Trusting the Regulator: South Korea's PIPA Amendment Lets AI Train on Personal Data Without Consent

What the amendment does

According to reporting by the IAPP, the amendment inserts a new set of provisions into PIPA — Articles 28-12 to 28-15 — creating a dedicated legal pathway for using personal data in AI development. The core mechanism is simple to describe and complex to apply: a personal-information controller may use personal data it holds, beyond the original purpose of collection and without obtaining fresh consent, for the development and improvement of AI systems, provided the Personal Information Protection Commission (PIPC) approves the specific project.

Approval is not automatic. The PIPC must be satisfied that four conditions are met:

Anonymisation or pseudonymisation is impractical. The purpose of the AI development cannot realistically be achieved using anonymised or pseudonymised data.
Appropriate safeguards are in place. The controller has put in place technical, organisational and managerial measures to protect the data and the rights of data subjects.
There is a public interest or social benefit. The project serves a public interest or produces benefits for society, rather than purely private advantage.
The risk of unfair infringement is markedly low. The likelihood that the processing unfairly infringes data subjects' rights and interests is significantly low.

The PIPC is also given powers to oversee approved projects, including monitoring compliance with the conditions and the safeguards on which approval was granted. The amendment is designed to enter into force six months after promulgation, giving the Commission time to prepare implementing rules and procedures.

The legislative path

The amendment passed the National Assembly's Political Affairs Committee on 14 May 2026 and cleared the Legislation and Judiciary Committee on 29 July 2026 with bipartisan support — a notable sign of political consensus in a polarised legislature. A plenary vote was expected in mid-August 2026. Readers should check the latest status of the text, as final wording and implementing rules will determine how the pathway operates in practice.

Why Korea is doing this

Several forces converge behind the amendment.

The AI Basic Act. South Korea's Framework Act on the Development of Artificial Intelligence and Establishment of a Foundation for Trust — widely called the AI Basic Act — entered into force on 22 January 2026. It commits the state to promoting AI development and building national AI capabilities. Data availability is a precondition for that ambition.
"Sovereign AI" and the Korean language. Korea wants competitive domestic foundation models that perform well in Korean and reflect Korean contexts. High-quality Korean-language data is limited compared with English, and much valuable data sits inside companies and public bodies where it was collected for other purposes.
The limits of existing tools. PIPA is traditionally one of the world's most consent-centred data-protection laws. The 2020 "three data laws" reform allowed pseudonymised data to be processed without consent for statistics, scientific research and public-interest archiving. But developers argue that pseudonymisation degrades data for some AI purposes, and that re-obtaining consent from millions of people is impractical.
Legal uncertainty. Without a clear pathway, companies faced a choice between not using data, using it and risking enforcement, or relying on untested interpretations. A regulator-approved route promises predictability.

How it fits into Korea's data-protection architecture

The amendment does not arrive in a vacuum. The PIPC has spent several years building a toolkit for AI:

Pseudonymised data. Since 2020, PIPA has distinguished pseudonymised data, which can be used for research and statistics without consent under safeguards, including combination through designated specialised institutions.
The 2023 amendment. A major revision, in force from 2024, modernised consent rules, introduced rights relating to automated decisions and unified standards for online and offline controllers.
Guidance on publicly available data. In 2024 the PIPC issued guidelines on processing publicly available personal information for AI development, clarifying when web-scraped data may be used under legitimate-interest-type reasoning and what safeguards are expected.
Prior adequacy review. The PIPC introduced a system allowing companies to consult the regulator in advance about the privacy design of new AI services and receive a view on compliance before launch.
Active enforcement. The PIPC has shown it is willing to act against large AI and platform companies — for example by fining Meta in 2024 over the collection of sensitive information for advertising, and by restricting new downloads of the DeepSeek app in 2025 until privacy concerns were addressed.

Against that background, the new articles are an evolution rather than a rupture: they extend the logic of "regulator-guided innovation" from advisory opinions to binding legal approvals.

