By Asutay Duhan Meydan, Attorney at Law
Bakırköy 2nd Commercial Court of First Instance, Case No. 2023/940, Decision No. 2025/971, 17 October 2025
A 2025 judgment of the Bakırköy 2nd Commercial Court of First Instance offers an important indication of how Turkish courts may approach web scraping in Turkey, particularly where publicly accessible platform data are systematically extracted for commercial use. A 2025 judgment of the Bakırköy 2nd Commercial Court of First Instance offers a useful indication of how Turkish courts may approach systematic web scraping of publicly accessible platform data. The court found unfair competition where a technology company collected structured real-estate data through automated means, continued doing so after formal notices and IP blocking, and incorporated the data into its own commercial analysis service.
The decision is not an AI training judgment. Its relevance lies in an earlier question that AI providers increasingly need to answer: does the fact that data are publicly accessible make their automated commercial extraction legally unrestricted under Turkish law?
A company collects data from a publicly accessible website.
The data are largely entered by users rather than created by the platform. No password is bypassed. No closed database is technically penetrated. The collecting company does not merely republish the material; it adds its own analysis and uses the data as an input for a different technology service.
At first sight, this looks like a relatively strong defence to a scraping claim.
The Bakırköy 2nd Commercial Court of First Instance nevertheless found unfair competition.
That is what makes its 17 October 2025 judgment worth reading.
The decision is not an artificial-intelligence training judgment, and it should not be presented as one. The court did not decide whether copyrighted works may be used to train a large language model, whether machine learning constitutes a permitted form of text and data mining, or whether web scraping is generally unlawful under Turkish law.
But the dispute sits unusually close to those questions.
The defendant itself described its service as a high-technology and artificial-intelligence-based system. It collected information from several online sources, processed that information and provided users with property risk analyses. The court’s concern was not that the defendant had copied the plaintiff’s business model. The concern was how the defendant had obtained and commercially used the underlying data.
For AI providers, that distinction matters.
The legal problem may arise before a model is trained and before a model produces a single output.
It may arise at the point of acquisition.
Key Takeaways
- Publicly accessible data are not necessarily free for systematic commercial extraction under Turkish law.
- The court distinguished web scraping from unlawful technical intrusion into a closed system.
- Continued data collection after formal notices and IP blocking was a material part of the court’s assessment.
- The defendant’s use of the collected data in a different, technology-driven analysis service did not prevent a finding of unfair competition.
- The judgment does not decide whether AI model training itself is lawful; its significance lies in the data-acquisition stage that may precede model training.
The dispute
The plaintiff operated an online real-estate listing platform.
The defendant provided a different service. According to its defence, a user could enter details concerning a property and receive a detailed risk-analysis report produced through technologies developed by the defendant. The defendant stated that its system monitored numerous sources, including property-listing websites, auction and enforcement portals, maps, public decisions, newspapers, earthquake-risk information and traffic data.
Its position was therefore clear: it was not running another classified-advertising platform.
It was building an analytical product on top of information gathered from a number of sources.
That distinction did not persuade the court to dismiss the claim.
Court-appointed experts identified the use of information taken from the plaintiff’s platform, including listing numbers and dates, property type, gross and net area, category, number of rooms, building age, floor information, occupancy status, fees, credit eligibility, exchange information and map-location data.
The defendant then added technical analysis and used the resulting information in its own service.
This factual structure is important.
The case was not simply:
copy → republish.
It was closer to:
collect → structure → analyse → create a downstream service.
That is a structure AI lawyers will recognise.
The technical finding: scraping, but not hacking
The technical distinction drawn by the experts is one of the strongest parts of the case.
They found that the defendant’s software used web scraping to obtain data from the plaintiff’s website.
They also found that the relevant information was externally accessible. For that reason, the experts stated that the activity could not technically be characterised as penetration of the plaintiff’s systems or an unlawful technical intrusion merely because the information had been collected automatically.
That finding deserves emphasis.
There was no determination that the defendant had broken into a closed database.
There was no finding that authentication had been defeated.
The information was available from the outside.
Yet that did not end the legal inquiry.
The distinction is fundamental:
Technical accessibility and legal entitlement are not the same question.
A crawler does not acquire a legal licence merely because a server responds to its request.
For AI companies, that proposition should be obvious in compliance terms, but it is often lost in discussions that reduce scraping to a binary technical question: either the data were behind a security barrier or they were “public”.
The Bakırköy judgment suggests that Turkish unfair-competition law may require a more substantive analysis.
The facts became considerably worse after notice
The decision should not be read without the chronology.
The plaintiff sent two notarial notices objecting to the data collection.
IP addresses associated with the activity were then blocked.
According to the expert findings accepted by the court, the defendant continued collecting data by using different IP addresses.
