Busan’s Twelve-Year School Account Faces a Test Beyond Access

PENZ brings commercial cloud services under a common educational identity, while BeAT offers AI tutoring to high schools. The city has begun building the infrastructure. Its larger undertaking is to protect students’ work, develop their judgment and show what they learn when the technology is no longer doing the work for them.

In May 2025, a student at Busan International High School asked the education authorities to help make Canva available for assignments and club activities. The proposal, later described in the Busan Metropolitan City Office of Education’s own account of the project, concerned a practical problem: educational software could be useful to a student without being accessible through the institutional arrangements available to her school. The education research institute was already planning a unified account system and subsequently considered how to include Canva. It would exaggerate the student's role to say she created the system, but her request captured something important about the way digital tools had entered public education. Schools could encourage pupils to become competent users of software while leaving their ability to obtain it dependent on separate accounts, licenses and negotiations with the companies supplying the tools.

Busan had reasons to prefer coordination over another stand-alone education platform. The office said that 87 percent of teachers in a survey it cited in July were already using Google Classroom or Microsoft Teams, although the published announcement provided insufficient detail to independently judge the survey. Teachers had developed lessons and routines around products students also encountered outside school. Replacing those applications wholesale would have created more upheaval; arranging access across the education system offered a way to support familiar work while reducing the repetitive administration of accounts. There was an equity argument, too. Schools and families do not all have the same capacity to purchase subscriptions or navigate eligibility rules, even when students are expected to produce similar kinds of assignments.

PENZ, introduced across Busan in September 2026, was the institutional response. Its published instructions distinguish enrollment in Education Digital Onepass, the national education authentication service, from PENZ membership and registration with individual providers such as Google, Microsoft, Adobe and Canva. Rather than building replacements for every application, the education office established a common gateway to a collection of existing services. In July, it attached an unusually broad promise to the arrangement: a student would keep an account from the first year of elementary school through high school graduation, with a learning portfolio continuing for twelve years without the data loss associated with changing school accounts. Integration with the national school administration system, NEIS, was also intended to automate account creation and grade progression, freeing teachers from routine account management.

The first results concerned access, as might be expected of a newly opened gateway. By September 17, the education office reported more than 10,000 registrations and 200,000 cumulative visits, together with orientation sessions attended by 1,410 school administrators and staff members. Those figures, reported by the operator rather than independently measured by this newsroom, demonstrate initial uptake of a public service. They do not establish that earlier schoolwork has survived a transfer or that teachers have saved a particular number of hours. Still, the underlying achievement should not be dismissed. Making educational tools available through a coordinated public arrangement can solve problems that neither a family nor an individual teacher ought to have to negotiate repeatedly.

The twelve-year promise introduces a different responsibility. A pupil remains the same person as grades and schools change, but the pupil’s teachers, classes, software entitlements and access to shared material do not remain fixed. A report written independently, a presentation edited with classmates and a teacher’s reference document may all appear in the same workspace while carrying different permissions. Retaining the student's work need not justify continued entry into a former class’s restricted files; ending an old license need not extinguish the educational value of an essay or project. Busan's approach makes consistent management possible, but a shared identity cannot settle ownership or decide which permissions should survive a change of school. Those are institutional judgments, and the public authority has made their reliable coordination part of the value it promises.

Bringing AI Into the Classroom

BeAT, the education research institute's AI tutoring service, raises the stakes because it draws the work of learning itself into an application. Its teacher-facing service opened in April, followed by student access in June, before the broader PENZ rollout. The institute describes an environment in which teachers can provide instructional material and configure tools for their classes, while students ask questions, upload work, seek feedback and download selected conversations as Word documents. It is designed to be closer to a course than a general-purpose chatbot: the material a teacher chooses can help shape explanations, and a pupil can return to a difficult idea without waiting for another opportunity to ask in class. Those are plausible educational advantages, especially for students who hesitate to expose confusion before their peers. Their realization depends on what happens during the exchange, rather than on the existence of the software alone.

A July 16 hearing of the Busan City Council’s Education Committee offered a useful picture of the program's first weeks. Yu Young-ok, director of the Busan Education Research and Information Institute, told council members that 137 of the city's 147 high schools had enrolled in BeAT and that 1,310 teachers and 1,838 students were using it. She put the potential high-school student population at approximately 70,000. The student service had opened only in late June, making any judgment about mature demand premature. What the testimony exposed was the distance between administrative distribution and classroom practice. A school can join a system without making it part of every course, and a student's first experiment with an AI tutor tells us little about whether the student has learned something that can be demonstrated independently.

Yu also explained that adoption for student instruction could be slowed by school operating committee procedures. The delay should not automatically be treated as resistance to innovation. Under rules governing the selection of qualifying learning-support software, schools must examine privacy safeguards and submit the adoption of such products for committee review. The Education Ministry's guidance makes clear that a common login does not transform different applications into a single educational choice; tools handling student information or supplying specified curriculum-related learning content can require scrutiny of their individual functions and evidence, including publicly developed software. Central procurement can reduce repetitive contracting and help schools obtain supporting documentation, but it does not relieve a school of deciding whether a particular application belongs in its teaching. Nor does a compliant checklist settle whether the answers the application produces are accurate or instructive.

