X.AI LLC v Ellison and the Litigation Risk Facing AI Providers in Turkey
By Av. Asutay Duhan Meydan / Attorney at Law
Meydan AI & Tech Law
14 September 2026
The decision in X.AI LLC v. Ellison is easy to misstate. The Minnesota federal court did not hold that the State’s new AI nudification law is constitutional. It did not decide that an AI provider is finally liable for every unlawful image generated by a user. It did not even resolve whether Grok Imagine, its users or xAI itself have a First Amendment interest strong enough to invalidate the law.
The court decided a narrower question. While the constitutional action continues, should Minnesota be prevented from enforcing the law against xAI?
The answer at first instance was no.
That procedural answer matters well beyond Minnesota. It shows that a provider may have substantial constitutional objections and still lose the first urgent application because it moved late, described its losses in general terms or failed to establish irreparable harm with sufficient evidence. By the time the merits are reached, the provider may already have redesigned part of its product, restricted functions in one jurisdiction and absorbed compliance costs that cannot easily be recovered.
Turkey should study that sequence before Bill No. 2/3358 becomes law. The current text would impose six-hour removal duties, deepfake labelling requirements, content-verification and hallucination controls, administrative fines, access restrictions and possible suspension of activity. Yet Article 10 would bring all of this into force on the date of publication. There is no general adaptation period.
For AI providers, that is the point at which the American dispute becomes a Turkish legislative warning. A serious law needs serious safeguards. It must define the prohibited harm, identify the actor who controls that harm, give providers a workable period to adapt and preserve an effective route to judicial review. Otherwise, Turkey may reproduce the hardest part of X.AI v. Ellison: the provider will be required to comply first and litigate later, while the court asks why the company was not ready sooner.
1. What Minnesota Regulated
Minnesota enacted H.F. 1606, now codified as Minn. Stat. § 325E.91, on 7 May 2026. It entered into force on 1 August 2026.
The law prohibits a person controlling a website, application, software program or other service from allowing a user to access or use that service to “nudify” an image or video, or from nudifying an image or video on a user’s behalf. The relevant image must depict an identifiable person and add an intimate part that did not appear in the original image. It must also be realistic enough for a reasonable person to believe that the depicted body part belongs to that person.
The statutory definition of an “intimate part” is unusually broad. It includes the primary genital area, groin, inner thigh, buttocks and breast. The law is therefore directed at non-consensual sexual deepfakes, but its words can reach further than the ordinary meaning of nudity. It also contains an exception where use of the service requires technical skill. In practical terms, its principal target is an easily accessible consumer product that can produce realistic sexual images from ordinary prompts.
The Minnesota Attorney General may seek injunctive relief and civil penalties of up to USD 500,000 for each violation. The amount changes the provider’s risk calculation immediately. When liability can be counted image by image, a provider cannot wait for a settled interpretation before changing the product.
The legislative record was built around documented harms. Minnesota heard evidence concerning non-consensual sexual images of women, AI-generated child sexual abuse material, use of such images for blackmail and rapid circulation of sexualised images among students. The District Court treated that material as evidence of the public interest supporting the legislation.
xAI responded that Grok Imagine is used for a much wider field of expression, including art, family photographs, product mock-ups, journalism, religious imagery, memes and political satire. It also stated that its policies prohibit nudifying real persons, placing them in sexual settings without consent and sexualising children. According to the material submitted by the company, xAI had suspended more than 50,000 accounts and made more than 70,000 reports to the National Center for Missing and Exploited Children in 2026.
Minnesota’s answer was direct: post-generation moderation does not necessarily prevent the harm. Once a realistic sexual image has been created and distributed, closing an account or removing one copy may be too late. The State therefore chose to regulate access to the generation capability itself.
This is the central provider-liability question in the case. Is a provider’s duty satisfied by terms of service, reporting and enforcement after misuse, or may the State require the provider to prevent a defined category of output at the model or product layer?
2. Why the District Court Refused Preliminary Relief
xAI challenged the statute both on its face and as applied, relying principally on the First Amendment. It requested a temporary restraining order and then a preliminary injunction. The temporary request was rejected on 31 July 2026. On 4 September 2026, Judge Donovan W. Frank denied the preliminary-injunction motion in X.AI LLC v. Keith Ellison, Civil No. 26-3425 (DWF/DTS), Docket No. 54.
