The regulator that ordered a model destroyed

Case in one sentence

Korea’s data protection regulator found that Kakao Pay had sent the personal data of 40 million users to Alipay every day for nearly five years so that Alipay could compute, for Apple, a score predicting whether each user would have enough money to pay — and rather than only fining the companies, it ordered the model itself destroyed.

Executive summary

On 22 January 2025 the Personal Information Protection Commission decided sanctions against three companies over a single data pipeline.¹

Apple had engaged Alipay Singapore E-Commerce Private Limited for system integration work on payment methods, including calculating a “Non-Sufficient Funds” score — a customer-specific score indicating whether a customer would have enough money to cover their App Store transactions. Kakao Pay supplied the data that made the score possible.¹

It did so twice over. Between April and July 2018 it sent Alipay the data of about 16 million users on three occasions, in 24 categories, specifically to build the scoring model. Then, from 27 June 2019 to 21 May 2024, it sent the data of 40 million users — every one of its users, de-duplicated — every single day, to keep the scores current.¹

The regulator’s findings were specific. Kakao Pay obtained no consent for the cross-border transfer. Apple never listed Alipay in its privacy policy as an overseas entity entrusted with processing, though it did list a different payment company. And Kakao Pay sent the data of Android users too, despite iOS users being under 20% of those who had registered Kakao Pay as an Apple payment method.¹

The sanctions: KRW 5.97 billion against Kakao Pay, KRW 2.4 billion plus a KRW 2.2 million fine against Apple, correction and publication orders against both.¹

Against Alipay — which had received the data rather than collected it — the Commission issued no fine at all. It issued a correction order to destroy the model. Its stated reason was to address the root cause of the violations.¹

In June 2026 the Seoul Administrative Court dismissed all of Kakao Pay’s claims and upheld the penalty.²

What no source establishes is whether the model was destroyed.

Research question

When a model is built from data that should never have been supplied, what is the appropriate remedy — and can a regulator that orders the model destroyed ever know whether it was?

Organization and operating context

Three companies in three jurisdictions, connected by one payment flow. Kakaopay Corp. is a Korean payment provider. Apple Distribution International Limited is Apple’s Irish entity. Alipay Singapore E-Commerce Private Limited is the third-party payment platform Apple engaged.¹

The commercial logic is ordinary. Apple bundles small App Store transactions and settles them together, which creates a risk that a customer’s payment method will have insufficient funds when settlement comes. A score predicting that risk is a reasonable thing to want. Apple entrusted the system integration to Alipay, including the API work that lets Korean payment methods work with its services, and the NSF calculation came with it.¹

What made this a case rather than a product is where the training data came from and who agreed to it. Alipay was the only route by which NSF scores were calculated among the domestic payment methods linked to users’ Apple IDs, and Kakao Pay was the only Korean provider whose data flowed through that route.¹

The situation before AI

Korea’s Personal Information Protection Act requires separate consent for a cross-border transfer of personal data, and requires an organisation entrusting processing to an overseas entity to say so in its privacy policy. The Commission’s own summary of the law is that where a third party assumes responsibility for processing, the arrangement stops being entrustment and becomes a transfer needing its own legal basis — typically consent.¹

Apple’s privacy policy did name an overseas system-integration entrustee: NHN KCP, a payment gateway. It did not name Alipay. Users therefore had no way of knowing Alipay was processing their data.¹

The AI intervention or event

Building the model. From April to July 2018, on three occasions, Kakao Pay extracted the records of about 16 million users in the form Alipay requested and sent them across. Twenty-four types of data: hashed internal customer numbers, phone numbers, email addresses, and the features that would carry the predictive signal — join date, whether the account had been verified against the holder’s ID, balance, and counts of top-ups, payments, and remittances over the previous seven days.¹

Those last fields are the case in miniature. They are a behavioural profile of a person’s money over a rolling week, and they were the training data for a model estimating whether that person would come up short.

Running the model. From 27 June 2019 until 21 May 2024, Kakao Pay sent the personal information of 40 million users daily. Alipay generated a score for each user, updated the model on the incoming data, and returned scores to Apple on request.¹ Five years of continuous retraining on a national population’s payment behaviour.

