RESILIENCE SUPERSEMINAR · REGGIO EMILIA, MARCH 9-11, 2026
«EXPLORING THE FRONTIER OF SCIENCE WITH AI AND THE ROLE OF HUMANITIES»

From Text to Sound: AI-Assisted Hermeneutic Tools for the Study of Religious Traditions

Research questions, visual complexity, sonic immersion, and cross-disciplinary transfer

Francesco Mariano
Multimedia Artist & Lecturer in Sound Design and Interaction Design
Academy of Fine Arts, Macerata · G.B. Martini Conservatory, Bologna

Abstract

This paper presents two progressive web applications (PWAs) developed as research tools for the study of religious traditions: SacredTexts AI, which uses artificial intelligence to analyze sacred texts and visualize them as interactive graphs with generative soundscapes, and SacredSounds Radio Hub, which aggregates 73 religious radio stations from 12 traditions into a navigable interface with AI-assisted comparative analysis. The paper recounts the development journey of both projects, from the perspective of an outsider to religious studies, highlighting discoveries that emerged at the boundaries between technology, epistemology, and the study of religions: AI safety filters as a form of implicit theology, the graph as a new reading surface, sonic curation as an epistemological act, real-time audio as a hermeneutic dimension. The contribution addresses the two central questions of the seminar: what more complex research questions can we formulate through new technologies, and how can this complexity be channeled into tools useful to other domains.

Keywords: artificial intelligence, sacred texts, computational hermeneutics, graph visualization, religious radio, comparative study of religions

I am not a theologian, a biblical scholar, a historian of religions, or a philologist. I am a musician, a composer of electroacoustic works, and a lecturer in creative technologies (Sound Design, Interaction Design, audiovisual programming) at Italian conservatories and academies of fine arts.

What I present here is the account of a journey born from the professional curiosity of an interaction designer who chose sacred texts as a testing ground for comparing different generalist artificial intelligence systems.

The application grew through a series of questions I asked myself along the way: questions that, I now realize, directly touch upon the two central problems this seminar addresses.

II. SacredTexts AI: Text as Computational Object

Starting points: the questions that generated the tool

The project began as an exercise in comparing AI systems. As an interaction design lecturer, I was interested in how different large language models (GPT-4, Claude, Gemini) would respond to the same structured prompt when asked to perform a complex analytical task. I needed a domain challenging enough to reveal differences between models: one that required sensitivity to historical context, awareness of multiple linguistic and cultural layers, the ability to maintain structured output, and the management of material that pushes against the safety filters these systems impose.

Sacred texts proved to be an almost perfect stress test. A passage from Genesis contains Hebrew poetry, cosmological assertions, centuries of rabbinic interpretation, theological implications across three Abrahamic traditions, and an internal structure that resists simplification. Asking an AI to "analyze" such a text means asking it to navigate a space of extraordinary density, and the way different models navigate that space reveals their architectures, their training biases, and their implicit epistemologies.

The initial questions were therefore technical and comparative:

• Given the same structured prompt, how do different LLMs decompose a sacred text into thematic clusters?
• What systematic differences emerge in how they handle cross-traditional references?
• Where do they hallucinate, and where do they refuse to proceed?

But while building the tool to answer these questions, new ones appeared: questions I could not have formulated without the tool itself. This, I believe, is the fundamental dynamic this seminar intends to examine.

The architecture of the tool

SacredTexts AI is a progressive web application (PWA) that uses Google's Gemini language models (user-selectable among four variants, from Flash-Lite to Pro) to analyze passages of sacred texts belonging to eight religious traditions. The analysis is visualized as an interactive graph with cinematic mode, where concepts, themes, and cross-references become visible animated spatial relationships. It includes a dual audio system: real audio samples specific to five religious traditions (Gregorian chant, Hebrew cantillation, Sufi ney flute, Om with tanpura, Tibetan singing bowl) played via native Web Audio API, and a probabilistic algorithmic generator where dragging graph nodes modulates musical parameters in real time.