Under the American test applied by the court, preliminary relief is an extraordinary remedy. The applicant must address four matters: likelihood of success on the merits, likely irreparable harm without relief, the balance of equities and the public interest.
The court began with delay. The law had been signed on 7 May. xAI moved for emergency relief only three days before it became effective. Its explanation was that legislation takes time to reach corporate management and that management must authorise the resources required for litigation. The court was not persuaded. It described xAI as a sophisticated and well-resourced litigant capable of acting quickly. In the court’s view, the timing weakened the claim that the threatened harm was truly urgent and irreparable.
The court then considered the alleged harm.
First, the potential USD 500,000 penalty was monetary. The court treated monetary loss as ordinarily reparable, particularly because xAI had already disabled the relevant capability for Minnesota users and a penalty was therefore unlikely. The court also stated that penalties could be recovered if the law were later invalidated.
Second, xAI relied on engineering and compliance costs. Its product team had designed, tested and deployed geographically targeted controls within the Grok Imagine image-editing pipeline. Engineers had been diverted from other work. The court accepted that work had been performed but found the evidence of continuing cost vague. The adjustment had already been made, and the record did not quantify the future burden with sufficient precision.
Third, xAI argued that increased moderation would cause the loss of Minnesota users and related business. The moderation rate for image-editing prompts had reportedly increased by more than ten per cent after the controls were introduced. The court considered the resulting claim of user loss speculative. It was supported principally by one declaration and did not establish an actual or imminent commercial loss. The court also observed that the statute applied to competing companies offering comparable nudification tools, which weakened the allegation of a provider-specific competitive disadvantage.
Fourth, xAI argued that the loss of First Amendment freedom, even briefly, constitutes irreparable harm. The court accepted the general importance of that principle but refused to treat a pleaded constitutional violation as an automatic right to preliminary relief. The company still had to show a clear and likely irreparable injury. Its delay remained relevant to that assessment.
The court could have stopped there. It nevertheless considered the balance of harms and the public interest. It found that both weighed heavily in Minnesota’s favour. The State had enacted a nearly unanimous civil law to address the undisputed harm caused by photorealistic sexual images. It was entitled, for the purpose of the interim analysis, to argue that enforcement against users after circulation was insufficient and that prevention at the generation stage was necessary.
The court expressly left the constitutional merits open. Minnesota disputed xAI’s standing to assert the rights of users, disputed whether xAI had its own protected expressive interest and argued that the law could survive the applicable level of constitutional review. The court called those issues complex and stated that they would receive full consideration in the pending proceedings. The preliminary-injunction decision therefore preserves the law while the merits remain unresolved. It does not validate the law.
3. What xAI Says the District Court Got Wrong
xAI filed its notice of appeal on 4 September. It then sought an injunction pending appeal from the District Court; that application was refused on 9 September. On 11 September, the company applied to the United States Court of Appeals for the Eighth Circuit in Case No. 26-2806. The appellate filing uses the company’s new name, SpaceXAI LLC, although the District Court docket remains under X.AI LLC.
The appeal begins with a procedural criticism. xAI says the District Court did not analyse likelihood of success on the First Amendment claim. In its view, the court could not properly weigh the public interest without asking whether Minnesota was likely enforcing an unconstitutional restriction. A State has a strong interest in applying a valid law. It has no legitimate interest in applying an unconstitutional one. For xAI, the missing merits analysis affected the rest of the decision.
The company also challenges the finding of delay. It brought a pre-enforcement challenge before the law became effective and approximately two and a half months after enactment. The authorities cited by the District Court involved delays of approximately sixteen months or several years. xAI argues that time spent assessing new legislation before it begins to cause legal harm should not be treated in the same way as inactivity while an alleged violation is already continuing.
There is a broader institutional point behind that submission. A complex provider must identify the affected product, obtain a technical assessment, determine the legal entities involved, retain counsel, preserve evidence and secure authority for litigation. Requiring a company to file within days of signature may encourage incomplete proceedings. According to xAI, the diligence standard does not become stricter merely because a claimant has substantial resources.
