The New York Times v. OpenAI: How the Same Copyright Dispute Could Be Decided Under Turkish Law

By Av. Asutay Duhan Meydan – Attorney at law
Meydan AI Law

Introduction

The dispute between The New York Times and OpenAI is frequently reduced to an oversimplified question:

May an artificial intelligence company train its models on material that is publicly accessible on the internet?

That formulation is incomplete.

The real dispute concerns a chain of legally distinct acts: obtaining copyrighted material, creating and maintaining datasets, making temporary or permanent copies during model training, incorporating information derived from those works into a commercial system, generating outputs that may reproduce protected expression, and deploying products that may compete with the copyright owner in existing or developing markets.

Each stage must be examined separately.

Public availability does not, by itself, amount to a licence. At the same time, copyright does not grant a publisher ownership over facts, historical events or general knowledge. The legal difficulty lies precisely between those two propositions.

This article examines The New York Times v OpenAI under Turkish copyright law, with particular focus on AI training, reproduction rights, database protection and potential remedies under FSEK.

This article first examines the principal allegations and defences in The New York Times Company v. Microsoft Corporation and OpenAI under United States law. It then considers how a materially similar dispute might be pleaded and decided before a Turkish court.

This is not a prediction of the outcome of the pending United States proceedings. Nor is it a definitive opinion on the liability of any particular company. The purpose is to identify the legal questions that a Turkish court would likely be required to answer if a national publisher brought an equivalent claim against a generative AI provider.

1. What The New York Times Is Actually Alleging

The Times filed its action against OpenAI and Microsoft on 27 December 2023. Its pleaded claims originally included direct copyright infringement, contributory and vicarious infringement, alleged violations of the Digital Millennium Copyright Act, common-law misappropriation and trademark dilution.

The complaint is built around two separate stages.

The first is the training stage. The allegation is that the defendants collected and stored large quantities of internet content, including Times articles, placed those materials into training datasets and used copies of the articles during the model-training process.

The second is the output stage. The Times alleges that, in response to certain prompts, the relevant products can reproduce substantial portions of its articles verbatim or nearly verbatim. It further alleges that search-integrated products may provide extensive paraphrases or quotations without directing the user to the Times website in the manner of a conventional search engine.

This distinction is fundamental. An output may be non-infringing even where an unauthorised reproduction occurred during dataset construction or training. Conversely, the lawfulness of the original training process would not necessarily excuse a later output that reproduces a substantial part of a protected work.

The Times’ economic theory is also broader than simple plagiarism. In substance, it says:

1. The Times invests substantial financial and institutional resources in reporting, editing and verifying original journalism.
2. That journalism is monetised through subscriptions, advertising and licensing.
3. OpenAI allegedly used those materials without purchasing the relevant rights.
4. OpenAI then commercialised products capable, in certain circumstances, of delivering the value of that journalism to users.
5. Those products may reduce the need for users to visit the Times website, subscribe to its services or obtain a licence.

Accordingly, the alleged injury is not limited to the taking of words. It includes the alleged appropriation of an expensive journalistic input and the deployment of that input within a competing commercial product.

2. The United States Copyright Framework

2.1 Prima facie infringement

Under United States copyright law, the copyright owner generally holds exclusive rights that include reproduction and the preparation of derivative works. A person who violates those exclusive rights may be liable under 17 U.S.C. § 501.

In broad terms, the Times must therefore establish ownership of valid copyrights and the copying of protectable elements of the relevant works.

News articles require careful treatment in this respect. Facts themselves are not protected by copyright. Neither are ideas, historical events or information merely because a publisher discovered them first. What copyright may protect is the author’s original expression: the particular selection, organisation, phrasing, narrative structure and presentation of those facts.

The federal court has already applied this distinction in the consolidated litigation. In April 2025, it dismissed certain claims concerning summaries of articles published by the Center for Investigative Reporting because those outputs presented the underlying facts with different phrasing, structure, style and tone. The court found that those summaries were not substantially similar to protectable expression in the original articles.

That ruling is significant. It demonstrates that a detailed AI-generated summary is not automatically an infringing copy. The relevant comparison is not between the information conveyed by the two texts, but between their protected expressive elements.

2.2 Fair use

Even where copying has occurred, 17 U.S.C. § 107 may excuse it as fair use. The statute directs courts to consider:

1. the purpose and character of the use, including its commercial nature;
2. the nature of the copyrighted work;
3. the amount and substantiality of the portion used; and
4. the effect of the use upon the actual or potential market for the copyrighted work.

