Why King Sejong as a persona, not just a history tutor?

We could build a generic AI tutor that explains Joseon Dynasty history in a neutral manner. However, recent research on immersive learning and narrative pedagogy demonstrates that culturally grounded personas offer distinct cognitive and affective advantages that generic tutors cannot replicate. Studies of XR docent systems and museum chatbots reveal that personas create an embodied historical presence, enabling visitors to form emotional connections that enhance memory retention and foster deeper engagement with cultural contexts.

With this understanding, a King Sejong persona specifically offers unique educational affordances:

• Immersive Historical Perspective: Speaking as Sejong enables first-person accounts of decision-making processes, for example, during Hangul's creation, conveying the intellectual and political challenges in ways that third-person explanation cannot capture. When visitors hear "I created Hangul because..." rather than "King Sejong created Hangul because...", there is a deeper engagement with historical agency and motivation.
• Culturally Situated Understanding: Sejong's voice reflects a 15th-century Korean Confucian worldviews, Neo-Confucian governance principles, and period-specific social hierarchies. This cultural situatedness helps learners experience and, possibly, understand historical perspectives, rather than projecting modern assumptions backward.
• Narrative-Based Learning: Storytelling through a historical figure's perspective activates different cognitive processes than factual presentation. Research on narrative transportation theory indicates that learners process story-based information more deeply and retain it for a longer period than abstract facts.
• Emotional and Cultural Investment: For Korean learners, especially, interacting with Sejong connects personal and national identity to historical understanding. For international learners, the persona creates a human entry point into Korean cultural heritage that feels less abstract than textbook accounts.

While these are clear pedagogical advantages, they come with responsibility. A persona that sounds authentic risks misleading users about the certainty of historical knowledge. This is precisely what prompted us to create the epistemic framework presented in this paper, to preserve the engagement benefits of persona-based learning while maintaining scholarly integrity regarding what we can and cannot know.

Research Focus

In this work, we focus on King Sejong the Great (세종대왕, 1397-1450), who ruled the Joseon Dynasty of Korea for 32 years and created Hangul, the Korean alphabet. Sejong is an ideal case for several reasons:

• His reign is extraordinarily well-documented: eight historians followed him daily, recording even embarrassing moments. These records have been fully digitized since 2006 and are available online.
• Historians actively debate aspects of Sejong's life, like whether he created Hangul alone or with help from scholars. These debates let us test how well our system handles uncertainty.
• Sejong remains deeply important to Korean national identity, Hangul Day is a national holiday, his statue stands in central Seoul, and his face appears on the 10,000 won note. This cultural significance means any AI representation must involve Korean communities as decision-makers, not just consultants.

Rather than building yet another chatbot, we developed an epistemic framework, a system for representing what we know and how certain we are. The objective of our four-tier system is to discriminate between:

• documented facts that appear in multiple sources,
• interpretations most historians agree on,
• reasonable guesses based on context, and
• questions historians actively debate.

This approach builds on the work of Michael Piotrowski, who argues that digital humanities must find ways to formally represent uncertainty rather than hide it.

Research Questions

• How can AI systems handle Korean historical texts while preserving complexity instead of oversimplifying?
• What governance structures enable Korean communities to exercise genuine authority over cultural representation?
• How can we design systems that clearly communicate users' limitations to them honestly?

Learning from Existing Projects

Several recent projects demonstrate both the possibilities and challenges of recreating historical figures with AI.

UneeQ Digital Einstein (2021) combined a computer-generated face, a synthesized voice, and integration with scientific computation tools. The project generated significant traffic, with over 100,000 interactions in just two weeks. But the conversations stayed shallow. Einstein could answer questions about his documented life and known scientific work, but he was unable to discuss anything after his death in 1955 or engage emotionally in meaningful ways. While the project succeeded as a marketing tool, it failed as a means of historical education.