The over-collection. Kakao Pay sent everyone. iOS users accounted for under 20% of the users who had registered Kakao Pay as a payment method for Apple services, and Android users have no App Store payments to score at all. Their data went anyway.¹

The remedy. The Commission fined the company that supplied the data and the company that benefited from the score. For the company that built the model it chose something different: a correction order to destroy it, on the reasoning that the model was constructed on unlawfully transferred data and destruction addresses the root cause.¹

Outcomes and economics

All from the Commission’s own release.¹ The monetary total is roughly KRW 8.37 billion, about USD 5.8 million at the rates the Commission used.

The scale on the other side of the ledger: 40 million people’s data, moved daily for 1,790 days.

Kakao Pay sued to overturn the sanctions. In June 2026 the Seoul Administrative Court dismissed every claim.² Kakao Pay maintained it had transferred the data under lawful procedures and with strong encryption to meet a duty to prevent fraudulent payments, said it regretted the judgment, and said it would decide its response after reviewing the written ruling.²

A separate review of the same conduct under the Credit Information Use and Protection Act was noted as pending when the PIPC decided.¹ It has since produced a larger number than the privacy regulator’s. In April 2025 the Financial Supervisory Service announced a 15 billion won (about $10 million) sanction recommendation against Kakao Pay for handing roughly 40 million users’ data to Alipay without consent — a finding it drew from an inspection of the company’s foreign exchange trading. The final amount rests with the Financial Services Commission, and no source read here confirms the FSC’s determination, so this case treats the 15 billion won as recommended rather than imposed.³

Two details from that action are worth keeping. The FSS can set a penalty at up to 3% of revenue under the credit-information law, a ceiling well above what the PIPC applied. And Kakao Pay’s position to the financial regulator is the same one it later argued in court: that the data went across to provide Apple App Store payment services to its own users.³

On the destruction order there is no economic figure and no confirmation. Nobody has published what the model was worth, whether it was destroyed, or what became of the scores already generated.

Causal assessment and competing explanations

This case is labelled descriptive. It records a regulatory finding, the conduct behind it, and a court’s review. It makes no claim that the model caused any measurable outcome for any user — indeed no source read here says what Apple did with an NSF score once it had one, whether any transaction was declined because of one, or whether the scores were accurate.

That absence is worth sitting with. The entire enforcement action concerns how the model was built and fed, not what it did. A reader looking for harm to individuals from the scoring itself will not find it documented here, and this case does not supply it by implication.

One competing explanation deserves a hearing, because the company advanced it and a court rejected it. Kakao Pay’s position is that the transfers were lawful, encrypted, and made in service of a genuine duty to prevent fraudulent payments for Apple services.² The court’s answer was not that fraud prevention is illegitimate but that consent was absent and could not be inferred: users could not be regarded as having agreed to their information being used as a credit-assessment metric, and the transfer of all users’ data, Android included, told against the necessity argument.²

A second reading — that this is really a cross-border data case with an AI model incidentally attached — is half right and is why the case is filed as governance-regulatory. The violations found are consent and disclosure violations under transfer rules. What makes it belong in this library is the remedy, which is aimed at the model rather than at the data.

Failures, limitations, and governance

• The data flow outlived any plausible consent. Whatever a user agreed to in 2018 or 2019, five years of daily transfers of a whole national user base to a foreign processor is not something anyone was asked about.¹
• Everyone was in the training set, including people the product could not apply to. Android users’ payment behaviour fed a score for App Store transactions they could not make.¹
• The disclosure gap was specific, not systemic. Apple’s privacy policy named one overseas entrustee and omitted the one that mattered, so users could not have discovered Alipay’s role by reading it.¹
• Responsibility diffused across three companies in three jurisdictions. The collector, the beneficiary, and the builder were separate entities, and only the collector and beneficiary were fined. The builder — the one holding the asset — paid nothing.
• The remedy has no published verification. An order to destroy a model raises an obvious question that no source in this record answers: how would anyone know? Weights can be copied, derived scores persist, and successor models trained on the same signal are not obviously covered.
• Derived outputs are not addressed. The order concerns the model. Scores already generated and delivered to Apple are not discussed in the release read here.¹
• The second regulator’s action is recommended, not final. The FSS proposed 15 billion won in April 2025 — more than twice the PIPC’s penalty — but the Financial Services Commission sets the final figure, and no source read here records it.³