No single factor is automatically decisive.

OpenAI’s central position is that using publicly accessible material to train a model is a transformative technological use. It argues that the model does not ordinarily function as an article database and does not sell copies of Times journalism. Rather, it uses a broad corpus to identify linguistic and informational relationships and to produce a system capable of performing many different tasks.

OpenAI also argues that the examples relied upon by the Times were not representative of ordinary product use. In its motion to dismiss, it alleged that the outputs were obtained through tens of thousands of attempts, the use of specially designed prompts and the provision of parts of the articles that the tester was attempting to extract. These remain allegations made by OpenAI in the litigation; they are not findings of fact.

The Times, by contrast, is likely to argue that the use is commercial, that entire works were copied, that the system can reproduce protected material and that OpenAI’s products operate in markets that overlap with the publisher’s subscription, search, syndication and licensing activities.

2.3 The importance—and limits—of the Google Books precedent

OpenAI relies heavily on decisions such as Authors Guild v. Google, in which the scanning of millions of books and the display of limited text snippets through Google Books were held to be fair use.

The Google Books service served a search and research function. It did not ordinarily provide the public with complete substitutes for the books. The Second Circuit regarded the search functionality as highly transformative and found that the limited snippets did not create a significant market substitute for the protected works.

The analogy is relevant, but it is not conclusive.

A conventional book-search system tells the user that a work exists and may assist the user in locating it. A generative system may provide a self-contained answer. Where the answer supplies the substance of a protected article, the user may have no practical reason to visit the publisher.

That factual difference may become decisive under the fourth fair-use factor.

The Supreme Court’s decision in Andy Warhol Foundation v. Goldsmith also warns against treating “transformative use” as a universal answer. The Court emphasised that purpose must be evaluated in relation to the particular challenged use and whether the secondary use serves a substantially similar commercial purpose.

The correct question is therefore not whether generative AI is transformative in the abstract. The question is whether the particular use of the particular copyrighted works, at the particular stage challenged by the plaintiff, is sufficiently different and justified.

2.4 The developing AI-training decisions

In 2025, federal district courts in Bartz v. Anthropic and Kadrey v. Meta accepted fair-use arguments concerning particular uses of copyrighted books in model training. However, those decisions do not establish that all AI training is lawful.

In Bartz, the court distinguished the transformative use of books for training from the separate creation of a permanent library using unlawfully obtained copies. The source and method by which a work was acquired therefore remained legally important. In Kadrey, Meta prevailed on the record presented by those particular plaintiffs, including deficiencies in their evidence of relevant market harm.

Neither decision binds the Southern District of New York in the Times litigation. More importantly, neither removes the need for a work-specific, use-specific and market-specific analysis.

The United States Copyright Office has adopted the same cautious position. Its 2025 report concluded that some uses of copyrighted works in generative AI training may qualify as fair use and others may not. It specifically distinguished non-commercial research and non-substitutional uses from commercial exploitation of large quantities of copyrighted expression to produce competing content, particularly where material was obtained through unlawful access.

3. How a United States Court Could Rule for The New York Times

A decision favouring the Times could rest on the following reasoning.

First, the court could find that OpenAI made unauthorised reproductions of complete Times articles when acquiring, storing, processing or repeatedly presenting those articles to the model during training.

Second, it could find that the use was not sufficiently transformative in relation to the challenged commercial purpose. A model may be technologically innovative while still using a particular work for a purpose that materially overlaps with the publisher’s purpose: supplying users with informative written accounts of current and historical events.

Third, the use of entire articles could weigh against OpenAI. Although copying an entire work may sometimes be justified by a genuinely transformative technical purpose, the necessity and proportionality of that copying would need to be demonstrated.

Fourth, market harm could weigh heavily in the Times’ favour. The publisher can identify established markets for subscriptions, syndication, archival access and content licensing. It may also attempt to prove an emerging and reasonable market for AI-training licences. The Copyright Office has recognised that existing or likely-to-develop licensing markets may be relevant under the fourth factor.

Fifth, the court could distinguish Google Books on the basis that Google displayed controlled snippets and generally directed users towards the relevant books, while a generative product can provide a complete, fluent and immediately usable response within its own interface.

Finally, repeated verbatim or near-verbatim outputs—if shown to arise under realistic conditions rather than extraordinary extraction attacks—could support the conclusion that the system does not merely learn unprotected facts and linguistic patterns. It may retain and commercially reproduce protected expression.