On the merits, xAI characterises Grok Imagine as a tool for creating protected expression. Pictures, films, drawings and other visual works fall within the First Amendment. The use of AI does not remove that protection. The company also argues that designing and operating the model involves expressive and editorial choices of its own. It therefore asserts both its own rights and, under third-party-standing doctrine, the rights of users whose expression may be chilled.
The principal constitutional objection is that § 325E.91 is content based. Liability depends on what the finished image depicts. A hairstyle edit is permitted, while an edit showing an inner thigh or uncovered chest may expose the provider to an enormous penalty. xAI therefore says the law must satisfy strict scrutiny: Minnesota must demonstrate a compelling interest and narrow tailoring through the least restrictive means.
xAI accepts Minnesota’s legitimate interest in preventing non-consensual intimate deepfakes. Its argument is that the enacted text is not confined to that interest.
The law does not expressly require lack of consent. It can apply to a person generating an image of themselves or to an image generated with the depicted person’s consent. It does not require dissemination. An image created once, seen by no third party and immediately deleted may trigger the same provider exposure as an image distributed worldwide. It does not contain an express scienter requirement directed to the provider. It offers no clear safe harbour for a company that prohibits misuse, maintains filters, receives reports and acts against violating users. Nor does it state specific exceptions for artistic, medical, educational, scientific, journalistic, religious or political expression.
The definition of intimate parts gives the overbreadth argument a concrete form. An inner thigh, the chest of a male swimmer or a person in shorts may fall within the text even where the resulting image is neither pornographic nor abusive. xAI relies on these ordinary examples to argue that the statute burdens a substantial amount of lawful expression in relation to its legitimate reach.
The company also points to less restrictive legislation. Minnesota’s existing deepfake law and the federal TAKE IT DOWN Act use narrower definitions and include safeguards concerning consent, dissemination and protected categories of material. If those mechanisms can address non-consensual intimate imagery, xAI argues, the State cannot justify a wider provider-level prohibition.
Finally, xAI disputes the District Court’s treatment of irreparable harm. Minnesota users are already unable to use functions available elsewhere. The company continues to maintain a jurisdiction-specific version of the product and cannot necessarily recover those operational costs from the State. It describes the choice as immediate: comply with a law it considers unconstitutional or expose itself to potentially catastrophic liability. The Eighth Circuit has not yet finally resolved these arguments as of 14 September 2026.
4. Does Bill No. 2/3358 Contain a Comparable Rule
There is no exact equivalent of Minnesota’s nudification prohibition in Bill No. 2/3358. The Turkish proposal does not contain a provision drafted specifically around the generation of realistic intimate body parts from the image of an identifiable person.
That does not mean the problem is absent. Three provisions could create comparable, and in some respects broader, provider exposure.
Article 3 would add a new paragraph to Article 8 of Law No. 5651. It requires access-blocking and content-removal measures concerning AI-generated material that violates personality rights, threatens public security or has been falsely manipulated through deepfake technology to be carried out within six hours. Content providers and developers of AI systems would be jointly responsible for the obligation.
The wording raises an immediate control problem. A model developer may generate an output, but the user may download and publish it on an unrelated platform. The developer may have no technical power to remove the later publication. The proposal nevertheless places the developer within the same joint-responsibility sentence without distinguishing creation, hosting, dissemination and control.
The use of “or” also matters. Read literally, the provision may capture any falsely manipulated deepfake, even where it does not violate personality rights or threaten public security. That reading sits uneasily with Article 7, which assumes that deepfake material may remain available if it carries a clear, understandable and indelible “Generated by Artificial Intelligence” label. One provision appears to require removal within six hours, while another regulates lawful availability through labelling. The line between them is not clear.
Article 6 would amend Article 7 of the Cybersecurity Law No. 7545. It requires AI service providers to ensure transparency and auditability of training datasets, establish content-verification mechanisms to prevent false and manipulative information, implement algorithmic controls to reduce hallucination risk, develop human-approved oversight for high-risk applications and conduct periodic cybersecurity vulnerability testing. Failure may attract an administrative fine of up to TRY 5 million. Serious violations threatening public order may result in temporary suspension of activity.