What this case demonstrates

1. A regulator can treat a trained model as the thing to be remedied rather than as an incidental artefact of a data breach, and Korea’s did so explicitly to address what it called the root cause.¹
2. The remedy can fall on a party that was neither the collector nor the primary beneficiary. Alipay received data it did not gather, for a customer that was not the data subject, and it is the party ordered to destroy an asset.
3. Model deletion is not unprecedented. The US Federal Trade Commission required deletion of “models and algorithms” developed from unlawfully used photos in the Everalbum settlement four years earlier.⁴ What is newer in Korea is the target: a third-party processor rather than the company that misled its own users.
4. Continuous retraining turns a one-off consent problem into a standing one. Three transfers built the model; 1,790 daily transfers kept it current, and each day was its own violation.
5. Over-collection is legible after the fact and was decisive here. Sending Android users’ data to score App Store payments was the fact the court cited when it upheld the penalty.²
6. A privacy policy that names some overseas processors and omits one is worse than uninformative, because it looks complete.
7. Enforcement of a distributed pipeline requires naming every link. The Commission issued a different remedy to each of the three companies, matched to what each had done.
8. Judicial review is where these findings become durable. The PIPC’s decision is now a court-tested one, at least as to Kakao Pay.²

What this case does not demonstrate

• It does not show the model was destroyed. No source read here confirms compliance, describes verification, or says who would check.
• It does not show any user was harmed by an NSF score. No source describes a declined transaction, an adverse decision, or the model’s accuracy.
• It does not establish what Apple did with the scores it received beyond requesting them.
• It does not resolve the case as to Apple or Alipay. The June 2026 ruling read here concerns Kakao Pay’s suit; nothing in this record says whether the other two challenged their orders.
• It does not settle the matter finally even for Kakao Pay, which said it would decide its response after reviewing the written judgment.²
• It does not establish the final penalty under credit-information law. The FSS recommended 15 billion won; the FSC’s determination is not in any source read here.³
• It does not describe machine learning architecture. The sources say a model was built, fed daily, and updated; none says what kind.
• It does not establish that model destruction is an effective remedy. This case is evidence that a regulator will order it, not that ordering it accomplishes anything.

Evidence assessment

Grade B — an authoritative primary decision, a court ruling reached through one press account, and an unverified central remedy.

The core of the record is strong. The PIPC published its own findings company by company, with dates, volumes, data categories, the legal provisions breached, and the sanction imposed on each.¹ That is a regulator’s account of its own decision, which is primary evidence of the decision even where it is an interested account of the conduct. The sanction against Kakao Pay has since survived judicial review, which is the strongest available corroboration that the findings were sound.²

Four things hold it at B.

The central remedy is unverified. The case is interesting because a model was ordered destroyed. Whether it was destroyed is unknown, and nothing read here indicates that anyone has checked or how they would. A case whose headline fact is a remedy, with no evidence of the remedy’s execution, cannot be graded as though the remedy were established.

The judicial review is a single relay. The Seoul Administrative Court’s written judgment was not read. It reaches this case through one newspaper report, which quotes the ruling but does not reproduce it, and which discloses that it was translated from Korean with the help of generative AI tools before editing. Kakao Pay’s own statement in that article notes it had not yet reviewed the written judgment either.²

Everything is in translation, and the regulator says so. The PIPC’s English release states on its face that it is an unofficial translation of a Korean original. No Korean-language source was read for this case. Figures and legal characterisations should be read with that in mind.

A second regulator has now been heard, through press. The Financial Supervisory Service’s April 2025 sanction recommendation is the closest thing in this record to an independent finding on the same conduct: a different regulator, a different statute, a larger number.³ It is reported rather than read — the FSS’s own decision document is not in this record — and it is a recommendation, not a final penalty. It does not move the grade, because the case’s central fact remains an unverified destruction order, but it does retire the claim this case previously had to label Unknown.

The figures disagree at the margins. The PIPC gives Kakao Pay’s penalty as KRW 5.97 billion; the Korean press report gives 5.96 billion in its headline and 5.97 billion in its body; a policy tracker cited during research gives 5.968 billion and a combined total of KRW 8.3752 billion, against the PIPC’s itemised KRW 2.4 billion for Apple where the press gives 2.41 billion. The differences are rounding, and this case uses the regulator’s own numbers throughout, but they should be recorded rather than smoothed. The decision date is similarly reported as both 22 and 23 January 2025; the PIPC’s release states the plenary decided on 22 January and is itself headed 23 January.

