expand the scope of these statutory provisions.

Legislation and Judicial Decisions Under Article 31

Article 31 is particularly important for legal AI providers:

“The reproduction, distribution, adaptation or exploitation in any manner of officially published or announced laws, Presidential decrees, regulations, communiqués, circulars and judicial decisions is free.”

This provision supplies a clear statutory basis for using the legal materials it identifies.

However, a judicial decision and a publisher’s treatment of that decision are not necessarily the same material. Original summaries, annotations, classifications and commentaries may be protected separately. Database rights and contractual access restrictions may also require an independent assessment.

Article 31 therefore permits the use of official judicial decisions. It does not automatically authorise the extraction of every commercial product in which those decisions appear.

III. The Position Emerging From Turkish Case Law

There is currently no publicly accessible Turkish judgment directly resolving whether generative AI training constitutes permissible use under FSEK. Nevertheless, two decisions of the Court of Cassation’s General Assembly of Civil Chambers demonstrate how Turkish courts approach protected expression and the quotation exception.

Court of Cassation General Assembly of Civil Chambers, E.2017/2628, K.2021/865, 29 June 2021

The dispute concerned the inclusion of artworks in an independent publication addressing a scientific and cultural subject.

The General Assembly examined the independence of the later work, the explanatory function of the images, proportionality and attribution. It also accepted that a scientific work need not address an exclusively academic audience.

In the circumstances of the case, the artworks performed an ancillary function consistent with the purpose of the new work. The use was accepted as falling within the quotation exception.

The decision demonstrates that Article 35 contains genuine interpretative space. Turkish courts do not necessarily apply quotation rules through a purely mechanical measurement of the material used.

However, the judgment does not create a general fair use doctrine. It supports a contextual interpretation within the boundaries of an existing statutory exception. It cannot establish that an industrial training corpus is lawful merely because AI development has a scientific dimension.

Court of Cassation General Assembly of Civil Chambers, E.2022/219, K.2023/631, 14 June 2023

The dispute concerned material taken from a biographical and historical publication and used in a documentary.

The General Assembly required a proper examination of both the original work and the allegedly appropriated portions. It distinguished protected authorial expression from scientific facts, historical events, abstract ideas and methods.

The Court also required the claimant’s entitlement to the economic rights to be established through the underlying contractual documents. A general assertion of ownership or a limited administrative document was not sufficient to complete the examination.

The judgment was a reversal based on insufficient examination. It was not a final determination that the contested use was lawful.

Applied to AI disputes, the decision requires precision. Similarity involving facts, methods or historical information does not establish the appropriation of protected expression. The claimant must also prove entitlement to the economic rights invoked.

However, this distinction does not eliminate a separately established reproduction of a protected work during training. The provider must answer the copying allegation at the stage at which the copying occurred.

IV. The Three-Step Test Does Not Create a General Fair Use Defence

Article 13 of the Agreement on Trade-Related Aspects of Intellectual Property Rights provides:

“Members shall confine limitations or exceptions to exclusive rights to certain special cases which do not conflict with a normal exploitation of the work and do not unreasonably prejudice the legitimate interests of the right holder.”

The test has three cumulative elements:

1. The limitation must concern a certain special case.
2. It must not conflict with the normal exploitation of the work.
3. It must not unreasonably prejudice the legitimate interests of the right holder.

The three-step test guides the scope and interpretation of copyright limitations. It does not expressly give each user an independent right to create a new exception whenever immediate financial loss cannot be demonstrated.

A Turkish court may consider proportionality, constitutional freedoms and the competing interests involved when interpreting an existing provision. However, such interpretation cannot automatically replace the legislature or introduce the complete American fair use doctrine into FSEK.

Arguments based on scientific freedom, public interest or the technical necessity of temporary copies may carry greater weight where the material was accessed lawfully, the copying was strictly necessary, the copies were temporary and safeguards prevented redistribution.

The same arguments become weaker where a commercial provider permanently accumulates protected works, cannot identify their origin or uses them to develop a product competing within the same economic environment.

V. A Defensible Position Requires More Than a Fair Use Label

An AI provider’s strongest defence will generally begin with authorisation, entitlement, protected subject matter and proof of actual use.