A pool system with seven API keys and round-robin rotation ensures stability during high-usage sessions. When a user submits a sacred text, Gemini decomposes it into a structured JSON object: a central node, thematic clusters, child concepts, keywords, cross-references, historical context across nine dimensions, confidence scores, and three epistemological research questions. The prompt governing this analysis is itself a carefully crafted artifact: 93 lines of instructions defining the AI's role, its constraints, the output schema, and the epistemological posture it must adopt.

OpenAI's GPT-4 was used during development for comparative testing and prompt refinement, running the same texts through different models to evaluate fidelity, hallucination rates, and adherence to the output schema. This triangulation is itself a methodological point.

The prompt as hermeneutic tool

I want to dwell on the prompt, because I believe it is one of the most overlooked elements in the current discourse on AI and the humanities.

The system prompt I designed for the text analysis engine is not a simple instruction. It is a hermeneutic framework encoded in natural language. It assigns the AI a role ("comparative religion scholar, expert historian, and information architect"), establishes epistemological rules ("never invent concepts not present in the text," "provide in-depth historical context," "assign confidence scores to every assertion"), defines the output ontology (a graph structure with typed relationships), and sets constraints on granularity ("quotations must be 10 words or fewer").

Designing this prompt required me to make explicit decisions about what constitutes a valid analysis of a sacred text. This is an inherently epistemological act. What categories should the AI use to classify themes? I chose four: theology, ethics, mysticism, practice. This is a choice with consequences: it privileges certain readings over others, and any alternative categorization would produce a different graph topology.

The prompt, in other words, is not neutral infrastructure. It is an executable theory of reading. And every modification produces measurably different outputs. This means that prompt design for humanistic applications is not an engineering task: it is an academic one, requiring the same rigor we apply to the design of a research methodology.

III. Three Discoveries from Text

Working on SacredTexts AI as a person without training in religious studies produced a series of encounters I would not have had from within the discipline. I describe three, because each generates a research question of the type this seminar intends to identify.

1. Safety filters as implicit theology

In the early stages of development, I discovered that Google's AI refused to analyze certain passages from the Quran, the Hebrew Bible, and the Vedas. The API returned a SAFETY flag and blocked the response. Passages describing divine punishments, wars, or severe judgments triggered the model's harm prevention filters, specifically the categories "hate speech" and "harassment."

At one level, this is a technical problem with a technical solution (the application includes an option to relax these filters for academic study). But at another level, it reveals something profound: the AI system carries with it an implicit judgment about which parts of sacred texts are "safe" and which are "dangerous." This judgment was not formulated by theologians: it was formulated by engineers in California applying general-purpose content moderation rules to texts that have been the subject of commentary and interpretation for millennia.

When an AI classifies a passage as "hate speech," it is performing an act of interpretation with theological, cultural, and political implications, yet it is an interpretation that passes as a neutral technical constraint.

This discovery is not merely a problem for religious studies. It is a case study in the broader challenge of applying universal content moderation to culturally specific material. The same problem arises in legal analysis (judgments describing violent crimes), medical texts (descriptions of self-harm for clinical purposes), historical research (primary sources containing offensive language), and conflict studies (analysis of extremist rhetoric).

Sacred texts, because they combine extreme cultural sensitivity with extreme interpretive complexity, are probably the hardest case. A solution that works for computational analysis of the Quran will work for legal proceedings, medical reports, and historical archives. In this sense, religious studies are not consumers of AI safety research: they are its drivers, precisely because their materials are the most demanding.

2. The graph as a new reading surface

When the AI's analysis is rendered as an interactive graph (with nodes representing concepts, edges representing relationships, and spatial proximity indicating thematic affinity), something unexpected happens. The visual topology of the graph itself becomes an object of interpretation.