This matters considerably.
The case is therefore not authority for the proposition that one automated request to a public website constitutes unfair competition.
It involved continued, commercial data extraction after the source had expressly objected and implemented technical blocking measures.
That sequence changes the character of the dispute.
For counsel advising a data-intensive company, there is a material difference between:
accessing publicly available information in the absence of an objection
and
continuing systematic extraction after formal notice and technical blocking.
The latter produces not only a different substantive argument but a much stronger evidentiary record.
Notice.
Knowledge.
Blocking.
Continuation.
Alternative IP addresses.
Those facts allow a court to see conduct rather than merely technology.
The defendant’s strongest argument: the content belonged to the users
The defendant raised another important defence.
The plaintiff was, in substance, a hosting platform. The underlying listings were created by users. Therefore, the defendant argued, the platform could not simply treat all information appearing on the website as its own commercial product.
That argument is legally serious.
A user-submitted fact does not become the platform’s intellectual creation merely because the platform displays it.
The court, however, did not analyse the dispute solely at the level of each individual data point.
Instead, it focused on what the plaintiff had done with the data as a collection.
The expert analysis, subsequently relied upon by the court, referred to the plaintiff’s processing and presentation of the information, its classification structure, filtering mechanisms, user-experience investment and technical organisation. The fact that the underlying information originated from users did not, in the court’s assessment, eliminate the platform’s own investment in turning that information into a structured and commercially usable system.
That is an important distinction for AI training disputes:
The absence of protection over an isolated datum does not necessarily answer the legal question concerning the dataset from which that datum was extracted.
One may not own the fact that a property has three bedrooms.
The legal analysis of a systematically constructed database containing millions of classified, filtered and commercially organised records is a different matter.
This distinction is likely to become important in disputes involving price databases, financial data, product catalogues, job listings, legal databases, news archives and other structured sources used in machine-learning pipelines.
The court’s legal route: unfair competition
The operative basis of the judgment was the unfair-competition regime under the Turkish Commercial Code No. 6102 (“TCC”).
Article 55(1)(c) identifies unauthorised exploitation of another’s business or work products as a form of unfair competition.
The provision includes, in subparagraph 3, taking over another person’s market-ready work products by technical means of reproduction and exploiting them without an appropriate contribution of one’s own.
The language is particularly relevant to automated extraction.
The court found that the defendant had used the plaintiff’s ready data through technical reproduction methods and held that the conduct constituted unfair competition under Article 55(1)(c). It ordered determination and cessation of the unfair competition and removal of its consequences.
But this is also where the judgment becomes more difficult.
And more interesting.
The question the judgment does not fully answer: what counts as an “appropriate contribution”?
Article 55(1)(c)(3) is not drafted as a simple prohibition on copying.
The statutory language contains an important qualification:
“without an appropriate contribution of one’s own.”
Turkish scholarship on the provision treats the adequacy of the defendant’s contribution as fact-sensitive. The relevant inquiry may include the work, expenditure and modifications made by the party taking over the original work product. A party may, at least in principle, argue that it did not simply appropriate a market-ready product but made a sufficiently material contribution of its own.
That creates a genuine difficulty in the Bakırköy case.
The court itself recorded that the defendant did more than simply reproduce the listings.
Technical analyses were added to the collected data.
The defendant’s final product was different from the plaintiff’s listing service.
Nevertheless, the court found Article 55(1)(c) applicable.
The judgment does not, however, develop a detailed test explaining why the defendant’s additional analytical contribution was legally insufficient for purposes of Article 55(1)(c)(3).
That omission should not be concealed.
It is one of the points on which an appellate court could provide useful clarification.
There are at least two possible readings.
The first is that the defendant’s downstream analysis was insufficient because the economically relevant appropriation had already occurred when a structured and market-ready body of data was systematically taken and incorporated into another commercial service.
The second is narrower: the outcome may have been heavily influenced by the defendant’s continued collection after formal notices and IP blocking, meaning that the court’s finding should not be abstracted from those particular facts.
The judgment does not clearly choose between those theories.
For that reason, it would be unsafe to extract a general rule that:
“Adding new analysis to scraped data can never amount to an appropriate contribution.”
The court did not formulate such a rule.
What it did establish, on the facts before it, is that adding technical analysis did not save this defendant from an unfair-competition finding.
That is a narrower proposition, but a useful one.
The “sui generis database” point requires similar caution
The expert reports characterised the plaintiff’s organised data structure as having acquired the character of a sui generis database, referring to the platform’s systematic organisation, filtering, categorisation and investment.
The court adopted that reasoning when rejecting the defendant’s argument that user-generated listings could not constitute the plaintiff’s protectable work product.
But the procedural basis must be stated correctly.