The four conditions under the microscope

1. Impracticality of anonymisation or pseudonymisation

This condition embodies the principle of data minimisation. It forces applicants to explain why less identifying data will not do. The difficulty is technical: for many machine-learning tasks, pseudonymised data works well; for others — such as models that must understand context in free-text records — removing identifiers can undermine the purpose. The PIPC will need technical expertise to evaluate these claims rather than accepting assertions of impracticality at face value.

2. Appropriate safeguards

Safeguards will likely include access controls, secure processing environments, restrictions on outputs, testing for memorisation and data leakage, retention limits and audit trails. A key question is whether safeguards must also address the trained model itself — for example, preventing the model from regurgitating personal data in outputs.

3. Public interest or social benefit

This is the most open-textured condition. Medical diagnostics, public-safety applications and Korean-language models for public services may readily qualify. But what about a commercial chatbot that also benefits consumers? A broad reading could turn the condition into a formality; a narrow one could exclude much private-sector innovation. How the PIPC defines "social benefit" will largely determine the scope of the pathway.

4. Markedly low risk of unfair infringement

The fourth condition requires a risk assessment focused on data subjects. It resembles the balancing test in the GDPR's legitimate-interests analysis, but with a demanding threshold: the risk must be not merely acceptable but "markedly low." Sensitive data, children's data and data that enables profiling are likely to face heightened scrutiny.

"Trusting the regulator": the logic and the risks of ex ante approval

The most distinctive feature of the Korean model is institutional. Under the GDPR, a controller relying on legitimate interests conducts its own assessment and bears the risk if a regulator or court later disagrees. Korea instead places the regulator at the front of the process.

Advantages

Legal certainty. Approved projects gain a clear legal basis, reducing the risk of retroactive enforcement.
Consistency. A central authority can apply consistent standards across sectors.
Expertise and dialogue. Approval procedures can improve privacy design through structured interaction with the regulator.
Public legitimacy. Using data without consent is more defensible when an independent authority has examined and approved the conditions.

Risks

Regulatory capacity. Evaluating complex AI projects requires technical staff and time. Delays could frustrate innovation; rushed reviews could weaken protection.
Capture and political pressure. When national AI competitiveness is a political priority, the regulator may face pressure to approve.
Transparency. If approvals and their conditions are not published, data subjects and civil society cannot scrutinise how their data is used.
Weakened individual control. Approval replaces individual consent with collective, administrative judgement. The question is what rights remain for the individual — for example, a right to object, or to learn whether their data was used.
The model problem. Once personal data has shaped model weights, removing its influence ("machine unlearning") is technically difficult. Approval decisions are effectively irreversible for the data subject.

The constitutional backdrop: informational self-determination

Korea's Constitutional Court has long recognised a right to informational self-determination, derived from constitutional guarantees of human dignity and privacy. That right does not make consent an absolute requirement; the legislature may restrict it for legitimate purposes, provided the restriction is proportionate. The amendment's four conditions can be read as an attempt to build proportionality into the statute: necessity (impracticality of less intrusive data), suitability and safeguards, a legitimate aim (public interest) and balance (low risk). Whether the regime survives constitutional challenge will depend heavily on how rigorously the PIPC applies those conditions in practice.

A worked example: two applications, two outcomes

Consider how the conditions might play out in two hypothetical applications. A university hospital wants to train a model on years of radiology reports to detect early signs of lung disease. Pseudonymisation is possible for structured fields, but free-text reports contain contextual details that are hard to strip without losing clinical meaning. Access would be limited to a secure environment, outputs would not include patient information, and the public-health benefit is clear. Such an application appears well placed to satisfy all four conditions.

Now consider an e-commerce platform seeking to train a general-purpose shopping assistant on customers' purchase histories and chat logs. Much of its purpose could be achieved with pseudonymised or aggregated data; the benefit is primarily commercial; and the data enables detailed profiling. Even if consumers gain convenience, the application would struggle with the first, third and fourth conditions. The contrast illustrates that the pathway is designed less as a general licence for training than as a gateway for projects with a strong public-interest case.