Under that reasoning, the court could conclude that OpenAI’s commercial benefit was obtained through copying that exceeded what was reasonably necessary for the claimed transformative purpose and caused, or was likely to cause, harm in markets belonging to the copyright owner.

4. How a United States Court Could Rule for OpenAI

A decision favouring OpenAI could be equally structured.

The court could first characterise model training as an intermediate computational use that differs fundamentally from the expressive purpose of journalism. The Times writes articles to communicate particular reports to human readers. OpenAI may argue that the training copies were used to calculate statistical relationships and develop a general-purpose model, rather than to republish those articles.

Second, the predominantly factual character of news reporting could weigh more favourably towards fair use than the copying of novels, music or visual art.

Third, copying the complete work could be regarded as technically necessary for the transformative training purpose, just as complete copies were tolerated in earlier search, indexing and software cases.

Fourth, OpenAI could prevail by demonstrating that ordinary outputs are not substitutes for Times articles. The system may frequently provide short answers, general summaries or new text that does not reproduce protectable expression. The possibility of obtaining infringing material through exceptional, adversarial or terms-violating prompts might not establish that the ordinary function of the product harms the market for the original works.

Fifth, the Times would still be required to prove causation and market injury rather than relying on the general proposition that generative AI competes for attention. A decline in website traffic is not automatically a copyright injury. The alleged loss must relate to a market that copyright law recognises and protects.

The court could therefore find that the training use is transformative and fair, while separately preserving liability for individual outputs that reproduce substantial protected passages. This divided result is legally possible and, in my view, more plausible than treating the entire system as one indivisible act.

5. The Present Procedural Position

The April 2025 ruling did not determine whether OpenAI infringed the Times’ copyrights. It decided motions to dismiss.

At that procedural stage, the court was required to assume well-pleaded factual allegations to be true and determine whether they stated legally plausible claims. It allowed important direct and contributory copyright claims to continue, while dismissing the common-law misappropriation claim and certain other claims in whole or in part.

In June 2026, the Times sought to amend and narrow parts of its case following a change in the governing standard for contributory infringement. It proposed dismissing an alternative contributory-infringement count and its trademark-dilution count while revising its contributory claim against Microsoft.

As of 4 August 2026, the dispute remains pending without a final judgment on copyright liability or fair use. Recent proceedings have included contested discovery issues and allegations by the publisher plaintiffs concerning OpenAI’s handling of relevant evidence. OpenAI disputes those allegations. They should not be treated as established facts unless and until the court makes the necessary findings.

6. A Turkish Version of the Dispute

AI provider liability under Turkish law

Assume that a major Turkish news publisher maintains:

• an extensive digital archive;
• subscription-only investigative journalism;
• content-licensing and syndication operations;
• a database developed through substantial investment; and
• express terms reserving the right to reproduce or commercially reuse its content.

The publisher alleges that OpenAI, directly or through third-party datasets, copied a substantial portion of this archive, used the material in the development or fine-tuning of its models, and subsequently provided Turkish users with verbatim passages, detailed substitutes for subscription articles and inaccurate texts falsely attributed to the publisher.

The publisher seeks an injunction, cessation of the alleged infringement, compensation calculated by reference to a hypothetical licence and recovery of the commercial benefit obtained from the use.

The Turkish court would need to avoid deciding the dispute through general statements about whether AI “learns like a human being.” That analogy does not resolve the legal issue.

The court would have to ask:

• Which specific works were used?
• Who owns or is authorised to exercise the relevant rights?
• How were those works obtained?
• What copies were created?
• Where did the legally relevant acts occur?
• Did the model reproduce protected expression or merely unprotected information?
• Which statutory exception is relied upon?
• What commercial market was affected?
• What evidence connects the alleged use to the claimed loss?

Those questions require technical and financial evidence.

7. The Principal Claims Under Turkish Law

7.1 Protection of individual articles

Under Law No. 5846 on Intellectual and Artistic Works—the FSEK—a text must bear the individuality of its author to qualify as a protected work. Written scientific and literary works may fall within the statutory categories of protection.

The facts contained in a news report would not, as such, belong exclusively to the publisher. However, an original investigative article may contain protectable elements in its language, structure, selection, sequence and presentation. The claimant would therefore need to identify the protected expression rather than merely asserting ownership of the reported facts.

