AI Provider Liability in Turkey: Key Differences Between Turkish and Foreign AI Providers Under Bill No. 2/3358
By Av. Asutay Duhan Meydan / Attorney At Law
Meydan AI & Tech Law
Introduction
Turkey still does not have a horizontal artificial intelligence statute comparable to the EU AI Act. AI providers are instead subject to a fragmented legal framework built around data protection, internet regulation, cybersecurity, civil liability, consumer protection, criminal law and intellectual property.
Bill No. 2/3358 may materially change that structure. The proposal, submitted to the Grand National Assembly of Turkey on 7 November 2025, remains before parliamentary committees as of 9 September 2026. It is therefore not current law. If enacted, however, it would introduce AI-specific provisions into several existing statutes rather than creating a standalone AI Act.
In our previous analysis of the Bill, we focused primarily on its operational consequences for AI providers: six-hour removal obligations, deepfake labelling, emergency access restrictions, training-data auditability, hallucination controls and, most importantly, the possibility of temporary suspension of an AI service. The resulting enforcement structure may be described as an escalation from content removal, to access blocking, and ultimately to suspension of activity.
A second question is equally important:
Who bears these risks when the AI provider is established in Turkey, and what changes when the provider is a foreign company with no Turkish legal entity?
The answer is more complex than corporate residence.
A Turkish company registered with a Turkish tax office will ordinarily be easier for regulators, courts and users to reach. A foreign company may face a more complicated enforcement path. But tax registration is not the general legal test for AI liability, and foreign incorporation is not immunity from Turkish law.
The relevant question is not simply where the company was incorporated. It is what role the company performs, what it controls, whom it serves, where the relevant legal effect occurs and which Turkish statutory regime is engaged.
1. The Real Dividing Line Is Not the Tax Office
For the purposes of this analysis, a Turkey-based AI provider means an AI company operating through a Turkish legal and commercial presence: for example, a Turkish company or locally established operation that contracts with customers, processes data or provides AI services from Turkey and is registered within the Turkish tax system.
A foreign AI provider, by contrast, means a company incorporated abroad that makes an AI model, application, API or other AI service available to users or businesses in Turkey without necessarily maintaining a Turkish subsidiary, branch or ordinary operational establishment.
This distinction is commercially important, but legally incomplete.
Turkish legislation does not apply one universal jurisdictional test to every AI-related dispute. Different connecting factors apply under different laws.
A provider may become relevant under the Personal Data Protection Law because it determines the purposes or means of processing personal data; under civil law because its conduct causes damage; under consumer law because it markets a subscription to a Turkish consumer; under Law No. 5651 because it falls within a particular internet-service classification; or under criminal law because an offence or its result occurs in Turkey.
Therefore:
Tax presence affects the practical route of enforcement. It does not, by itself, create or eliminate substantive AI liability.
This distinction becomes particularly important where a foreign foundation-model company operates through a Turkish subsidiary or where a Turkish AI company integrates a foreign model. The correct analysis cannot stop at the corporate chart. It must determine which entity controls the model, the user relationship, the data processing, the safety architecture, the disputed output and the regulatory response.
2. Turkey-Based AI Providers: Direct Regulatory Exposure
For a locally established AI provider, the principal difference is not necessarily that more laws apply. It is that Turkish law can generally be applied more directly.
The company may have a Turkish address for service, local management, assets, contractual relationships, employees and an established regulatory presence. Administrative investigations, civil claims and compliance requests therefore have an immediate domestic addressee.
Civil and Contractual Liability
Even without Bill No. 2/3358, AI-generated harm may already fall within ordinary Turkish civil-law principles.
Article 49 of the Turkish Code of Obligations No. 6098 requires a person who culpably and unlawfully causes damage to compensate that damage. Where the relationship is contractual, Article 112 provides a separate basis where an obligation is not performed or is improperly performed, unless the debtor proves that no fault is attributable to it.
This does not mean that an AI provider automatically becomes liable whenever its system produces an incorrect, harmful or unlawful answer.
A claimant would still need to establish the relevant legal basis, including matters such as breach, unlawfulness, fault where required, damage and causation. The fact that AI generated the immediate output should not replace ordinary liability analysis.
The same principle applies to personality rights. Articles 24 and 25 of the Turkish Civil Code permit prevention, cessation and determination claims, together with potential compensation and other remedies, where personality rights are unlawfully infringed. This can become particularly relevant to AI-generated voice, image, likeness, impersonation and deepfake disputes.