That question matters because BeAT's strongest advertised function concerns the point at which a learner has not yet succeeded. In the council hearing, Yu described a student photographing an incorrectly solved mathematics problem and receiving an explanation of the error, followed by questions intended to encourage another attempt. The account was a description of functionality, not an observed demonstration of learning gains. Yet it identifies an important educational choice. An AI system could spare the student the work of finding a solution, or it could help the student understand where a line of reasoning broke down and return to the problem with greater independence. Both might make a homework session feel easier. Only the latter necessarily advances the skill the lesson was meant to develop.

Research provides a reason to take the difference seriously. In a field experiment published in Proceedings of the National Academy of Sciences in 2025, researchers studied nearly 1,000 students in mathematics classes at a high school in Turkey. Students using GPT-based help performed better during assisted practice, but those given a relatively unrestricted, chatbot-like tutor scored 17 percent below a control group on a later examination taken without AI. A second version, designed to offer teacher-informed hints instead of readily giving answers, largely mitigated that decline; it did not produce a statistically significant advantage on the unaided examination. The experiment was not a test of BeAT, and its findings should not be treated as an estimate of the effect on Busan students. It demonstrates something more relevant to the policy decision: success at an AI-assisted task can conceal weaker acquisition of the skill required to perform the task alone, while the way assistance is designed can materially change that outcome.

Busan has chosen a model that could address part of this problem. The institute describes the use of teacher-supplied sources, error analysis and guiding questions, and Yu told a September budget committee hearing that BeAT operated as a GPT-based service in Microsoft's Azure environment. The teacher's role is consequently more than encouraging students to write better prompts: someone has to select suitable material, decide how much help is pedagogically appropriate and teach pupils to inspect an answer that sounds authoritative even when it is mistaken. A response based on a teacher's documents may be easier to trace to the curriculum, but that does not guarantee a correct interpretation of the documents. The capacity to challenge the explanation remains part of what students are supposed to learn.

Teaching Judgment, Not Just Use

It would be inaccurate to claim that Busan introduced AI without making provision for ethics or digital literacy. The education office announced guidance for teachers and learners and school-based AI ethics instruction, while its 2026–2030 financial plan separately allocates approximately 1.75 billion won to standards for safe AI use and digital literacy. The corresponding performance schedule calls for 1,400 hours of education in 2026 and 8,000 hours over five years. In October, the research institute also announced visiting BeAT sessions, initially covering nine high schools in thirteen meetings, with activities that included criticizing generated answers and rebuilding drafts in students' own words. These programs establish that educational preparation forms part of the stated policy. Their schedules and targets do not, by themselves, tell us how far the preparation reaches, how consistently it enters the curriculum or which competencies students acquire from it.

The education office's teacher survey makes the distinction between technical familiarity and instructional capacity particularly concrete. Of 1,931 educators responding to an April survey intended to shape professional training, 96.6 percent reported experience using generative AI, but 58.8 percent said they used it in lessons, while 51.7 percent reported difficulty because they did not know how to apply it. These are figures from a training-needs survey, not a representative measure of all teachers' ethical competence. They nevertheless suggest that personal use was widespread among respondents while integrating AI into teaching remained a substantial challenge. The office planned staged training in prompt engineering and classroom application for approximately 1,900 educators and staff. Better prompts can make a tool more useful, but a teacher designing worthwhile AI-assisted instruction needs other judgments as well: when a generated explanation should be rejected, how assistance should be disclosed in assessed work, whether pupils have understood the source material and when using the tool is less educationally productive than working without it.

Those questions belong to the purposes of education rather than a narrow definition of online safety. A student may need to recognize a fabricated citation, refuse to upload another person's private material and understand what personal information an application collects. The same student also needs to know how to distinguish an argument developed through independent reading from an argument accepted because a fluent system produced it. In one case AI may lower the barrier to exploring an unfamiliar idea; in another it may allow a pupil to submit a polished response without acquiring the capacity to defend its reasoning. The teacher's challenge is to shape the conditions of use so that a machine's availability expands the pupil's intellectual agency rather than replacing the difficult parts of learning with an attractive performance of competence.

UNESCO's 2024 AI competency frameworks offer one way to examine the ambition, though they are not binding standards for Busan. The student framework addresses human-centered judgment, AI ethics, the understanding of AI techniques and the design of systems; the teacher framework also places pedagogy and professional development alongside technical knowledge. Such a framework cannot certify the quality of an individual lesson. Its usefulness here is that it distinguishes knowing how to operate an application from learning to evaluate its outputs, recognize its limits and make accountable decisions about its use. Busan's training plans contain elements of that broader aim, including critical review and revision. The public question is whether those skills are taught and evaluated as part of normal learning across schools, or remain concentrated in demonstrations, optional workshops and the accounts of particularly enthusiastic users.

