Executive Summary

Until now, analyzing or using personal data in Korea has run, in practice, through a gate: pseudonymization, the masking of identifiers such as names and resident registration numbers. The path to using raw data was narrow, and pseudonymization served as the last safeguard protecting privacy.

The problem surfaced as the kind of data changed. With tabular, structured data, erasing identifiers does not badly damage statistical value. With video, voice, and images, the story is different. The moment a face is blurred, a voice is altered, and a background is wiped, the very signals AI needs to learn, the subtle shifts in expression or the paths people took through a disaster scene, vanish along with them. Industry has argued that pseudonymization keeps data safe at the cost of destroying its training value.

The Third Basic Plan's exemption is aimed squarely at this point. For personal data that has already been collected lawfully, it creates an exception that allows use in raw form for the purpose of developing AI. It is more accurate to say that the safeguard was not removed outright, but that a different form of control was set as a condition, opening the door precisely where the nature of the data means pseudonymization undermines the purpose.

The exemption should not be read as unconditional approval. A company that wants to use raw data has to pass through two gates. One is the PIPC's review; the other is a risk-factor assessment that weighs, in advance, the risks that will arise from using the data. The door has opened, but a checkpoint stands in front of it.

That checkpoint did not appear out of nowhere. Back in December 2024, the PIPC had already established an AI Privacy Risk Assessment and Management Model, helping AI companies diagnose and manage their own risk. This Basic Plan is closer to an expansion built on that foundation, widening the scope of application to reach raw data.

The scope of what gets assessed widens along with it. Emerging technologies enter the frame: the accountability structure of agentic AI that decides and acts on its own, the rights protections owed by physical AI that gathers data from its surroundings in real time, and the prevention of cyber threats and data manipulation that abuse AI. At the same time, three sectors, telecommunications, education, and employment, are designated as high-risk areas, where the PIPC and the relevant ministries inspect and manage jointly. That is why any company holding data should first work out which gate it is standing in front of.

The phrase running through this Basic Plan is a shift from uniform regulation to a principle-based system proportionate to the level of risk. Instead of drawing the same prohibition line across all data, it clears a path for low-risk uses and concentrates control on high-risk ones. The center of gravity moves from "do not use" to "manage in a way that fits the risk."

The direction also moves toward prevention. Judging that heavier penalties after an incident can only go so far, the Commission is pursuing, in parallel, a CEO accountability rule that names executive responsibility, a coercive fine that raises the effectiveness of investigations, and a consent-decree scheme that waives sanctions when a company puts sufficient recurrence-prevention measures and victim compensation in place. Sharpening punishment and widening the path to lawful use sit inside a single plan. The PIPC itself described this three-year plan as centered on redesigning the personal data governance system for the AI environment and establishing a prevention-first protection framework.

Once the path to using raw personal data opens, the shape of the shield a company can raise before regulators changes too. Until now, the safest answer was a denial that it had neither collected nor used the data. But the moment such use enters the realm of the lawful, "we didn't use it" becomes a choice that falls behind in competition. The new shield becomes "this is how we managed it," a demonstration of who used which data, through what risk assessment, and how.

That demonstration cannot be made in words. Where the data came from and within what consent and purpose scope it was collected, who accessed it under what authority during training, and which risk assessment it cleared, all of this has to be attached to the data as its history. Being able to unfold that lineage the moment a reviewer asks is the practical condition for passing the exemption.

Here the standing of data governance and provenance changes. They are no longer a nice-to-have management activity but a necessary condition for using raw data lawfully. Compliance moves from an after-the-fact response to a matter of infrastructure, something that has to be designed at the very entrance where data enters the pipeline.

The Basic Plan also carries mechanisms to help companies: an AX Reassurance Support Center that resolves legal uncertainty, an expansion of regional data safe zones for handling pseudonymized and anonymized data securely, and an upgrade of the MyData platform that will widen from an initial ten fields into welfare, care, and healthcare. Yet however dense this support is, the responsibility for proving the right to use the data still rests, in the end, with the company that holds it.

So the checklist for any organization that wants to handle raw personal data is clear.

• A system that records the source of the data, its lawful basis, and the scope of consent
• A process that folds a risk assessment into the stage before training and runs it on an ongoing basis
• A lineage-tracking structure that keeps access and processing history traveling with the data

The detailed implementing rules have yet to be announced, and much of the pass criteria and scope for risk assessments remains to be spelled out. But the direction is already set. The three years starting in 2027 are less about regulation that blocks data from being used than about regulation that requires proof it was used well. For organizations that prepare that proof in advance, the door stands open; for those that do not, it becomes a checkpoint they cannot pass. Thank you for reading.

Official Documents

• 1.Personal Information Protection Commission. (2026). Third Basic Plan for Personal Information Protection (2027–2029): Promoting Trust-Based AI Innovation. Announced at the Ministerial Meeting on Economic Affairs (2026-07-03).
• 2.Personal Information Protection Commission. (2024). AI Privacy Risk Assessment and Management Model.

Industry & Press

• 3.Seoul Economic Daily. (2026). "Korea Overhauls Data Privacy Framework for AI Era." en.sedaily.com
• 4.ZDNet Korea. (2026). "PIPC to Shift to Flexible AI-Applied Regulation: Third Basic Plan for Personal Information Announced." zdnet.co.kr
• 5.Aju Business Daily. (2026). "PIPC Announces Third Basic Plan for Personal Information Protection: A Shift to AI-Era Tailored Regulation." ajunews.com
• 6.The Korea Economic Daily. (2026). "PIPC Reforms the Personal Information Protection System for the AI Era." hankyung.com
• 7.Financial News. (2026). "Government Pursues a Consent-Decree Scheme for Personal Data Breach Incidents." fnnews.com