This is the closest Turkish provision to an ex ante design obligation. Minnesota addresses one output category. Article 6 asks providers to prevent false and manipulative information and reduce hallucinations across potentially broad areas of model behaviour. Yet the proposal does not define the required performance threshold, the acceptable residual error rate, the relevant technical standard or the evidence that will establish compliance. No general-purpose model can guarantee that every output is true. A statutory duty framed as “prevent” rather than “take reasonable and proportionate measures to mitigate” may become an undefined result obligation.
Article 7 imposes the deepfake label. BTK may fine content providers and developers between TRY 500,000 and TRY 5 million. Systematic and intentional violation may result in access blocking of the content provider. Material that disrupts public order, harms personality rights or serves political manipulation is to be blocked urgently, with a possible criminal referral.
The Bill therefore contains provider-level duties capable of producing the same practical conflict seen in Minnesota: restrict the product immediately or operate under the threat of severe sanctions and service-level measures. The Turkish text also omits several safeguards at the centre of xAI’s appeal. It does not consistently define knowledge, intent, causation, consent, dissemination, actual control or a safe harbour based on reasonable preventive systems.
Article 2 adds a separate concern. It proposes an amendment to Article 125 of the Turkish Criminal Code under which a user directing an AI system towards criminal conduct is treated as the perpetrator, while the sentence for a developer whose design or training “enables” the offence is increased by one half. “Enables” is not defined. The provision does not say whether it requires intent, knowledge, conscious disregard of a specific risk or merely technical capability. This is difficult to reconcile with the personal nature of criminal responsibility under Article 20 of the Criminal Code and the legality principle under Article 38 of the Constitution. It also sits inside the offence of insult while referring more broadly to conduct criminalised “under this Code.”
The drafting defects are not theoretical. Article 8 of the Bill lists threat as “Article 28” of the Criminal Code, although the offence of threat is regulated by Article 106. A statute capable of criminal exposure, service suspension and urgent blocking cannot enter into force with uncertainty at this level.
The most serious timing problem appears in Article 10. The Bill would enter into force on publication. Minnesota gave providers almost three months between signature and effectiveness, and the court still criticised xAI for waiting. Bill No. 2/3358 presently gives no statutory transition period at all.
5. What Turkey Should Correct Before Enactment
The Bill should first contain a staged entry-into-force regime. Definitions, institutional mandates and the authority to prepare secondary rules may enter into force on publication. Provider obligations carrying fines, blocking or suspension should not. A general transition of at least six months, and a longer period for training-data governance or product-level technical redesign, would be more realistic. The implementing standards should be published before the compliance period begins.
This is not a request to delay protection for victims. Urgent measures concerning child sexual abuse material and clearly defined non-consensual intimate imagery can operate under existing criminal and internet-law mechanisms while the broader provider framework is prepared. The point is to separate a clearly prohibited harm from open-ended duties concerning every false, manipulative or hallucinatory output.
The Bill should then distinguish the relevant actors. A foundation-model developer, deployer, API integrator, content provider, hosting provider, social-network provider and end user do not exercise the same control. Liability should follow the technical and legal ability to prevent, detect, remove or distribute the disputed material. Joint responsibility should not be imposed on a developer for content hosted elsewhere unless the developer had legally relevant control, notice or participation.
The prohibited deepfake category should be defined with greater precision. For intimate imagery, the law should address identifiability, lack of consent, realistic depiction, knowledge or a defined fault standard, and dissemination or a concrete risk of dissemination. It should contain express treatment of self-created and consensual material, as well as medical, scientific, educational, journalistic, artistic and public-interest uses. Child sexual abuse material requires a distinct and stricter regime.
Provider duties should be framed around reasonable, proportionate and state-of-the-art measures. An obligation to eliminate hallucinations or prevent all false information is not technically measurable. A workable rule would require documented risk assessment, testing, incident response and mitigation appropriate to the model’s purpose, capability and risk level. Compliance should be assessed against published standards rather than hindsight after one harmful output.
A safe-harbour structure is also necessary. It should not immunise a provider that knowingly designs or promotes an unlawful function. It should, however, recognise a provider that maintains effective policies, age and identity controls where justified, safety filters, reporting channels, traceability, prompt and output logging subject to data-protection limits, rapid response and independent testing. Without such a structure, the law may punish the provider that can demonstrate governance in the same manner as the provider that did nothing.