A densely connected cluster suggests a passage with high internal coherence. A conceptual node positioned between two clusters, connected to both, suggests a bridging theme that the linear text does not make explicit. Cross-references between traditions, shown as colored lines traversing the graph, make visible a structure of relationships that exists in the AI's analysis but would be difficult to perceive in a written comparative essay.

The graph is not a summary of the text. It is a spatial transposition of a particular reading of the text: the AI's reading, governed by the hermeneutic framework of the prompt. But because it is spatial, it is also explorable: the user can zoom, pan, click nodes to see details, drag nodes to reorganize relationships. This interactivity transforms the reader from a consumer of analysis into an active explorer of an analytical space.

3. Research questions as recursive inquiry

Perhaps the most significant feature of the application, and the one most directly relevant to this seminar's concerns, is that for every textual analysis, the AI is instructed to generate exactly three epistemological research questions. These are not summaries or comprehension questions. The prompt explicitly requires the identification of tensions, paradoxes, historiographic gaps, and non-obvious connections within the analyzed passage, formulated as open academic questions useful for further research.

What makes this mechanism powerful is its recursive nature. Every generated question can be fed back into the analysis engine as new input. The user clicks "Analyze this question," and the AI produces a new graph, with its own structure, its own cross-references, and its own three new research questions. The process can be iterated indefinitely, creating a branching tree of inquiry that deepens with each cycle.

Consider a concrete example. Analyzing Genesis 1:1-5, the AI might generate: "What is the relationship between the pre-existing tohu va-vohu and the creative act of bara: does the text imply a creation ex nihilo or an ordering of pre-existing chaos, and what are the historiographic implications of each reading?" This question, when reanalyzed, produces a graph focused on ancient Near Eastern cosmogonies, the parallel with the Enuma Elish, and the theological stakes of the creatio ex nihilo debate, generating three further questions that a scholar of Second Temple Judaism might spend years investigating.

AI enables not only more complex individual questions, but recursive chains of inquiry where every answer generates new questions at increasing levels of specificity: a form of systematic deep exploration of the interpretive space with no manual equivalent.

IV. SacredSounds Radio Hub: From Text to Sound

Where SacredSounds comes from

SacredSounds Radio Hub was born from a question that presented itself while I was developing SacredTexts AI. Working with texts (deconstructing them into conceptual graphs, sonifying them through algorithmic generators), I became aware of an absence. Sacred texts, in the application, were analytical objects: textual fragments filtered through the computational mediation of AI, made visible as nodes and edges in a two-dimensional space. But religious traditions do not live only in texts. They live in voices, songs, recitations, liturgies transmitted in real time by communities scattered across the world.

In SacredTexts AI, sound was generative: algorithmically produced from graph data. In SacredSounds, sound is real: live audio streams from over 70 religious radio stations across 12 traditions, broadcast by communities reciting, singing, and praying right now.

If SacredTexts AI asks the question "how does a machine read a sacred text?", SacredSounds poses a complementary one: how can one build a listening space that fosters interreligious understanding without reducing the complexity of traditions to a catalog?

Curation as epistemological act

In SacredTexts AI I described the prompt as a hermeneutic framework encoded in natural language, an executable theory of reading. SacredSounds presents an analogous epistemological problem, but shifted from the domain of text to that of sound: the curation of sonic sources.

Selecting 73 radio stations from 12 religious traditions is not a neutral act. Every choice of inclusion and exclusion is an interpretive decision. How many stations for Christianity? How many for Buddhism? The disproportion in the number of stations available online reflects asymmetries of power, technological access, and media presence that have nothing to do with the theological relevance or spiritual depth of a tradition.