The publisher would also need to establish its chain of title or its authority to exercise the relevant economic rights. Moral rights, including attribution and protection against unauthorised modification, may continue to belong to the individual journalist or author, depending on the contractual and employment structure.

7.2 The reproduction right

Article 22 FSEK defines reproduction broadly. It covers direct or indirect, temporary or permanent copying by any method.

This language is capable of covering copies made during web scraping, dataset preparation, storage, tokenisation, model training or fine-tuning. It would not be sufficient for the defendant simply to argue that the final model does not contain human-readable copies of the articles.

The first question would be whether legally relevant copies were made during the process. The second would be whether those copies were authorised or fell within a statutory limitation.

In my assessment, Article 22 would be the central provision in a Turkish claim concerning the training stage.

7.3 Adaptation and communication to the public

Where a generated output reproduces or rearranges protected expression, the claimant could also invoke the adaptation right under Article 21 and, depending on the method of deployment, the right of communication to the public under Article 25.

Not every summary would constitute an adaptation or infringing reproduction. A summary written in materially different language and containing only unprotected facts may fall outside the copyright owner’s exclusive rights.

However, an output that reproduces substantial passages, follows the distinctive structure of the original or provides the article in a slightly altered form would present a materially stronger claim.

7.4 Attribution, alteration and hallucinated publications

Articles 15 and 16 FSEK protect the author’s interest in attribution and against unauthorised alteration of the work.

These provisions could become relevant where a system:

• reproduces an article without identifying its author;
• attributes an altered version to the journalist;
• adds fabricated passages while presenting the result as authentic;
• or falsely states that the publisher reported something it never published.

The claimant would nevertheless need to distinguish between interference with an existing work and the creation of an entirely separate false statement. A hallucinated article that was never written may raise personality, reputation, trademark or unfair-competition issues without necessarily constituting an alteration of a copyrighted work.

7.5 Newspaper, news-reporting and quotation exceptions

This is where the Turkish analysis differs sharply from the United States framework.

Turkish law does not contain an open-ended fair-use doctrine equivalent to 17 U.S.C. § 107. It instead contains specifically regulated exceptions.

Article 35 governs quotation and requires, among other conditions, that the quotation be limited, identifiable and connected to its lawful purpose.

Article 36 permits the quotation of daily news and contains specific rules for certain articles concerning current social, political and economic matters. It also requires the identification of the relevant newspaper, journal, agency, date and author.

Article 37 permits the use of parts of works in connection with current events, provided that the use remains within the scope of informing the public and does not prejudice the legitimate interests of the right holder or conflict with normal exploitation of the work.

Article 38 permits reproduction for personal use without a profit-making purpose, subject again to the right holder’s legitimate interests and normal exploitation. A commercial AI provider could not plausibly convert its own industrial-scale reproduction into “personal use” merely because the eventual end user is an individual.

These exceptions could justify particular outputs. They do not appear to grant a general authorisation to collect an entire commercial archive and use it for model development.

In particular, an exception designed to permit reporting on a current event should not automatically be interpreted as permission to create a permanent commercial training corpus containing years of archived journalism.

7.6 Database rights

A publisher’s strongest Turkish claim might not be limited to copyright in individual articles.

Under Article 6 FSEK, an original database may be protected where the selection and arrangement of its contents bears the required individuality. Separately, Additional Article 8 provides a sui generis right to a database producer that has made a qualitatively or quantitatively substantial investment in obtaining, verifying or presenting the contents of the database.

That producer may prevent the permanent or temporary transfer of all or a substantial part of the database to another medium and its distribution or communication to the public.

A systematic extraction of a publisher’s archive could therefore trigger two different inquiries:

1. whether individual articles were reproduced; and
2. whether all or a substantial part of an investment-protected database was extracted or reutilised.

This distinction may be particularly important where the defendant argues that many individual news items contain limited copyrightable expression. Even where the underlying data is not protected as such, the systematic taking of a substantially invested database may still engage Additional Article 8.

7.7 Unfair competition

The publisher could also plead unfair competition under Articles 54–56 of the Turkish Commercial Code.

Article 54 prohibits deceptive or otherwise dishonest commercial conduct affecting relations between competitors, suppliers and customers. Article 55 includes the unauthorised exploitation of another person’s work product and, more specifically, the taking and exploitation of market-ready work products through technical reproduction methods without an appropriate contribution of one’s own.

Whether model training falls within Article 55(1)(c)(3) has not been definitively resolved. Nevertheless, a publisher could argue that the provider technically reproduced commercially ready journalistic products, incorporated their economic value into its own service and avoided the production or licensing costs borne by lawful market participants.