One analytical correction belongs here rather than in the narrative. The practitioner analysis cited frames model deletion as putting South Korea “at the leading edge of global privacy practice”, noting the FTC has “occasionally” required algorithm disgorgement.⁵ Reading the FTC’s Everalbum announcement directly shows the precedent is squarely on point: the 2021 proposed order required deletion of the models and algorithms developed from users’ photos.⁴ The Korean order is genuinely distinctive, but for a different reason than novelty of the remedy — it was directed at a processor that received the data from someone else, rather than at the company that had misled its own users.

Conflicts are ordinary and disclosed. The PIPC is describing its own action and writes its takeaways as industry guidance. The FTC release is the same kind of document from the same kind of body. The practitioner analysis is professional commentary by someone working in the market it describes, and only its ungated portion was read.

Material claims

Direct quotations

“Meanwhile, Alipay’s model for calculating NSF scores was built upon the unlawful provision and cross-border transfer of personal information. To address the root cause of violations of the PIPA, the PIPC issued a correction order on the third-party payment platform to destruct the model.”

— Personal Information Protection Commission, in the sanctions section of its own decision announcement; the English text is the Commission’s unofficial translation of the Korean original¹ · locator: Section II, Sanctions, paragraph 3, “Alipay: A correction order”

“Separately, a correction order was issued on Alipay to destruct a model for calculating Non-Sufficient Funds (NSF) scores built upon the personal information provided without obtaining consent from users”

— Personal Information Protection Commission, the subheading of the press release¹ · locator: subheading, immediately beneath the title

“The plaintiff did not obtain consent from users of its payment service when it provided their information to Alipay. It is hard to conclude that the data subjects recognized, or specifically and clearly agreed to, the information being used by Apple as a kind of credit-assessment metric to evaluate customers’ ability to pay.”

— The Seoul Administrative Court, quoted in press reporting of the June 2026 ruling; the written judgment was not read² · locator: article of 2026-06-11, first quotation from the ruling

“Users cannot be regarded as having consented to the effective nullification of their right to control their own personal information in the process of calculating NSF information.”

— The Seoul Administrative Court, dismissing Kakao Pay’s claims² · locator: article of 2026-06-11, second quotation from the ruling

“Kakao Pay fully explained why the PIPC’s sanctions were unfair, and we regret that the court made a different judgment. We plan to decide our future response after reviewing the written judgment.”

— Kakao Pay, in a statement after the ruling² · locator: article of 2026-06-11, company statement

“The proposed order also requires the company to delete models and algorithms it developed by using the photos and videos uploaded by its users.”

— U.S. Federal Trade Commission, describing the Everalbum settlement four years earlier⁴ · locator: press release of 2021-01-11, second paragraph

Revision notes

• 2026-09-13 — Source review. Added the Financial Supervisory Service’s April 2025 sanction recommendation, which retires the claim this case had recorded as Unknown: the financial regulator did find violations in the same conduct, and proposed 15 billion won — more than twice the PIPC’s penalty — under a statute allowing up to 3% of revenue. The figure is treated as recommended rather than imposed, because the Financial Services Commission sets the final amount and no source read here records its determination; a February 2026 report of a final figure was found only behind a paywall and is therefore not cited. The grade stays at B: the case’s central fact is still an unverified order to destroy a model. A further rounding variant is recorded — the Korea Times gives the PIPC fine as 5.9 billion won against the PIPC’s own 5.97 billion.
• 2026-09-13 — Initial publication at Grade B. Built on the Commission’s own English press release, which states on its face that it is an unofficial translation of a Korean original; no Korean-language source was read. Held at B rather than A for four reasons stated in the evidence assessment, the first being that the case’s headline fact is a destruction order whose execution nothing in the record verifies. deployment_stage is recorded as unknown rather than retired for the same reason: the model was ordered destroyed and no source says it was. Labelled descriptive — the record concerns how the model was built and fed, not what it did to anyone. Records the figure and date discrepancies rather than smoothing them, and corrects one analytical claim by reading the FTC precedent directly: model deletion had precedent, and what distinguishes the Korean order is that it fell on a third-party processor rather than on the company that misled its own users. First case in the library set in Asia.