I tried to manage this disproportion with explicit criteria:

• Institutional authenticity: prioritizing stations run by recognized religious communities (Radio Vaticana, Chabad.org, SikhNet, Darbar Sahib) rather than commercial aggregators.
• Internal diversity: representing multiple currents within each tradition. Christianity includes Radio Vaticana (Catholic), Ancient Faith (Orthodox), BBN (Protestant Evangelical). Islam includes different Quran reciters (Alafasy, Abdulbasit, AlSudais), each with an approach to tajwīd that carries a specific interpretive tradition.
• Original language: stations in Arabic for Quranic recitation, in Hebrew for Torah study, in Punjabi for Sikh Gurbani, in Pali for Buddhist sutras, because language itself is a vehicle of meaning that translation inevitably alters.
• Technical verifiability: every URL automatically tested to ensure the student actually finds a functioning station.

But these criteria, however explicit, do not eliminate the problem. They make it visible. And making it visible, I believe, is already a contribution: every catalog of sources for the study of religions carries with it an implicit hierarchy, and declaring one's selection criteria is the first step toward being able to discuss them.

Curation, in other words, is the sonic counterpart of the prompt: a design act that determines what the user will be able to experience, and what will remain excluded.

V. Three Discoveries from Sound

As with SacredTexts AI, the development of SacredSounds produced unexpected encounters with questions that transcend the technical domain.

1. The asymmetry of online presence as cultural data

Christianity has 17 stations and the Bahá'í Faith has one. This disproportion is not the result of my curation: it is a faithful reflection of the online radio presence of different traditions. The infrastructure of audio streaming (servers, bandwidth, technical standards) is a product of the technological West, and religious traditions access it unequally.

Sikhism, with 7 stations (5 of which managed by SikhNet alone), represents an interesting case: a relatively small community with a sophisticated digital infrastructure, built by the Sikh diaspora in North America. Indigenous traditions, by contrast, have 2 stations: not because their sonic practices are less rich, but because the model of online radio is culturally foreign to many native communities.

This imbalance is research data, not a catalog defect. Mapping the online radio presence of religious traditions implicitly maps the relationship between spirituality, technology, and globalization.

2. Real time as hermeneutic dimension

SacredTexts AI operates on texts that are, strictly speaking, timeless: the Bhagavad Gita is the same at whatever moment the user submits it for AI analysis. SacredSounds introduces a variable that SacredTexts AI does not have: time.

The student who accesses Darbar Sahib at 4 AM (Indian time) hears the Asa di Var, the morning prayer. At 9 PM they hear the Rehras Sahib, the evening prayer. The content changes because religious practice is organized in time. Radio, unlike text, is not indexable: one cannot "search" for a specific passage in the audio stream as one searches for a verse in a database.

This non-indexability is a technical limitation, but also an epistemological quality. It forces a listening experience that more faithfully mirrors the experience of the believer: subject to time, to chance, to the unpredictability of what is being broadcast at that moment.

3. Sonic translation as unsolved problem

SacredSounds is bilingual for the interface (Italian and English), but stations broadcast in over 20 languages: Arabic, Hebrew, Hindi, Sanskrit, Punjabi, Japanese, Chinese, Turkish, Pali, Amharic. The application provides descriptions and context, but the sound remains in the original language.

This is intentional, but it generates a tension. The Western student listening to Sheikh Alafasy's Quranic recitation does not understand Arabic; what they perceive is the musicality, the rhythm, the vocal ornamentation: the aesthetic dimension of recitation, stripped of semantic meaning. Is this an impoverishment or an enrichment? The Islamic tradition itself distinguishes between tajwīd (the art of correct recitation) and tafsīr (interpretation): the sound of the Quran has an intrinsic value that precedes the understanding of its meaning.

This tension between semantic comprehension and sonic experience is, I believe, one of the central questions for any digital tool that claims to mediate the encounter between religious traditions.