Urgent orders require procedural safeguards. The competent authority, evidential threshold, scope and duration of an emergency measure should be stated in the statute. A reasoned notice should identify the content, feature or system concerned. Except in narrowly defined emergencies, the provider should have an opportunity to respond and cure. Emergency blocking or suspension should receive prompt judicial review. Measures should target the relevant content or function where technically possible before the entire service is made unavailable.
Sanctions also require calibration. The law should explain whether fines attach per output, per notice, per continuing breach or per compliance failure. It should address turnover, gravity, duration, fault, mitigation, cooperation, recurrence and the number of affected persons. Temporary suspension should be reserved for serious and continuing violations where a narrower measure is insufficient.
Finally, territorial scope must be clear. A foreign provider serving users in Turkey may have no local subsidiary, while a Turkish integrator may depend on a foreign foundation model. The Bill should identify when offering a service in Turkey creates jurisdiction, whether a local representative is required and which entity must answer a regulatory notice. Otherwise domestic companies will be easier to sanction while the legal route against foreign providers remains uncertain.
6. What AI Providers Should Do Before the Bill Becomes Law
Providers should not wait for publication. Under the current Article 10, publication and legal effect would occur on the same day.
The first step is legislative engagement supported by technical evidence. Providers, sector associations, universities and civil-society organisations should submit written opinions to the principal and secondary parliamentary committees while the text is still amendable. The purpose should be specific: propose a transition period, define the regulated actors, separate non-consensual intimate imagery from lawful synthetic content, specify fault and control standards, reconcile Articles 3 and 7, correct the criminal-law provisions and establish rapid judicial review.
The second step is entity and product mapping. Each group should identify which entity develops the model, operates the user-facing service, contracts with Turkish users, determines data processing, hosts outputs, responds to notices and controls safety systems. A Turkish subsidiary should not carry an obligation that only the foreign parent can technically perform. Contracts with model vendors, cloud providers, distributors and enterprise deployers should provide access to the information and cooperation required for Turkish compliance.
The third step is a Turkey-specific control plan. Providers should identify the model functions most exposed to the Bill: image editing, voice cloning, realistic video, political content, search summaries, high-risk decision support and open-ended generation. They should determine whether a feature can be restricted by modality, endpoint, user class or geography without disabling the entire service. The ability to isolate a function may become the difference between targeted compliance and market withdrawal.
The fourth step is a six-hour response system. A notice received in Turkey must reach a team with legal authority and technical access immediately. The provider should know who verifies the notice, who preserves evidence, who decides whether the material is unlawful, who can restrict an output or feature, and who communicates with BTK or another authority. A general global queue operating on a twenty-four or forty-eight-hour cycle will not satisfy a six-hour statutory obligation.
The fifth step is evidence. X.AI v. Ellison shows that a statement such as “our engineers spent substantial time” may be insufficient. Providers should record engineering hours, diverted projects, infrastructure costs, moderation changes, false-positive rates, user impact, revenue effects and the continuing cost of jurisdiction-specific controls. If urgent judicial relief becomes necessary, the court will need sworn, quantified and contemporaneous evidence showing why the harm cannot be repaired later.
The sixth step is governance evidence. Training-data lineage, auditability, red-team results, model evaluations, hallucination-mitigation measures, human oversight and cybersecurity testing should be documented before enforcement begins. A control that cannot be demonstrated may have little value during an investigation.
The final step is litigation readiness. Providers should prepare draft grounds addressing legal certainty, proportionality, freedom of expression, property, freedom of enterprise, competence, procedure and causation. They should identify the competent court for each possible act and preserve the ability to request a constitutional referral. This does not mean filing a speculative case. It means avoiding the position in which the first days of an enforcement crisis are spent identifying the legal entity, locating the relevant logs and deciding which court has jurisdiction.
7. If the Same Dispute Were Litigated in Turkey
The procedural structure would differ fundamentally from the American case.
A private AI provider cannot bring a direct annulment action before the Constitutional Court merely because Parliament has enacted a statute. Abstract review of a statute is reserved by Articles 150 and 151 of the Constitution to the President, the two parliamentary groups with the largest membership and at least one fifth of the total number of members of Parliament. The action period for substantive review is sixty days from publication.
For providers, this makes the legislative stage unusually important. If the authorised political actors do not seek abstract review, the company ordinarily needs a concrete administrative or judicial application of the law before it can build its own case.
