Busan’s Public-Sector AI and the Problem of Defining “Success”

Busan, South Korea — Public institutions are increasingly promoting basic uses of artificial intelligence as evidence of innovation, even when the underlying tasks would not be considered performance achievements in most private organisations. Busan Metropolitan City’s recently published “AI Practical Use and 2025 Learning Achievements Casebook” exemplifies this pattern, presenting routine administrative automation as a milestone in public-sector reform.

The casebook, distributed across city departments and affiliated agencies, highlights examples such as automated document summarisation, keyword-based file classification, chatbot-style responses to civil complaints, and AI-assisted drafting of standard administrative texts. City officials have described the publication as a practical manual for civil servants and as a summary of results from a year-long programme of AI training and internal learning groups.

Yet the functions showcased in the casebook are neither novel nor technically demanding. Comparable capabilities have been embedded for years in widely used office software and cloud-based productivity tools, and are routinely employed by small and medium-sized enterprises as part of everyday operations. In those contexts, such tools are treated as baseline infrastructure, not as indicators of organisational transformation.

The issue raised by Busan’s publication is therefore not whether public offices should adopt artificial intelligence, but how administrative success is being defined. When the delayed uptake of commonplace technology is framed as innovation or leadership, it signals a gap between contemporary standards of organisational performance and the criteria applied in public-sector AI policy.

That gap matters because artificial intelligence in government carries responsibilities that extend beyond efficiency. Administrative systems issue authoritative information, shape citizens’ rights and obligations, and operate under heightened expectations of accountability. Against that backdrop, the elevation of entry-level automation to the status of policy achievement invites closer scrutiny of what governments choose to celebrate—and what they choose not to examine.

Routine Automation Miscast as Administrative Achievement

A closer examination of the initiatives highlighted reveals that most fall into the category of routine administrative automation rather than institutional innovation. The tasks emphasised—summarising documents, classifying files by keywords, generating standardised text, and providing scripted responses through chatbots—require no fundamental redesign of administrative processes. They automate existing work without altering how decisions are made, reviewed, or authorised.

In organisational terms, such functions are typically regarded as enabling tools. In the private sector, including among small and medium-sized enterprises with limited technical capacity, these capabilities are deployed as part of basic operational infrastructure. They reduce repetitive workload but do not, on their own, qualify as performance outcomes. Their introduction is rarely documented as an achievement because it does not change how the organisation is governed or how responsibility is assigned.

By contrast, the framing adopted here treats the introduction of these functions as evidence of progress in AI-driven administration. This approach effectively lowers the bar for what is recognised as reform. The concern is not that public offices are adopting widely available tools, but that adoption itself is presented as a result, rather than as a preliminary step toward more substantive change.

Public-sector innovation is generally assessed by its impact on institutional behaviour. That includes whether new systems modify approval chains, introduce enforceable standards, redistribute authority, or establish clearer lines of accountability. None of these elements are apparent in the examples described. Instead, the use of AI remains confined to assisting individual tasks within unchanged administrative structures.

The material also makes little distinction between outputs generated through training exercises and systems that are operationally embedded. Applications developed in internal learning groups are presented alongside routine practices, without clarification as to whether they are deployed at scale, supported by dedicated budgets, or subject to formal oversight. In administrative terms, this distinction is critical. Training outputs are temporary and exploratory by nature, while operational systems imply continuity, responsibility, and public accountability.

By collapsing these categories, the publication creates an impression of advancement that rests on activity rather than institutional capacity. What is presented as achievement reflects participation in a technological trend, not the completion of a reform process. The result is a narrative of progress that prioritises visibility over verifiable change, and demonstration over durability.

From Routine Automation to Administrative Risk

The significance of these initiatives lies less in their technical ambition than in where they are applied. Several of the highlighted uses—automated responses to civil complaints, AI-assisted drafting of administrative guidance, and chatbot-based information services—operate in domains where government communication carries formal authority. In such settings, even minor errors can escalate into administrative disputes, appeals, or loss of public trust.

Research on public-sector AI consistently identifies citizen-facing automation as high-risk, regardless of technical complexity. Studies by the OECD and national audit bodies have shown that systems producing automated guidance or responses externalise error: mistakes are no longer contained within internal workflows but are transmitted directly to the public. When scaled across large volumes of interactions, low error rates can translate into systemic misinformation.

Behavioural research further complicates the picture. Empirical studies in public administration and human–computer interaction document the prevalence of automation bias, a tendency for officials to place undue confidence in machine-generated outputs. Under conditions of time pressure and workload constraints, AI-generated text is more likely to be accepted with limited scrutiny, particularly when systems are framed as efficiency tools. In administrative contexts, this increases the likelihood that flawed outputs are issued as official information.

Data protection and information security introduce additional exposure. Automated complaint handling and speech-to-text systems routinely process personal data, including sensitive information. Public-sector cybersecurity agencies have warned that AI systems integrating external models or tool-based frameworks expand attack surfaces, increasing vulnerability to data leakage, prompt manipulation, and unintended data retention. These risks exist independently of whether the AI function itself is technically sophisticated.