VI. Artificial Intelligence in Comparative Listening

The Comparative Explorer

SacredSounds integrates an artificial intelligence module (the Comparative Explorer) that uses Google Gemini to generate comparative analyses between religious traditions. Unlike SacredTexts AI's analytical engine, which operates on individual textual passages, the Comparative Explorer works at a macro level: the user selects 2-3 traditions and a theme (sung prayer, sacred music, meditation, rituals and liturgy, sacred texts, festivals and calendar), or formulates a free question, and the AI produces a structured analysis with cross-connections, significant differences, a listening guide, and suggested stations.

The prompt governing this analysis is an artifact designed with the same care as SacredTexts AI's hermeneutic prompt. It defines the AI's role as "expert in comparative religious traditions and sacred music," establishes constraints on the output structure (7 mandatory fields in JSON format), and explicitly asks to connect conceptual analysis to specific radio stations in the catalog.

This connection is the key point: the AI produces not only propositional knowledge, but suggests a listening path. The answer to "what are the connections between Buddhist meditation and Christian contemplative prayer?" does not end in text, but includes "listen to Buddha FM and then Ancient Faith Radio to grasp the difference between meditative silence and sung prayer."

The output also includes a caveats section (explicit warnings about simplifications, scholarly controversies, and analytical limitations) representing a form of epistemological honesty encoded in the tool's design.

The free question as epistemological opening

Unlike preset themes, the Comparative Explorer's free question allows the user to pose questions the designer had not foreseen. A student can ask: "How does the concept of karma compare in Theravada Buddhism, Vedanta Hinduism, and Sikhism?": a question that crosses three traditions with a specificity no preset menu could have captured. The AI responds with a structured analysis citing specific texts, distinguishing interpretive schools, and suggesting stations for each tradition.

This openness to the unforeseen question is, in my experience, the point where the tool stops being an application and becomes a research environment.

VII. The Hermeneutic Cycle: Text, Graph, Sound

The combined use of both applications suggests a hermeneutic cycle that alternates between different cognitive modalities:

1. Text → Analysis: the student selects a sacred text in SacredTexts AI (e.g., the Gayatri Mantra) and explores the graph of connections generated by the AI.
2. Analysis → Listening: the same student moves to SacredSounds to hear the Gayatri Mantra recitation live on Divyavani Radio or Radio Sai Global Harmony.
3. Listening → Comparison: navigating between traditions, the student discovers sonic parallels: the Gregorian chant on Concertzender Early Music resonates with the Buddhist sutras on Lam Rim Radio in ways that text alone cannot convey.
4. Comparison → Research: returning to SacredTexts AI, the student analyzes the corresponding texts with the AI, generating recursive research questions that deepen the connections intuited through listening.

This cycle was not planned. It emerged from the combined use of both tools, and I believe it represents an example of what this seminar intends to explore: how digital tools generate research paths that were not possible, or thinkable, before their construction.

SacredSounds is not a separate project from SacredTexts AI. It is its expansion into the sonic domain: the passage from text to voice, from analysis to listening, from graph to map, from algorithmic sonification to real sound. One without the other would be incomplete: text without sound is abstraction; sound without text is opacity.

VIII. Research Questions That Could Not Have Existed

I now turn directly to the first question this seminar poses: what research questions can we formulate through AI that are more complex and articulated than those we had before?

Based on my experience building these two tools, I identify several categories of new questions. They are "new" not because no one has thought of them before, but because AI makes them operationally tractable: they can now be investigated systematically, at scale, and with reproducible methods.

The epistemology of AI-generated ontologies

When AI analyzes a passage and produces a graph structure, it is constructing an ontology: a model of what the text "is about" and how its components relate. But this ontology is not fixed: it varies according to the model used, the prompt design, the temperature setting, and even the specific API call (the same request can produce different results due to stochastic processes in generation).

This is a question that requires not a single AI analysis but hundreds, compared computationally. It is a form of systematic hermeneutic survey with no pre-AI equivalent.

Bias cartography

AI models are trained on corpora that represent the world unevenly. When Gemini analyzes a passage from the Dhammapada, its analytical framework has been shaped by training data in which Christian theological categories are vastly overrepresented compared to Buddhist academic literature in Pali. The output inevitably carries this imbalance.