Comparative perspective

European Union. The GDPR has no AI-specific consent exemption, but controllers can rely on legitimate interests, subject to a three-step test. The European Data Protection Board's Opinion 28/2024 set out how that test applies to AI models and when models can be considered anonymous. The GDPR's rules on compatible further processing and on scientific research also play a role. The European Commission's "Digital Omnibus" proposals of late 2025 suggested clarifying legitimate interests for AI development — a proposal still under debate. The EU AI Act adds data-governance obligations for high-risk systems but does not itself provide a legal basis for processing personal data.
Italy. The Italian data-protection authority fined OpenAI EUR 15 million in December 2024, citing among other things the lack of an adequate legal basis for training — a reminder of the enforcement risk under self-assessment models.
United Kingdom. Post-Brexit reforms have sought to make research and legitimate-interest processing more predictable, but the UK still relies on controller self-assessment rather than prior approval.
Japan. Japan's data-protection framework is comparatively flexible for AI development, and reforms under discussion have explored easing consent requirements for statistical and AI-related uses.

Korea's approach is therefore a genuine third way: more permissive than a strict consent regime, but more controlled than the self-assessment model. For a broader map of how jurisdictions regulate AI, see our global AI law index.

What it means for Turkey

For Turkey, the Korean debate is highly relevant because the KVKK shares PIPA's consent-centred heritage while aspiring to closer alignment with the GDPR.

No dedicated AI pathway. The KVKK contains no provision allowing personal data to be used for AI training without consent under regulatory approval. Controllers must find a legal basis among the conditions in Article 5, or Article 6 for special categories.
Legitimate interest. Article 5(2)(f) permits processing necessary for the controller's legitimate interests, provided it does not harm the data subject's fundamental rights and freedoms. It is the most likely basis for some AI training, but its application requires careful balancing and documentation, and its limits for large-scale training remain untested.
Purpose limitation. The KVKK requires processing for specified, explicit and legitimate purposes, and in a manner that is relevant, limited and proportionate. Repurposing customer data for model training therefore raises immediate questions.
Anonymised statistics. Article 28 excludes from the law's scope processing for research, planning and statistics through anonymisation with official statistics — a narrow exception that does not cover most commercial AI development.
No pseudonymisation regime. Unlike PIPA and the GDPR, the KVKK has no specific regime treating pseudonymised data more flexibly.
Special categories. After the 2024 amendments, special-category data can be processed on a limited list of grounds; there is no general AI or research ground.

We have examined these questions in depth in training data: the KVKK and GDPR questions to settle first, in our guide to KVKK and AI governance and, for cross-border scenarios, in training AI across borders. The 2026–2030 Turkey AI Action Plan, which we reviewed in "The State's Smart Route," emphasises data availability — making the legal basis for training a policy question Turkey cannot avoid.

Should Turkey copy Korea? A wholesale transplant would be premature. But elements of the model are worth considering: a regulatory sandbox or prior-consultation mechanism at the Personal Data Protection Authority for AI projects; published criteria for using legitimate interest in training; a pseudonymisation framework; and mandatory transparency about approved projects. Any such reform should be accompanied by strengthened regulatory capacity and effective remedies for individuals.

A compliance roadmap for companies

Map your training data. Identify the sources of personal data used in AI development, the original purpose of collection and the legal basis relied on.
Test less intrusive options first. Document whether anonymised, pseudonymised or synthetic data can achieve the purpose. In Korea this is a legal condition; elsewhere it is evidence of proportionality.
Articulate the benefit. Describe concretely who benefits from the AI system and how; vague claims will not satisfy a regulator.
Assess risks to individuals. Conduct a data-protection impact assessment covering training, memorisation, outputs and downstream uses.
Design model-level safeguards. Test for regurgitation of personal data, filter outputs and control access to model weights.
Be transparent. Update privacy notices to explain AI uses, and publish information about approved or legitimate-interest-based training where possible.
Prepare for divergence. Global companies may need different data strategies for Korea, the EU and Turkey. Build data governance that can track jurisdiction-specific permissions.