The judgment was not a standalone database-right infringement judgment under the Turkish Law on Intellectual and Artistic Works.
The relief was granted under the TCC unfair-competition provisions.
Accordingly, it would overstate the case to write that the Bakırköy court definitively established the scope of the sui generis database right under Turkish copyright legislation.
A more accurate reading is this:
the structured and investment-backed nature of the database was part of the court’s reasoning as to why the platform’s commercially organised data could constitute a protectable work product for unfair-competition purposes.
For AI counsel, the distinction matters because causes of action should not be collapsed into one another.
Copyright.
Database rights.
Unfair competition.
Contract.
Data protection.
They may overlap factually, but they are not interchangeable.
Public accessibility was not treated as a complete defence
Perhaps the most commercially relevant aspect of the decision is therefore not the technology used.
It is the court’s treatment of public accessibility.
The defendant argued that the data were accessible to everyone.
The experts agreed, in technical terms, that the collection did not amount to system intrusion merely because the information was publicly reachable.
The court nevertheless imposed civil consequences because it considered the manner and commercial purpose of the extraction to fall within unfair competition.
The correct lesson is therefore not:
“Public data cannot be scraped in Turkey.”
That proposition is far too broad.
The better formulation is:
Public accessibility does not by itself establish an unrestricted right to systematically appropriate and commercially exploit another undertaking’s organised data resources.
That is a much more defensible rule.
And for AI providers, it is also the more important one.
A 2026 appellate decision now points in the same direction
The Bakırköy judgment was delivered by a first-instance court and was expressly open to appeal.
As of the date of this review, I have not identified a publicly accessible appellate decision specifically resolving Bakırköy 2nd Commercial Court, E.2023/940, K.2025/971. The judgment should therefore be described as first-instance authority unless its appellate status is separately verified.
The broader Turkish case-law picture, however, has since become more interesting.
On 13 May 2026, the Istanbul Regional Court of Justice, 12th Civil Chamber, considered a separate dispute concerning large-scale extraction of content from a publicly accessible news-aggregation application.
The application was open and did not require membership. Nevertheless, millions of automated requests had been made, and the appellate court ultimately found the defendants’ undisclosed data acquisition to constitute unfair competition.
It overturned the first-instance dismissal and awarded TRY 250,000 in pecuniary damages and TRY 50,000 in non-pecuniary damages, together with other relief. That decision itself remained open to further appeal. Istanbul Regional Court of Justice, 12th Civil Chamber, E.2022/2001, K.2026/911, 13 May 2026.
The two cases are not identical and should not be treated as if they were.
But read together, they weaken a simplistic proposition that data become legally unrestricted merely because they can be accessed without a password.
The emerging question appears to be more contextual:
What was taken, how was it taken, at what scale, with what knowledge, from what kind of organised system, and how was it subsequently exploited?
That is much closer to the legal inquiry AI companies will face.
Damages: infringement is not the same thing as loss
The Bakırköy judgment also contains a useful warning for claimants.
The plaintiff established unfair competition.
It did not establish its claimed pecuniary loss.
The court found no sufficient evidence demonstrating that the defendant’s activities had caused a decline in website traffic or revenue. The plaintiff was given an opportunity to submit material supporting its damages claim but did not provide sufficient documentation.
The pecuniary damages claim was therefore dismissed.
The court nevertheless awarded TRY 50,000 in non-pecuniary damages, reasoning in part that a legal entity may possess protected commercial reputation and standing. It also ordered publication of the operative part of the judgment after finalisation, subject to the conditions set out in the ruling.
The evidentiary distinction is important:
Proof of unlawful acquisition is not proof of financial damage.
That will matter in AI training litigation.
A claimant may eventually prove that a corpus was scraped.
It may even prove that particular material entered a training pipeline.
The separate questions will remain:
What economic loss followed?
Was licensing revenue actually lost?
Did the provider avoid a licence fee that would otherwise have been payable?
Was traffic diverted?
Was database value diminished?
What causal link connects the acquisition to the claimed amount?
These questions should not be treated as an afterthought.
What this means for AI training
The Bakırköy case does not tell us whether model training is lawful.
It tells us something that may come first.
Consider a provider building a Turkish-language model.
It collects large volumes of information from a Turkish commercial platform. The individual data may be accessible without authentication. Some may consist largely of factual or user-generated material. The provider does not republish the source database as such. Instead, it preprocesses the material, combines it with other data and uses it in training or retrieval infrastructure.
The provider may be tempted to frame the legal issue exclusively in copyright terms:
Is the source material protected by copyright?
That may be the wrong first question.
The Bakırköy judgment points toward an earlier inquiry:
1. What exactly was acquired?
Individual facts?
Complete pages?