Article 52 of FSEK provides:

“Contracts and dispositions concerning economic rights must be in writing, and the rights forming their subject matter must be specified separately.”

A dataset supplier’s general assurance that content is “cleared” must therefore be tested against the actual rights granted. The identity of the licensor, the specified economic rights, the permitted purpose, the territory, the duration and any right of sublicensing must be examined.

An indemnity may allocate financial risk between the supplier and the provider. It does not create missing copyright authorisation against the actual right holder.

Territorial analysis is equally important. Under Article 23 of Law No. 5718, intellectual property rights are governed by the law of the country for which protection is sought.

Proceedings filed in Turkey do not automatically make Turkish law applicable to every training operation carried out abroad. Conversely, locating servers outside Turkey does not eliminate claims concerning relevant conduct or protection sought in Turkey. The alleged acts, the requested protection and the applicable law must be separated territorially.

The practical consequence is evidential. Providers should be able to identify their source material, permissions, relevant model versions and processing operations. A general statement that the model learned only patterns cannot replace an account of the copies made in order to conduct that learning.

VI. What Outcome Should AI Providers Expect in Turkey?

Where a claimant proves entitlement to protected works and establishes that those works were reproduced without authorisation for commercial AI training, the current statutory structure gives the claimant a substantial basis for relief.

The provider should expect scrutiny under Article 22 and may face claims for cessation and prevention of infringement, compensation, the payment mechanism under Article 68 and interim measures. A defence based solely on the transformative nature of the technology is unlikely to answer the statutory claim.

Article 68 permits a claim based on the fee that could have been requested had a contract been concluded or on the market value determined under FSEK, subject to the statutory multiplier. It does not create a fixed American-style statutory award for every work.

The nature and scope of the remedy must nevertheless correspond to the proven infringement. A finding concerning particular training copies should not automatically determine the legality of every output or justify the destruction or prohibition of an entire model.

The provider’s strongest response will be a documented account addressing authorisation, the claimant’s entitlement, the protected nature of the material, actual use, applicable law, causation and the proportional scope of the requested remedy.

VII. Conclusion

Fair use has no general statutory home in Turkish copyright law.

Some of its underlying concerns appear through specific limitations, the distinction between ideas and protected expression, the three-step test and the balancing of competing rights. Their legal force nevertheless depends on the applicable statutory provision and the facts of the dispute.

Existing Turkish law can reach unauthorised copies made during AI training. Its unresolved task is to determine, with sufficient certainty, the conditions under which particular training activities should be permitted and the extent to which courts should intervene when infringement is established.

For AI providers, the distinction has immediate consequences. A defence built around an imported fair use label is vulnerable. A defence built around the actual work, the actual right, the actual copying operation and the actual legal authority for that use is considerably stronger.

Asutay Duhan Meydan
Attorney at Law
Meydan AI & Tech Law

Sources

Legislation and International Instruments

1. Law No. 5846 on Intellectual and Artistic Works (FSEK). Official legislation page of the General Directorate of Copyright.
2. Turkish Code of Obligations, Law No. 6098. Consolidated text.
3. Law No. 5718 on Private International Law and International Civil Procedure. Consolidated text.
4. Agreement on Trade-Related Aspects of Intellectual Property Rights. World Trade Organization.
5. Berne Convention for the Protection of Literary and Artistic Works. WIPO treaty text.
6. United States Copyright Act, 17 U.S.C. § 107. United States Copyright Office.

Turkish Case Law

7. Court of Cassation, General Assembly of Civil Chambers, E.2017/2628, K.2021/865, 29 June 2021. Full judgment.
8. Court of Cassation, General Assembly of Civil Chambers, E.2022/219, K.2023/631, 14 June 2023. Full judgment.

Official Guidance

9. Republic of Türkiye, Ministry of Culture and Tourism, General Directorate of Copyright, General Questions. Official guidance.
10. Republic of Türkiye, Ministry of Culture and Tourism, General Directorate of Copyright, Remedies Available in the Event of Copyright Infringement. Official guidance.

All online sources were last accessed on 17 September 2026.