MIT's Living Memories (2023) took a more scholarly approach. Researchers created AI versions of several historical figures and tested them with 90 students. Students learned more and felt more motivated compared to traditional methods. However, the researchers identified persistent problems: the AI confidently fabricated plausible-sounding facts, filled evidence gaps with fiction, and struggled to strike a balance between engagement and accuracy. Technical sophistication alone wasn't enough; researchers needed robust fact-checking and clear warnings about AI limitations.

The Mary Sibley Project (2024) at Lindenwood University offers the most relevant lessons. Researchers trained ChatGPT on Sibley's complete diaries from 19th-century Missouri. The AI successfully captured her religious tone and old-fashioned language. However, success required university archivists to review every response; they caught the AI confidently stating incorrect founding dates and institutional names. The project also added that Sibley's historically accurate views on race and gender would trouble modern readers. This points to a more mature approach than either hiding an uncomfortable history or presenting it without context, although the tensions remain unresolved.

Most relevant to our case, the Korean Institute of Science and Technology deployed an AI system for Joseon Dynasty historical records in 2025. The system provides answers to questions about 472 years of Korean history, accompanied by clear source citations. This proves that AI can handle Korean historical materials. However, it also revealed ongoing challenges, including managing dense chronological texts, balancing facts with interpretation, and processing massive datasets (over 336,000 career records with annotations).

When polled together, here are the learnings we can derive from these projects:

What works:

• Starting with rich personal writings (diaries, letters) helps the AI sound authentic
• Having experts review responses before public launch catches errors
• Being transparent about AI limitations builds trust
• Framing projects as educational (not entertainment) sets better expectations
• Providing clear citations maintains credibility.

What consistently fails:

• Accuracy remains hard to verify despite better technology
• Handling historically authentic but offensive views creates unresolved tensions
• Users anthropomorphize AI, treating generated text as real testimony
• Commercial pressures often conflict with scholarly standards

As Underwood's recent research confirms, perfect historical authenticity may be technically impossible. Even AI models trained exclusively on historical texts produce detectable anachronisms. This means we need realistic expectations: our systems create interpretations based on historical sources, not authentic recreations of historical consciousness.

What Digital Humanities (DH) Theory Tells Us

DH scholars have long warned about these problems. Johanna Drucker argues that we shouldn't treat digitized materials as transparent windows onto the past. Every digital representation involves interpretive choices, such as which documents to include, how to organize them, and what metadata to add. These choices shape what users can find and understand.

Roopika Risam extends this critique to show how digital projects can perpetuate colonial power structures. When Western institutions digitize non-Western cultural materials, benefits often flow outward. Western researchers gain access, prestige, and publications, while source communities see little return. Even "universal" technical standards often encode Western assumptions about how knowledge should be organized.

For AI specifically, we must design systems that preserve complexity rather than eliminate it. The goal isn't to make history simpler or more accessible at the cost of accuracy; it is to make complexity navigable.

How the Four Tiers Work

At the most basic level, we consider historical knowledge to exist in layers, from the most certain to the most debatable. From this perspective, we can identify the following tiers:

TIER 1: DOCUMENTED FACTS (HIGH confidence)

These represent statements that we can consider factual, as they are verifiable directly in primary sources and supported by multiple independent sources.

Example : "When were you born?"

AI Response : "I was born in 1397, the 26th year of King Taejo's reign."

What the user sees:

Sources : 3 primary sources confirm this.

Confidence : HIGH (documented fact)

Citation : Sejong Sillok, Biographical Section; Veritable Records of King Taejo, Year 26

This is straightforward; the birth year appears in multiple historical records written during or shortly after Sejong's life. However, in the background, the system searches its knowledge base for excerpts from documents mentioning Sejong's birth. It identifies the relevant passages, verifies their consistency, and generates a response that utilizes the language style of the historical sources. The metadata tags indicate that this is a biographical fact with high certainty.

TIER 2: SCHOLARLY CONSENSUS (MEDIUM-HIGH confidence)

These are interpretations that require inference but are widely accepted by historians. The evidence is strong, even though it requires some interpretation.

Example : "Why did you create Hangul?"