Assume first that BTK imposes an administrative fine, orders the blocking of AI content or directs a provider to disable a function. Alternatively, assume that the competent cybersecurity authority orders temporary suspension under the proposed Article 6. The provider’s principal merits claim would be an action for annulment of the administrative act under Article 2 of the Administrative Judicial Procedure Act No. 2577. The act could be challenged for lack of competence, defective form or procedure, absence or insufficiency of legal cause, unlawful subject matter, improper purpose, disproportionality and violation of constitutional rights.
The urgent request would be a stay of execution under Article 27 of Law No. 2577, rather than an American preliminary injunction. The provider would have to establish two cumulative conditions: implementation of the act would cause damage that is difficult or impossible to repair, and the act is manifestly unlawful. The court must give reasons addressing both conditions.
The requested interim relief should be drafted around the particular measure. For a fine, the provider would request suspension of collection and enforcement. For an access restriction, it would request suspension of the blocking order. For temporary suspension, it would request preservation of the service until the merits are resolved. The application should explain why a narrower measure, such as disabling one modality or removing identified content, would protect the public interest without eliminating the service.
Delay is not written into Article 27 as an independent statutory factor. It may still affect credibility. A company claiming that suspension will destroy its market position will face an obvious question if it waited after notification, failed to activate an available objection mechanism or produced no evidence of daily loss. The practical lesson of Ellison therefore survives the change of legal system: urgency must be shown by conduct as well as words.
The evidence should be considerably more precise than a general assertion of compliance cost. The provider should document the number of affected users, contractual consequences, technical work required, continuing infrastructure cost, irreversible loss of trust, risk to enterprise customers, data migration, employee impact and whether later compensation would realistically restore the position. It should also show its existing safeguards and explain how the challenged act exceeds what is necessary to address the identified harm.
If the challenged measure is an individual BTK fine or order, the competent court would ordinarily be the administrative court determined under the statutory and territorial rules, commonly the Ankara Administrative Court for a centrally issued individual act unless a special jurisdictional provision applies. If BTK or another national authority issues a country-wide regulatory rule, guideline or communique, the Council of State may have first-instance jurisdiction under Article 24 of the Council of State Act No. 2575. The final route depends on the legal character of the challenged instrument, not simply the name of the authority.
Access blocking requires further care. If the order is an administrative act, administrative judicial review and a stay request are the proper route. If it is issued by a criminal judgeship under the judicial mechanism of Law No. 5651, the remedy is an objection under the criminal-procedure rules. Bill No. 2/3358 does not always identify with sufficient clarity who will make each urgent blocking or suspension decision. That uncertainty should be repaired before enactment because it affects both competence and the available remedy.
In the annulment action, the provider could ask the administrative court to refer the applicable statutory provision to the Constitutional Court under Article 152 of the Constitution. The provider would argue, depending on the measure, that the provision violates legal certainty and the rule of law under Article 2; proportionality under Article 13; freedom of expression under Article 26; property under Article 35; freedom of enterprise under Article 48; the right to an effective and fair judicial process under Article 36; and, for criminal exposure, the legality and personal-responsibility guarantees under Article 38.
The administrative court cannot itself annul the statute. If it considers the constitutional objection serious, it stays the proceedings and refers the question to the Constitutional Court. If it refuses, that refusal may be examined together with the final judgment in the ordinary appellate process. The Constitutional Court’s ruling on the provision will then determine the normative basis of the administrative dispute.
The provider’s main prayer for relief would therefore be structured as follows: stay execution without security where appropriate; annul the individual or regulatory administrative act; refer the governing statutory provision to the Constitutional Court; and, if loss has already occurred, award compensation through a full-remedy claim brought together with or following the annulment action under the procedural rules.
The first-instance judgment would generally proceed to the Regional Administrative Court through appeal. A further appeal to the Council of State would depend on whether the dispute falls within the categories open to cassation under Law No. 2577. Where the Council of State acts as the first-instance court for a country-wide regulatory measure, the appellate structure is adjusted accordingly.
After exhaustion of ordinary remedies, a private legal person may apply individually to the Constitutional Court only in relation to rights compatible with legal personality. The application is not a direct abstract challenge to the statute. It must concern a current and personal violation caused by public power after available remedies have been exhausted, and it must ordinarily be filed within thirty days of exhaustion. If the dispute remains within the scope of the European Convention on Human Rights, an application to the European Court of Human Rights may follow exhaustion of domestic remedies within the applicable four-month period.