For consumer-facing AI products, paid chatbot subscriptions, AI editing services and similar B2C products may also fall within the distance-contract regime of Consumer Protection Law No. 6502, including pre-contractual information obligations.
Data Protection and Model Inputs
For a Turkish legal person, the legal entity itself may be the data controller where it determines the purposes and means of processing. Appointment of a contact person does not transfer that responsibility away from the company.
This is particularly important for domestic AI companies that use user prompts, conversation histories, uploaded documents, biometric information, customer datasets, model-training or fine-tuning datasets, or employee and enterprise-user information.
Where a local AI company sends personal data to a foreign foundation-model provider or foreign cloud infrastructure, the cross-border transfer regime under Article 9 of Law No. 6698 must also be examined. The current framework includes mechanisms such as standard contractual clauses and binding corporate rules, subject to the applicable statutory requirements.
Bill No. 2/3358 would add another layer. Its Article 4 proposes to introduce requirements of anonymity, non-discrimination and legitimacy for datasets used in AI applications and would characterise discriminatory datasets as a data-security violation.
This should not be confused with copyright clearance. Compliance with personal-data rules does not establish that copyrighted training material was lawfully acquired or used. Copyright questions remain separately governed, principally by Law No. 5846.
Cybersecurity and Model Governance
Turkey’s Cybersecurity Law No. 7545 already applies broadly to natural and legal persons that operate, exist or provide services in cyberspace and imposes duties concerning cybersecurity measures, information requests, vulnerabilities and cyber incidents.
Bill Article 6 would go much further for AI service providers.
It proposes statutory duties concerning transparency and auditability of training datasets, mechanisms against false and manipulative information, algorithmic controls reducing hallucination risk, human approval mechanisms in high-risk applications and periodic cybersecurity vulnerability testing.
Non-compliance could result in an administrative fine of up to TRY 5 million, while serious violations threatening public order could lead to temporary suspension of activity.
For a domestic provider, this is the most straightforward regulatory exposure in the proposal. There is a local entity whose service can be inspected, ordered to comply, fined and potentially suspended.
3. Foreign AI Providers: No Turkish Company Does Not Mean No Turkish Liability
The more difficult issue concerns a foreign provider with no subsidiary or ordinary establishment in Turkey.
There is currently no general provision equivalent to the EU AI Act’s territorial-scope mechanism stating comprehensively when every foreign AI provider becomes subject to Turkish AI legislation. Bill No. 2/3358 does not adequately solve this problem either.
Yet it would be incorrect to conclude that foreign AI providers therefore remain outside Turkish law.
Personal Data: The Effects of Processing Matter
Law No. 6698 does not contain a GDPR-style territorial-scope article. The Personal Data Protection Board has nevertheless used an effects-based analysis in relation to foreign controllers.
In its 2025/601 decision, the Board expressly discussed its earlier approach under Decision No. 2019/10: where a breach at a foreign controller affects persons residing in Turkey and those persons benefit from the controller’s products or services in Turkey, Turkish data-protection obligations may become relevant. In the specific 2025 case, the foreign company fell outside the relevant processing scope precisely because the affected individuals were not using its services in Turkey.
Foreign data controllers that fall within the registration regime are also required to act through a data-controller representative in Turkey. Importantly, appointment of that representative does not transfer the underlying responsibility away from the foreign controller.
For global AI providers, this means that the absence of servers, employees or a subsidiary in Turkey is not necessarily decisive where Turkish users are actively using the service and their personal data are being processed.
Civil Claims Can Cross the Corporate Border
The same is true in private law.
Under Article 34 of the Turkish International Private and Procedural Law No. 5718, where the place of the wrongful act and the place where damage occurs are in different countries, the law of the country where the damage occurs may govern the tort. Article 35 contains specific connecting rules for personality-right violations committed through the internet or other mass communication tools and expressly extends those principles to certain personal-data-related personality claims.
A foreign AI provider can therefore face Turkish-law questions where a harmful output generated by a foreign system produces legally relevant damage in Turkey.
Consumer contracts create another route. Article 26 of Law No. 5718 protects the mandatory minimum rights available under the consumer’s habitual-residence law in qualifying international consumer contracts, while Article 45 provides Turkish jurisdiction in specified consumer disputes. A foreign governing-law or forum clause therefore cannot automatically be assumed to exclude every Turkish consumer-law consequence.
Criminal Exposure Requires a Different Analysis
Criminal responsibility must be treated more carefully.