Structured fields?
A commercially useful portion of an organised database?
2. By what method?
Ordinary browsing?
Crawler?
API?
Automated bulk extraction?
Multiple rotating IP addresses?
3. What did the source communicate?
No restriction?
Terms of use?
Crawler restrictions?
Formal notice?
Technical blocking?
4. What did the provider do after learning of the objection?
Stop?
Seek a licence?
Change the collection architecture?
Continue through alternative infrastructure?
5. What happened to the data?
Internal research?
Dataset creation?
Fine-tuning?
Pre-training?
Retrieval-augmented generation?
A paid analytical product?
Those facts may determine the legal characterisation before the court ever reaches the more fashionable questions concerning model weights and outputs.
Data provenance is therefore litigation infrastructure
Much of the current discussion around AI data provenance treats it as a governance concept.
That is too narrow.
It is also an evidentiary concept.
A provider should be capable of answering:
Which source did this data come from?
When was it collected?
Which crawler collected it?
What terms or restrictions applied at that time?
Was the source subsequently blocked or opted out?
Was collection stopped?
Which dataset incorporated the material?
Which model version used that dataset?
If those questions cannot be reconstructed later, the provider may face an evidentiary problem even where a substantive defence exists.
The Bakırköy case illustrates the opposite position.
The litigation record contained identifiable scraping conduct, notarial notices, IP blocking, continued collection through different addresses and expert examination of the parties’ technical systems.
That factual chronology gave the court something concrete to decide.
What the Bakırköy court did not hold
The limits should be stated expressly.
The court did not hold that all web scraping is unlawful in Turkey.
It did not hold that all public internet data belong to the website on which they appear.
It did not decide that AI model training constitutes copyright infringement.
It did not establish a Turkish equivalent of a general prohibition on text and data mining.
It did not find that scraping publicly accessible data was, by itself, hacking or unlawful system intrusion.
It did not establish a personal-data violation on the evidence before it.
It did not explain in detail when a downstream technical contribution becomes an “appropriate contribution” sufficient to take conduct outside Article 55(1)(c)(3).
And it did not issue a final appellate ruling binding higher courts.
Those limitations do not make the judgment unimportant.
They define its importance correctly.
The judgment should not be read as establishing a general prohibition on web scraping in Turkey.
Meydan assessment
The most useful proposition to take from the Bakırköy judgment is simple:
Publicly accessible is not synonymous with legally unencumbered.
The decision moves the analysis away from a crude public/private distinction.
The data were externally accessible.
Much of the underlying information had been supplied by users.
The defendant operated a different service.
It added technical analysis.
The experts did not characterise the extraction as technical penetration of a closed system.
Still, the court found unfair competition.
Why?
Because the relevant legal object was not viewed merely as a collection of isolated public facts. The court considered the commercial organisation of those facts, the platform’s investment in making them usable, the systematic method of extraction, the plaintiff’s express objections, the circumvention of IP blocking and the downstream commercial exploitation of the resulting data.
For AI providers, that sequence should change the order of legal due diligence.
The question should not begin with:
“Is training on this material transformative?”
It should begin earlier:
“How did this material enter our dataset?”
Then:
“What legal position did the source have in the organised dataset?”
Then:
“What did we know when we collected it?”
And only after that:
“What did the model do with it?”
The first serious Turkish AI-training dispute may ultimately be argued under copyright law.
But it may just as easily begin as a dispute about data acquisition, database investment, unfair competition, contractual restrictions or personal data.
The Bakırköy judgment is an early indication of why.
In AI litigation, the provenance of the data may matter before the intelligence of the model does.
Case Information
Court: Bakırköy 2nd Commercial Court of First Instance
Case No.: 2023/940
Decision No.: 2025/971
Decision Date: 17 October 2025
Nature of Action: Unfair Competition and Damages
Principal Provision: Turkish Commercial Code No. 6102, Article 55(1)(c)
Outcome: Partial acceptance; determination and cessation of unfair competition; pecuniary damages dismissed; TRY 50,000 non-pecuniary damages awarded; publication ordered subject to finalisation
Appellate Status: The judgment was expressly open to appeal before the Istanbul Regional Court of Justice. No publicly accessible appellate ruling specifically concerning this case has been identified as of this review.
Related Turkish Authority
Istanbul Regional Court of Justice, 12th Civil Chamber
E.2022/2001, K.2026/911, 13 May 2026
Large-scale data acquisition from a publicly accessible news application; unfair competition found on appeal; decision open to further appeal.
This article forms part of Meydan AI & Tech Law’s Turkish AI Case Review series. The series examines Turkish judgments that, although not always decided as “AI cases”, address legal questions that increasingly arise in AI development, training-data acquisition, platform liability and model deployment.