OpenAI would respond that the model is not a technical reproduction of the publisher’s ready-to-market articles, but an independently developed general-purpose system involving substantial investment, engineering and original contribution.

A court should not use unfair-competition law merely to recreate a copyright monopoly over unprotected facts. The claim would need an additional element of commercial dishonesty or unfair exploitation beyond conduct already addressed—and potentially permitted—by copyright law.

Article 56 allows claims for determination, prohibition, removal of the resulting situation and, where fault is established, damages. The court may also take into account the benefit that the defendant could have obtained through the unfair conduct.

7.8 Tort and unjust enrichment

The claimant could additionally rely on Article 49 of the Turkish Code of Obligations, which requires a person who culpably and unlawfully causes damage to compensate that damage.

Articles 50 and 51 govern proof and assessment of loss. The claimant must establish damage and fault, although the court may estimate loss equitably where its precise amount cannot be fully proven.

An unjust-enrichment claim under Article 77 could also be presented on a subsidiary basis where the defendant allegedly enriched itself from the claimant’s assets or labour without lawful cause.

These general provisions should not replace the more specific structure of the FSEK. They may, however, address residual economic injury not fully remedied by the special provisions, provided that double recovery is avoided.

8. Evidence and Territoriality

The greatest practical difficulty for a Turkish publisher would not necessarily be identifying legal provisions. It would be proving the facts.

The claimant would need evidence capable of showing that its works were included in a relevant dataset or otherwise copied during a legally relevant process. Similar outputs alone may be evidence of access or memorisation, but they do not automatically prove the complete history of data acquisition and training.

The court would likely require expert examination of:

• dataset documentation;
• scraping and acquisition records;
• archived copies;
• training and fine-tuning processes;
• model and output logs;
• reproducibility of the disputed outputs;
• the prompts required to obtain them;
• similarity between outputs and protected expression;
• filtering and anti-memorisation measures;
• publisher opt-out signals and access restrictions;
• existing content licences;
• and evidence of traffic, subscription or licensing losses.

Article 76 FSEK permits the court, where the claimant has submitted evidence sufficient to create a strong conviction as to the allegation, to require the user of protected works to produce documents showing that the necessary permissions were obtained or lists of the works used. Failure to produce the relevant documentation may create a presumption of unlawful use.

A separate issue is territorial connection. Training may have taken place on servers outside Türkiye. A Turkish claimant would therefore need to identify the acts for which protection is sought under Turkish law and their connection with Türkiye.

Article 23 of the Turkish International Private and Procedural Law provides that intellectual-property rights are governed by the law of the country under whose law protection is sought. Claims concerning effects in the Turkish market, Turkish rights, extraction from a Turkish database and outputs supplied to users in Türkiye could provide a basis for applying Turkish protection, but the precise territorial scope would remain a fact-sensitive issue.

A Turkish court should not assume jurisdiction over every stage of a global model-development process merely because the completed service is accessible in Türkiye. Equally, the provider should not be able to avoid all Turkish-law consequences simply by locating its servers abroad while deliberately supplying the disputed outputs to the Turkish market.

9. A Possible Turkish Judgment in Favour of the Publisher

A Turkish court could uphold the claim on reasoning substantially as follows:

The disputed articles contain original expression and qualify as protected works under the FSEK. The claimant has established its entitlement to exercise the relevant economic rights.

The evidence demonstrates that the defendant systematically copied complete articles and a substantial portion of the claimant’s digital archive during dataset creation and model development. These acts fall within the broad reproduction right under Article 22.

The defendant has not established that the mass and permanent commercial use falls within the quotation, news-reporting or personal-use exceptions. Those provisions permit limited uses serving specifically defined purposes. They do not authorise the wholesale ingestion of a commercial archive into a general-purpose model.

The systematic extraction also infringes the claimant’s rights as a database producer under Additional Article 8 because the archive was developed and maintained through substantial investment and a substantial part of its contents was transferred for commercial use.

The defendant’s products produced, under reasonably foreseeable conditions, outputs containing substantial protected passages and outputs capable of serving as substitutes for subscription or licensed articles. These uses exceed the communication of unprotected facts.

The defendant obtained a commercial benefit from the unauthorised exploitation, while the claimant lost licensing opportunities and faced interference with the normal exploitation of its journalism.