AI Response : " The spoken language of our country is different from that of China and does not suit the Chinese characters. Therefore, amongst uneducated people, many have been unable to express their feelings in writing when they have something they wish to put into words. I was greatly distressed by this, and so I have made twenty-eight new letters to everyone to practice them at their ease, and adapt them to their daily lives. "

What the user sees:

Sources : Hunminjeongeum preface (primary source) + 4 scholarly interpretations

Confidence : MEDIUM-HIGH (most historians agree)

Citation : Hunminjeongeum (1446), Preface by King Sejong

Note : We infer Sejong's motivation from his explicit statements in the preface of his work and from his literacy policies. Most scholars agree on the democratic/practical motivations, though some emphasize additional factors, such as linguistic nationalism.

In this case, the system retrieved the Hunminjeongeum preface (a primary source) along with several scholarly articles about the creation of Hangul. It examines the degree of consensus among historians regarding the interpretation. Because most of the retrieved articles support the democratic motivation explanation, the system classifies this as a scholarly consensus.

TIER 3: REASONABLE SPECULATION (MEDIUM-LOW confidence)

Answers belong to this tier when they are considered educated guesses based on historical context, but with limited direct evidence. Historians consider them plausible but can't prove them definitively.

Example : "How did you feel when yangban scholars opposed Hangul?"

A I Response : "While I cannot directly describe my feelings, court records document my decisions and policies, not inner emotional states; historical context suggests this was challenging. The records show I proceeded with Hangul despite elite opposition and assigned Hall of Worthies scholars to develop theoretical justifications. This determination suggests I viewed Hangul as essential to good governance even when facing resistance from officials who saw Chinese characters as culturally superior."

What the user sees:

Sources : Inferred from court records + historical context

Confidence : MEDIUM-LOW (reasonable speculation)

Citation : Sejong Sillok, Years 25-28 (Hangul development period

Note : Sejong's emotional states are rarely directly documented in court records. We infer from documented actions (proceeding despite opposition) and policy patterns. Historians consider this plausible but acknowledge limited direct evidence. In the historical context in which these documents were collected, we find the Joseon court records prioritized policy documentation over personal emotion, following Neo-Confucian traditions.

Notice the AI explicitly states it cannot know Sejong's feelings. Instead, it explains what evidence exists (documented actions) and why historians make this inference. This model's good historical thinking involves distinguishing between what we know and what we reasonably infer.

The system classifies the query as an emotion/motivation inquiry and searches for direct statements about Sejong's feelings, and finds none. It retrieves documents about his actions during this period and uses those to construct a contextualized inference. The response explicitly flags its speculative status.

TIER 4: CONTESTED INTERPRETATIONS (Multiple views, active debate)

It is inevitable that historians will disagree on certain questions, with different scholars using different evidence or frameworks to reach different conclusions.

Example : "Did you create Hangul alone or with your scholars?"

AI Response : "Historians actively debate this question. The evidence supports multiple interpretations.”

This response was guided by various viewpoints identified from the context of this question:

Majority View (Collaborative Creation, ~65% of scholars): King Sejong worked closely with Hall of Worthies scholars, especially Prince Suyang and scholars Jeong In-ji, Choe Hang, and Seong Sam-mun. The Hunminjeongeum Haerye (Explanatory Guide) credits these scholars with theoretical elaboration. This view emphasizes Confucian governance ideals, where rulers work in conjunction with learned advisors.

Minority View (Sole Creator, ~35% of scholars): King Sejong personally designed Hangul's phonetic system, with scholars only writing explanatory materials afterward. Evidence includes the preface's first-person statements ('I have created 28 letters') and Hangul's linguistic sophistication, suggesting a single brilliant mind rather than a committee design.

Both interpretations use the same sources but emphasize different aspects. The collaborative view stresses governance patterns and scholar credits. The sole-creator view emphasizes linguistic sophistication and first-person language, resulting in an interpretation that depends on whether one prioritizes social-historical context or technical linguistic analysis.