Data governance for training does not end with personal data; the legal pedigree of copyrighted content matters just as much, as our report on Anthropic's $1.5 billion copyright settlement shows.

What to watch next

Several developments will reveal whether the Korean model delivers. First, the final text and the PIPC's implementing rules: how applications are made, how long review takes and what evidence is required. Second, the first approvals: which sectors receive them, how "social benefit" is interpreted and whether decisions are published. Third, the rights left to individuals, particularly objection and transparency. Fourth, any constitutional or civil-society challenge. Finally, whether other jurisdictions — including Turkey — begin to experiment with regulator-approved pathways for AI training.

Frequently asked questions

Does the amendment let any company train AI on personal data without consent?

No. It creates a pathway that requires case-by-case approval from the PIPC, and approval depends on four conditions, including that anonymisation or pseudonymisation is impractical and that the risk to individuals is markedly low.

When does it take effect?

The amendment is designed to enter into force six months after promulgation. Its practical operation will depend on the PIPC's implementing rules.

Does it apply to sensitive data?

The details depend on the final text and implementing rules, but sensitive data will almost certainly face heightened scrutiny under the safeguards and risk conditions.

How is this different from the GDPR's legitimate interests?

Under the GDPR, the controller assesses legitimate interests itself and bears the risk if it is wrong. Under the Korean model, the regulator approves the specific project in advance on the basis of statutory conditions.

Can individuals object?

This is one of the key open questions. General PIPA rights continue to apply, but how they interact with approved AI projects — especially once data has shaped a model — will need clarification.

Could Turkey adopt a similar model?

It would require a legislative amendment to the KVKK. Before any transplant, Turkey would need to consider regulatory capacity, transparency, remedies and alignment with its broader GDPR-oriented reform agenda.

Does this affect foreign companies?

Foreign companies processing personal data of people in Korea are subject to PIPA. Those wishing to use the new pathway would need to meet its conditions and obtain approval like domestic controllers.

Expert Opinion

This section reflects my personal assessment as the founder of this site and an AI ethics & compliance counsel.

In my view, Korea's amendment is an honest answer to a question most jurisdictions are avoiding. Everyone knows that consent cannot realistically be obtained from millions of people for every AI training project, and that a great deal of training is already happening on uncertain legal ground. The GDPR world tolerates that uncertainty through self-assessment and after-the-fact enforcement. Korea chooses transparency about the trade-off: it says openly that some data may be used without consent, and it puts a public authority, bound by statutory conditions, in charge of deciding when. I find that more intellectually honest than pretending consent still governs everything.

But the model is only as strong as the regulator behind it. "Trusting the regulator" works if the regulator is independent, technically capable, transparent and willing to say no. If approvals become a rubber stamp for national AI ambitions, the law will have converted a fundamental right into an administrative formality. The four conditions are well chosen, particularly the requirement that less identifying data be shown to be impractical, but the "public interest or social benefit" test is dangerously elastic. I would want to see every approval, its conditions and its safeguards published, and a meaningful right for individuals to object and to know.

For Turkey, my recommendation is not to copy the Korean text but to learn from its structure. The KVKK Authority could start with a prior-consultation and sandbox mechanism for AI projects, publish clear criteria for legitimate-interest-based training, and introduce a pseudonymisation framework in future reforms. The goal should be the same one Korea is pursuing: legal certainty for responsible innovation, without asking people to surrender control over their data on trust alone.

This article is for information only and does not constitute legal advice. Facts about the amendment are based on publicly available reporting, principally by the IAPP; the final text, its status and implementing rules should be verified. The analysis and assessments are the author's own.