Article 8 of the Turkish Criminal Code provides for application of Turkish criminal law to offences committed in Turkey, and an offence may have a Turkish territorial connection where the relevant result occurs in Turkey.
However, Article 20 of the Criminal Code also establishes two fundamental principles: criminal responsibility is personal, and legal persons are not subject to criminal penalties, without prejudice to statutory security measures.
This distinction becomes essential under Article 2 of Bill No. 2/3358.
The proposed rule provides that a developer who enables criminal conduct through the design or training of an AI system may face an increased sentence.
The provision should therefore not be summarised loosely as “the AI company will be criminally liable.” Under Turkish criminal-law principles, the analysis must identify the relevant natural person, conduct and required mental element. Questions of design authority, knowledge, intent, participation and causation remain unavoidable.
For a foreign provider, additional territorial and international criminal-jurisdiction questions would also need to be resolved before liability could be imposed.
4. Bill No. 2/3358 May Create Similar Duties but Different Enforcement Risks
The central weakness of the Bill, when viewed through the distinction between domestic and foreign providers, is that it imposes substantial provider obligations without establishing a comprehensive territorial and representation framework for AI providers.
This creates a significant enforcement asymmetry.
For a Turkey-based AI provider, data-protection obligations, civil claims, consumer-law duties and the proposed AI-specific obligations can generally be directed against an identifiable domestic entity. The provider has a local legal personality, assets, management and an address against which regulatory and judicial mechanisms can operate. If Bill Article 6 enters into force in its current form, administrative fines and potentially temporary suspension of activity could therefore be imposed through a relatively direct enforcement route.
The position of a foreign provider without a Turkish legal entity is different, but not necessarily safer.
Under data-protection law, a foreign provider may still become subject to Turkish obligations depending on the processing activity and its effects on individuals in Turkey. In civil disputes, Turkish law or Turkish jurisdiction may become relevant where damage, personality-right infringements or consumer effects occur in Turkey. A foreign provider offering paid services to Turkish consumers cannot simply assume that a foreign governing-law clause will eliminate the application of mandatory Turkish consumer protections.
The practical difference becomes more visible in enforcement.
A monetary claim or administrative sanction against a provider with no assets or legal entity in Turkey may require additional cross-border steps. By contrast, Turkey retains domestic regulatory mechanisms capable of affecting access to an online service itself.
For that reason, the absence of a Turkish establishment may in some circumstances increase the practical importance of access blocking and service restriction as regulatory tools.
Bill Article 3 illustrates this problem. It would require the removal or blocking of certain AI-generated content within six hours and places AI developers within a joint responsibility structure even where the developer may not control the third-party environment in which the output was later published.
Article 7 creates a similar issue for deepfake content. It would require AI-generated deepfake material to carry a clear and non-removable AI label, expose developers and content providers to administrative sanctions, and permit blocking where violations become systematic and intentional.
These obligations may be relatively straightforward to address where the developer is a domestic company. Their application to a foreign model provider, however, raises more difficult questions. The Bill does not clearly establish when offering an AI service to users in Turkey alone is sufficient to subject a foreign provider to all provider-level obligations, nor does it establish a general local representative mechanism specifically for foreign AI providers.
Existing Law No. 5651 demonstrates why this matters. Additional Article 4 already establishes representation obligations for qualifying foreign social network providers that exceed statutory access thresholds in Turkey. Yet this is a sector-specific regime for social network providers. It cannot automatically be extended to every foreign chatbot, foundation-model company, API provider or generative AI service.
Accordingly, a foreign AI provider may simultaneously be outside one statutory classification and within another.
A foreign company may have no Turkish subsidiary, yet still be a foreign data controller for KVKK purposes. It may not qualify as a social network provider under Law No. 5651, while still providing AI services directly to Turkish users. Its output may also create a civil-law effect in Turkey even though the model itself was trained, hosted and operated abroad.
This exposes one of the most important structural problems in Bill No. 2/3358.
“AI provider” cannot itself be the end of the liability analysis.
The legal framework must distinguish between the developer of the underlying model, the company deploying it, the entity contracting with the end user, an API intermediary, a content provider and a platform distributing the resulting output.
Without such distinctions, identical statutory language may apply to actors that have radically different levels of technical and legal control over the relevant risk.
5. Precautions AI Providers Should Take Before the Law Changes
Bill No. 2/3358 remains a proposal, but the appropriate response is not to wait for the day of enactment. Article 10 currently provides for entry into force on publication and contains no general adaptation period.