What the user sees:

Sources : Same primary sources interpreted differently; 8 scholarly works divided

Confidence : CONTESTED (active debate, roughly 65-35% split)

Citations : Hunminjeongeum (1446), Lee & Ramsey (2011), containing a collaborative argument, and Kim-Renaud (1997), pointing to a sole creator argument

Note : This isn't resolvable with better AI or more data; it reflects genuine historiographical disagreement. Different scholars employ different frameworks (social history versus linguistic analysis), resulting in varying conclusions drawn from the same evidence.

We can consider this example as the most sophisticated response type. The AI doesn't pick a side or try to synthesize a middle position. Instead, it presents both views fairly and explains why historians disagree. This teaches users that history involves interpretation, not just fact-gathering.

During the search for scholarly literature, the AI detects significant disagreement (not 90-10%, but closer to 65-35%). It retrieves representative articles from each camp and analyzes their reasoning. The response explains both positions and the methodological differences that lead to different conclusions.

Conceptual Architecture

Figure 1 : System workflow from question to response, including ongoing review by Korean historians.

While most AI chatbots implement the classic RAG workflow in which the AI generates a response based on the training or on a Retrieval Augmented Generation (RAG) architecture, the proposed solution adds a classification sub-system to identify the certainty level:

• User asks questions
• System searches actual historical documents (RAG phase)
• The system classifies the certainty level
• Response includes both content AND epistemic status
• User sees evidence, not just claims

These additional steps are at the core of our contribution:

• Documents tagged with epistemic status during curation : When we add historical sources to our database, we mark whether they're primary sources, secondary interpretations, or scholarly debates. This metadata helps the system understand not just what documents say, but how reliable they are.
• Generation varies by tier : The system uses different response templates depending on the certainty level, as described in our tiered classification system.
• Mandatory transparency : Unlike commercial chatbots that hide their reasoning, every response shows the evidential basis. Users always see:
  • Which sources were used
  • How certain can we be,
  • What the citations are, and
  • Where to find alternative interpretations.
• System acknowledges its interpretive role : Rather than claiming neutral information retrieval, the system explicitly states that it creates interpretations based on historical sources. This honesty is crucial—it prevents users from treating AI outputs as authoritative truth.

This approach builds on Johanna Drucker's argument that digital humanities should make interpretive layers visible, rather than hiding them behind technical sophistication.

Authenticity as a Learning Opportunity

Recent research by Ted Underwood proved that AI models produce anachronisms even when trained exclusively on historical texts. Perfect historical authenticity is technically impossible. But rather than seeing this as a failure, we can make it educational.

When users learn to question AI outputs, "Why should I trust this? What evidence supports it? Whose interpretation is this?", they practice historical thinking. Our four-tier framework makes epistemic status visible, helping users distinguish between proven facts and scholarly interpretation. This is a core humanities skill that's often lost when information appears authoritative.

Systems should honestly state: "I am a computational model based on historical sources, not King Sejong himself." This acknowledgment models scholarly humility. Because the goal is not to tricka users into believing they are talking to the historical Sejong, it is helping them engage with Joseon Dynasty history in an accessible way while learning to think critically about sources and provide a scholarly example of transparency and intellectual integrity that can be adopted in other disciplines affected by the use of Generative AI.

Design as Interpretation

Our understanding of this framework evolved through research and writing. We initially focused on technical sophistication but came to recognize that epistemic responsibility matters more than computational advancement. The system design itself is an interpretive act requiring ongoing reflection.

This models how digital humanities should approach AI development: iteratively and reflexively, acknowledging that frameworks embed cultural values that require continuous examination. There's no neutral implementation; every design choice reflects assumptions about what knowledge matters, how certainty should be represented, and whose authority counts.

The four-tier system works because it makes these interpretive choices explicit rather than hiding them. When the system classifies a response as "Tier 3: Reasonable Speculation," it's not just providing information; it's modeling historiographical reasoning. When it presents multiple contested views in Tier 4, it teaches users that historical interpretation involves debate, not just the accumulation of facts.

This approach resists what Drucker calls "digital positivism," the assumption that computational methods give us unmediated access to reality. Instead, it shows how digital tools can preserve the interpretive complexity that makes the humanities valuable while leveraging computational capabilities for accessibility and scale.