Two related proceedings should be kept separate. A victim seeking prevention, removal or damages for an intimate deepfake may bring private-law claims before the civil courts, relying on personality rights and tort principles; interim protection may be sought under Article 389 of the Code of Civil Procedure. If a natural-person developer is prosecuted under the proposed criminal provision, the criminal court would decide individual criminal responsibility and could also refer the applicable rule to the Constitutional Court. Turkish law does not impose an ordinary criminal sentence on a legal person; Article 20 of the Criminal Code preserves only specifically authorised security measures. Neither route replaces the provider’s administrative challenge to a fine, blocking order or suspension.
Conclusion
X.AI v. Ellison has not yet answered whether Minnesota’s law is constitutional. Its immediate lesson is procedural. A provider can present a serious challenge and still remain subject to the law because it did not move early enough or prove the urgency with sufficient detail.
Bill No. 2/3358 would place Turkish providers in a more difficult position. It contains no direct Minnesota-style nudification provision, but Articles 3, 6 and 7 impose removal, prevention, labelling, blocking and suspension risks that can reach a wider field of AI activity. Article 10 would make those duties effective on publication, without a general compliance period. The text also leaves material questions of control, fault, consent, dissemination, safe harbour, territorial scope and judicial procedure unresolved.
Turkey still has time to correct this. The Bill remains before parliamentary committees. A transition period, precise actor definitions, control-based responsibility, measurable technical duties, narrow deepfake rules, proportional sanctions and prompt judicial review would protect victims while giving providers a rule they can actually obey.
Providers should use that time as well. They should engage with the legislative process, map their Turkish operations, build a six-hour response capacity, document technical controls and quantify the cost of jurisdiction-specific restrictions. If litigation becomes necessary, the first filing must already contain the evidence that Ellison found missing.
The question is no longer whether an AI provider will be asked to govern harmful outputs. The legal question is who controlled the relevant risk, what preventive step was reasonably available, and whether the State used a proportionate measure against the correct actor.
That answer should be written into the law before it has to be reconstructed in court.
Av. Asutay Duhan Meydan
Meydan AI & Tech Law
Sources and Legal Authorities
1. X.AI LLC v. Keith Ellison, U.S. District Court for the District of Minnesota, Civil No. 26-3425 (DWF/DTS), Memorandum Opinion and Order, Dkt. 54, 4 September 2026: https://www.courthousenews.com/wp-content/uploads/2026/09/Judge-denies-xAIs-preliminary-injunction-against-Minnesotas-ban.pdf
2. SpaceXAI LLC v. Keith Ellison, U.S. Court of Appeals for the Eighth Circuit, No. 26-2806, Motion for Injunction Pending Appeal, 11 September 2026: https://fingfx.thomsonreuters.com/gfx/legaldocs/zjvqlknbgpx/XAI%20APPEAL.pdf
3. Minnesota Statutes § 325E.91 and H.F. 1606, as reproduced and analysed in the District Court’s 4 September 2026 order.
4. Grand National Assembly of Turkey, Bill No. 2/3358, submitted 7 November 2025, status and official text: https://www.tbmm.gov.tr/Yasama/KanunTeklifi/12d348f9-77ee-4f09-8b78-019a5e27521f
5. Constitution of the Republic of Turkey, Articles 2, 13, 26, 35, 36, 38, 48, 125 and 150-153: https://www.anayasa.gov.tr/tr/mevzuat/anayasa/
6. Administrative Judicial Procedure Act No. 2577, Articles 2, 7, 12, 27, 45 and 46.
7. Council of State Act No. 2575, Article 24.
8. Law No. 6216 on the Establishment and Rules of Procedure of the Constitutional Court, Articles 40 and 45-47.
9. Turkish Criminal Code No. 5237, Articles 20, 106 and 125.
10. Law No. 5651 on the Regulation of Publications on the Internet and Combating Crimes Committed by Means of Such Publications.
11. Cybersecurity Law No. 7545 and Electronic Communications Law No. 5809.
12. European Convention on Human Rights, Article 10.