What heightens concern is the absence of visible mitigation mechanisms. Public descriptions emphasise functional capability but provide little insight into how errors are detected, how outputs are audited, or how accountability is assigned when failures occur. In administrative terms, capability without control constitutes a risk, not an achievement. Effective governance requires predefined thresholds for intervention, clear ownership of outcomes, and procedures for suspension or correction—elements that remain largely unarticulated.

The framing of such systems as successes further amplifies exposure. Once automation in high-responsibility areas is publicly validated, institutional incentives shift toward expansion rather than caution. Scrutiny diminishes, while replication becomes easier to justify. Under these conditions, systems that warrant careful monitoring can instead be normalised before their implications are fully understood.

The resulting risk is structural rather than technical. It arises not from the use of basic AI tools, but from their deployment in authoritative administrative roles without commensurate safeguards. In government, where information is treated as instruction and guidance as obligation, the margin for error is narrow. Treating routine automation in these domains as progress, without parallel investment in oversight, places institutional credibility at stake.

When Declared Success Blocks Verification and Redesign

Once an AI initiative is formally presented as a policy success, its status within the administrative system shifts. Provisional tools and limited experiments are recast as established outcomes through official publications, performance reports, and public communication. Subsequent evaluation is no longer treated as routine oversight but as a challenge to decisions already endorsed by leadership.

In public administration, success functions as an institutional commitment rather than a descriptive label. Recorded achievements become reference points in budget deliberations, performance assessments, and interdepartmental coordination. Revisiting an initiative that has been publicly validated requires reopening judgments that have already been formalised, a process that carries political and organisational cost and is therefore frequently avoided.

Early success designation also reshapes evaluation practice. Technology initiatives in government are often assessed at the point of introduction. When a system is declared successful at that stage, structured monitoring tends to weaken. Performance audits, error-rate tracking, and user-impact assessments are no longer mandatory components of the project lifecycle. Evaluation ends precisely when continued scrutiny becomes most necessary.

Responsibility distribution further discourages correction. Recognition for success is institutionalised and shared across leadership and departments, while operational risk is concentrated at the implementation level. Staff responsible for day-to-day operation manage the consequences of system errors without commensurate authority to suspend, modify, or redesign. Proposals to pause or revise a publicly endorsed system can therefore expose individuals to professional risk rather than institutional support.

Budgetary incentives reinforce this rigidity. Initiatives cited as successful become anchors for future funding and expansion. Replication across departments or service areas is easier to justify than reconsideration. At this stage, redesign is framed not as technical improvement but as a threat to budget continuity and policy credibility. The cost of reassessment increases, while incentives to proceed unchanged intensify.

Language surrounding artificial intelligence compounds these effects. When AI initiatives are framed as innovation, transformation, or future readiness, symbolic value accumulates around deployment itself. Technical criticism is more easily dismissed as resistance to progress, and administrative caution as reluctance to modernise. Debate shifts away from empirical performance toward narrative alignment.

The combined result is a familiar paradox in technology-led public policy. Systems that remain immature or lightly governed become harder to change precisely because they have been declared successful too early. Redesign, which should indicate learning and institutional maturity, is instead treated as an admission of failure. Verification, which should intensify as systems scale, is sidelined.

In this environment, the principal risk does not stem from experimentation, but from premature validation. When success is defined loosely, institutions become locked into trajectories that prioritise continuity over correction and presentation over performance. For public-sector AI, where authority and trust are central, such inversion carries consequences that extend beyond technology itself.

What Success Should Mean in Public-Sector AI

In public-sector artificial intelligence, the central question is no longer whether government can adopt new tools, but whether it can govern them. As AI systems become cheaper, more accessible, and easier to deploy, technological adoption has lost its value as a measure of progress. What distinguishes meaningful advancement from superficial change is the capacity to manage risk, assign responsibility, and sustain oversight over time.

The pattern examined here illustrates a recurring misalignment. Basic automation is treated as reform, experimentation as achievement, and visibility as validation. This inversion matters because it shifts institutional attention away from outcomes and toward appearances. When success is defined by deployment rather than performance, the administrative system rewards early declaration over continued scrutiny.

Public-sector AI operates under conditions that make such miscalibration especially costly. Government outputs are not suggestions or conveniences; they are authoritative signals that shape rights, obligations, and expectations. In this context, the absence of clear ownership, audit mechanisms, and redesign pathways is not a technical gap but a governance failure. Systems that cannot be paused, corrected, or withdrawn safely are not mature, regardless of how widely they are used.

A more rigorous understanding of success would therefore focus less on what AI systems can do and more on how institutions respond when those systems fall short. The ability to detect error, to intervene without reputational penalty, and to redesign without political resistance is a stronger indicator of administrative capacity than any catalogue of applications. Under such a standard, restraint becomes a form of competence, and revision a sign of institutional strength.

The broader implication extends beyond a single city or case. As public institutions rush to demonstrate relevance in an AI-driven environment, the temptation to lower performance thresholds will persist. Resisting that impulse requires redefining success not as the accumulation of tools, but as the development of durable governance structures capable of absorbing uncertainty.

Ultimately, progress in public-sector AI should be measured by the degree to which institutions are prepared to stand behind the consequences of automation. Where success is declared too easily, accountability weakens. Where verification and redesign are built into the definition of achievement, trust becomes sustainable. The difference lies not in technology, but in how governments choose to judge themselves.