The stakes of this distinction are heightened by the evidence the education office has offered for BeAT itself. In March, it said a pilot involving 27 classes during 2025 had verified the tutor's effectiveness, but the public announcement did not provide evaluation methods or measured learning outcomes that would allow an independent reader to examine the conclusion. At the July council hearing, Yu described the monitoring of use and consideration of interviews and other approaches to assessing classroom benefits during the broader rollout. Those accounts concern different stages of evaluation and need not contradict each other. They do, however, leave several different meanings of success in circulation: a school can approve a service, a student can enjoy using it, and an exercise can be completed more quickly without any one of those results demonstrating what the pupil can subsequently explain or solve. For an AI tutor, the last measure is central to the educational claim.

The Work and Records That Remain

The promises surrounding PENZ and BeAT meet in the information generated when students work. A common identity is useful partly because an unfinished project can eventually become a portfolio that follows a pupil into later schooling. But an AI tutor can also receive a photographed mistake, an exploratory question, a teacher's handout or a draft the student abandons. The different materials do not acquire the same educational status because the same account was used to produce or submit them. An assessed project may legitimately retain selected revisions that reveal the student's contribution, while a temporary question asked to overcome confusion may serve its purpose during the lesson. Requiring evidence of responsible AI use in schoolwork does not establish a reason to preserve every interaction as a permanent record of the pupil's thinking.

The research institute has not presented itself as operating outside privacy rules. Its published policy describes rights to request access, correction and deletion, subject to legal conditions, and sets out responsibilities for overseeing the processing of information by contractors. The institute has also described security measures for BeAT. Those commitments are relevant safeguards. They are not a substitute for distinguishing the administrative data needed to maintain a student's account from the files, conversations and feedback produced through instruction. Nor does identifying the GPT model or cloud environment establish that student conversations are retained for twelve years or that they are compiled into longitudinal behavioral profiles. The public evidence does not support either assertion. Responsible scrutiny concerns the rules actually applied to each category of work, the institutions empowered to retrieve it and the circumstances under which access or retention comes to an end.

Such distinctions become practical whenever a student's circumstances change. A pupil who moves to another class may still own a report but should not necessarily retain access to every restricted classroom document. A graduate may need a usable copy of a project even after eligibility for a publicly supplied software license has expired. BeAT's conversation-download function can allow useful exchanges to be deliberately preserved, but it does not settle the treatment of work that remains within the service. A durable portfolio therefore requires procedures for ownership, export and permission changes, not merely a continuing record of the user's identity. The education office has promised continuity over years of schooling; that promise can be assessed through ordinary transfers, grade progression and graduation long before a twelve-year cohort has completed its education.

What the Investment Should Demonstrate

The financial commitments make these distinctions consequential, though the size of a budget alone offers no proof of poor value. In July, Yu told the city council that BeAT had a 2026 budget of approximately 919 million won and that annual operating costs were expected to be around 800 million won. She explained that a low execution rate at that stage reflected the timing of contractor inspection and payment rather than the absence of an operating service. These are the director's budget figures and forecast, not audited final expenditure. The education office's wider five-year plan separately provides for broad distribution of generative AI services and for AI safety and digital-literacy education; those program lines cannot simply be added to BeAT's reported budget as though they were unrelated invoices. Coordination may reduce duplicated purchases, and broader access can provide public value even when the total cost does not fall. Both benefits still warrant evidence grounded in actual provision and use.

The office's published indicators are strongest at describing reach. PENZ registrations and visits show that the gateway has attracted users, while BeAT's initial school enrollment shows how rapidly administrative provision can expand. The planned hours of AI literacy education demonstrate that the authority has allocated attention and resources to a different part of the program. None of these figures, taken alone, can show that teachers have more time for instruction, that students can protect their work when access changes, that an AI-supported lesson advances independent reasoning or that the pupil recognizes an answer that should not be trusted. The appropriate measures need not be identical: a service designed to remove account-management burdens should demonstrate those savings, while a tutor intended to improve learning should be assessed against learning outcomes and a literacy program against students' demonstrated judgment. Publication of favorable experiences would be informative, but their significance depends on who participated, what was measured and how the results compare with students who were not using the service in the same way.

Busan began by addressing a genuine access problem, and its answer may make educational software more equitable and classrooms less encumbered by administration. The education office has also acknowledged the need for ethics, teacher preparation and forms of AI assistance intended to stimulate inquiry rather than deliver finished answers. The strongest case for the policy lies in bringing those elements together. An account that lasts twelve years can carry the student's selected work from one stage of education to another; the courses delivered through that account must also cultivate abilities the student can carry without the software. That demands a more exacting standard than the survival of a username or a rise in the number of logins. Public education succeeds when the tools it provides leave students with work they can legitimately keep, explanations they can test and judgments they are increasingly able to make for themselves.