On that basis, the court could order cessation and prevention of the infringement under Articles 66 and 69 FSEK. Depending on the relief sought and the evidence, it could also award the amount calculated under Article 68 by reference to up to three times the fee that would have been payable if a lawful licence had been obtained, together with compensatory or profit-based remedies available under Article 70.

Article 68 should not be described as American-style punitive damages. It is a statutory calculation mechanism tied to the hypothetical or established licensing value of the unauthorised use.

Where the risk of continuing or imminent harm is sufficiently established, the court could also consider provisional relief under Article 77, including an order requiring specified conduct to cease pending the final decision. Any such measure would need to be proportionate and technically enforceable.

10. A Possible Turkish Judgment in Favour of OpenAI

The Turkish court could dismiss the claim for materially different reasons.

First, it could find that the claimant failed to identify which protected articles were actually included in a relevant training dataset. General evidence that the model knows facts previously reported by the publisher would not prove copying of protected expression.

Second, the claimant might fail to establish ownership or authority regarding the specific economic rights asserted. The fact that an article was published on the claimant’s website would not, in every case, prove that the claimant holds all rights required for every form of relief.

Third, the court could find that the disputed outputs conveyed only facts, ideas and general information using different words, organisation and structure. In that event, there would be no reproduction or adaptation of protected expression.

Fourth, the outputs could fall within the specific quotation or news-reporting exceptions, particularly where they are limited, attributed, proportionate and connected to informing users about a current event.

Fifth, the claimant could fail to prove that a substantial part of its database was extracted. The existence of scattered articles within a vast training corpus would not necessarily establish qualitative or quantitative substantiality under Additional Article 8.

Sixth, the claimant might be unable to establish a sufficient territorial connection between the alleged training copies and Türkiye.

Seventh, the asserted economic loss might remain speculative. A publisher cannot recover the entire commercial value of an AI product merely because some of its works may have contributed to a training corpus containing billions of documents. The loss, hypothetical licence and defendant’s benefit would need to be connected to the works and conduct actually proven.

Finally, the court could reject the proposition that model parameters themselves constitute copies of the claimant’s works in the absence of technical evidence demonstrating that protected expression can be reconstructed from the model.

On those grounds, the court could dismiss the training-related claim while still ordering the removal or prevention of a limited number of outputs shown to reproduce protected passages.

11. My Assessment

The New York Times v OpenAI dispute therefore provides a useful test case for examining how Turkish copyright law may respond to commercial AI training.

In my view, the Turkish dispute would be more difficult for an AI provider in one respect and more difficult for the publisher in another.

It would be more difficult for the provider because Turkish copyright law does not contain the broad, open-ended fair-use balancing mechanism available under United States law. A defendant would have to identify a specific statutory exception or demonstrate that the conduct did not fall within the copyright owner’s exclusive rights in the first place.

The argument that “the use is innovative and socially beneficial” would not, on its own, create a new exception to Article 22 FSEK.

It would be more difficult for the publisher because abstract objections to the use of journalism in artificial intelligence would not be sufficient. The publisher would still need to establish protected works, standing, copying, territorial connection and legally cognisable loss.

The fact that an AI product is commercially successful does not prove that every item in its training corpus generated recoverable profit for every copyright owner.

The strongest Turkish case would therefore not be framed as:

OpenAI read our publicly available news and learned from it.

It would be framed more precisely:

OpenAI, without contractual or statutory authorisation, systematically transferred complete protected articles and a substantial part of an investment-protected database into a commercial model-development process; retained or used those copies beyond what was technically and legally necessary; produced outputs reproducing protected expression; and entered markets in which the claimant already licensed or commercially exploited the same content.

The strongest defence would be equally precise:

The claimant has not proven that the relevant model was trained on the identified works, that legally relevant copies were made within the scope of Turkish protection, that the outputs contain protectable expression, or that the alleged use displaced an established market belonging to the claimant.

The likely judicial approach should therefore be modular rather than absolute.

Dataset acquisition, temporary copies, permanent archives, pre-training, fine-tuning, retrieval-augmented generation, live web browsing and user outputs should not be placed into a single legal category. The same article may be used lawfully at one stage and unlawfully at another.

That is also the central lesson of the United States proceedings. The dispute cannot be resolved by deciding whether artificial intelligence is “good” or “bad” for journalism. Nor can it be resolved by saying that everything available online is free for commercial reuse.

Public availability is not consent.

But copyright is not ownership of facts.

Between those two rules lies the evidence—and ultimately, the judgment.