Providers serving the Turkish market should therefore prepare before the legislative framework becomes binding.
The first step should be to determine the Turkish regulatory perimeter entity by entity. A multinational group should identify which entity is the developer, AI service provider, contractual supplier, content provider, data controller, data processor or, where applicable, social network provider. The foreign parent and Turkish subsidiary should not automatically be treated as a single legal actor.
Responsibility should then be allocated contractually according to actual technical control. A Turkish AI company integrating a foreign foundation model should secure access to information necessary for Turkish compliance, including model documentation, data-processing details, incident information, safety updates, audit material and cooperation with regulatory requests. Contractual allocation cannot eliminate statutory liability, but it can prevent the local provider from being placed in a position where it bears legal obligations without having access to the information required to satisfy them.
AI providers should also establish a Turkey-specific regulatory response mechanism. A global trust-and-safety process designed around ordinary internal review periods may be insufficient if the Bill introduces a six-hour statutory deadline. The responsible team should know in advance who receives a Turkish regulatory notice, who performs the legal assessment and who has technical authority to restrict the relevant content, feature or service.
Training-data governance should also be made demonstrable. Providers should maintain dataset lineage, lawful-processing records, anonymisation procedures, discrimination assessments and documentation concerning cross-border personal-data transfers. If auditability becomes an express statutory obligation, compliance that cannot be evidenced may have little practical value in an investigation.
The same principle applies to model safety.
Hallucination mitigation, red-team testing, evaluation results, human oversight, cybersecurity testing and responses to identified failure modes should be documented systematically. If the legal framework begins to impose duties concerning model reliability and safety, the central issue may not be whether an AI system ever produced an incorrect answer. The more relevant question may become whether the provider can prove that reasonable technical and organisational controls were designed, tested and maintained.
Providers of generative image, audio and video systems should separately examine synthetic-content provenance. Watermarking, persistent labels, metadata and other traceability mechanisms may become necessary if the deepfake provisions of the Bill enter into force.
Election security, political manipulation and public-order risks should likewise be separated from ordinary content moderation. Bill No. 2/3358 repeatedly treats these matters as elevated regulatory concerns. Providers serving the Turkish market should therefore consider specific escalation procedures for election periods and other politically sensitive circumstances.
Foreign providers should additionally establish a functioning Turkish legal-response capability even where no general AI representative is currently mandatory. This includes reviewing whether KVKK representation and VERBIS obligations apply, determining whether the provider falls within Law No. 5651, maintaining access to Turkish legal counsel and ensuring that notices issued in Turkey can reach a team with both legal authority and technical capacity to respond.
Finally, providers should treat service continuity as part of legal compliance.
Where technically possible, a company should be capable of restricting a particular model function, modality, endpoint or category of output in Turkey without disabling its entire service. If Turkish AI regulation moves from individual content intervention toward access restriction and temporary suspension, the technical architecture of the product itself becomes part of regulatory risk management.
Conclusion
The distinction between a Turkey-based AI company and a foreign AI provider is real, but it should not be misunderstood.
For the domestic provider, Turkish incorporation, local tax registration, assets and management create a direct enforcement path. Regulatory instructions, administrative sanctions, civil proceedings and potentially service-level measures can operate against an immediately identifiable local entity.
For the foreign provider, the path may be less direct, but the legal exposure does not disappear. Turkish data-protection law may become relevant through effects on users in Turkey; civil and personality claims may be connected to the place where damage occurs; Turkish consumer rules can survive cross-border contractual structures; criminal law may apply where the statutory territorial requirements are met; and internet regulation can ultimately affect the availability of a foreign digital service in the Turkish market.
Bill No. 2/3358 would make this distinction more important because it moves Turkish regulation beyond the legality of individual outputs and toward model design, training data, hallucination controls, human oversight and service continuity.
Yet the proposal does not presently contain a sufficiently precise territorial architecture explaining when a foreign AI provider becomes subject to each new provider-level obligation or how those obligations are to be enforced against an entity without a Turkish establishment.
That point should be addressed during the legislative process.
For providers, however, the practical conclusion is already clear:
A Turkish tax registration may determine how easily an authority can reach the company. It does not determine where legal responsibility begins or ends.
In the Turkish market, the decisive questions will increasingly be:
Who controls the relevant risk? Who provides the service to the Turkish market? Where does the legal effect occur? And can the provider demonstrate that the risk was governed before it became harm?
Av. Asutay Duhan Meydan
Meydan AI & Tech Law