But this is not just a problem: it is a research opportunity. By systematically comparing AI analyses across different traditions, we can begin to map the cultural biases of training data. Where does the AI perform best? Where does it produce more hallucinations? Where does it fall back on Christian theological vocabulary when analyzing non-Christian texts? This produces a bias cartography valuable not only for religious studies but for any field using LLMs for cross-cultural analysis.

The ecology of religious listening in the digital age

How does the relationship with the sacred change when a Quranic recitation, a Sikh kirtan, and Gregorian chant are accessible from the same device, one click apart? Does technological contiguity produce cognitive contiguity? And does this contiguity truly foster interreligious understanding, or does it risk flattening differences into an aesthetics of "generic sacred"?

The performativity of sacred sound out of context

A radio station broadcasts for a specific community: Kol Chai broadcasts for the Israeli Jewish community, Radio Maria for the Italian Catholic community. The religious studies student accessing these stations is, by definition, outside the intended receiving context. Is academic listening participant observation? An act of appropriation? A new type of contemplative practice?

The prompt as variable in hermeneutic experiments

If we accept that the prompt is a hermeneutic framework, then we can treat it as an experimental variable. The same text can be analyzed through different prompts (one emphasizing historical context, another linguistic structures, a third ethical implications), and the resulting graphs can be compared. This means that prompt design for humanistic applications is not an engineering task but an academic one, requiring the same rigor we apply to the design of a research methodology.

Safety filters as a general AI governance challenge

The discovery that AI systems classify sacred texts as "hate speech" is not merely a religious studies problem. It is a case study in the broader challenge of applying universal content moderation to culturally specific material. Sacred texts, because they combine extreme cultural sensitivity with extreme interpretive complexity, are probably the hardest case. In this sense, religious studies are not consumers of AI safety research: they are its drivers.

IX. Transferability: From Religious Complexity to Other Domains

I now turn to the seminar's second question: how can this complexity be channeled into the development of technologies useful to other domains?

Multimodal knowledge representation

The combination of text, graph, and sound in these two projects is not decorative. Each modality encodes different aspects of the analysis. Text conveys propositional content. The graph conveys structural relationships. Sound conveys qualitative and perceived dimensions of exploration. Together, they create an epistemic environment richer than any single modality.

This model of multimodal knowledge representation, developed for sacred texts, has evident applications in fields ranging from medical diagnosis (where complex patient data could be visualized and sonified for pattern detection) to urban planning (where the topology of social, economic, and environmental relationships could be explored spatially and acoustically).

The humanities, in this case, are not borrowing visualization techniques from data science. They are generating new approaches to multimodal representation that data science can adopt.

The tool as question generator

The most transferable characteristic of both projects is not a specific technology, but a pattern: the digital tool as a generator of research questions that were not formulable before its construction. This pattern, building to discover, is applicable to any domain where the complexity of the object of study exceeds the capacity of a single researcher to explore it manually.

X. Limits, Responsibilities, Open Questions

The application includes a disclaimer: its outputs have not been validated by specialists in any of the traditions it covers. The AI does not "understand" the texts: it generates plausible analytical structures based on statistical patterns in its training data. The gap between plausible and correct is the space where human expertise remains irreplaceable.

My lack of training in religious studies meant that I made design decisions a specialist would not have made: some productive, and some certainly naive. The tool would be better if it had been built in collaboration with scholars of every tradition. I hope this seminar can be the beginning of such collaborations.

The democratization of analysis carries risks. A student using this tool might come away with the impression of "understanding" the Bhagavad Gita because they have seen its conceptual graph and read the AI's thematic summary. This is a seductive illusion. The tool must be carefully positioned: not as a substitute for reading, but as an invitation to read more deeply, as a provocation that generates questions rather than answers.

I do not know whether these are research tools or objects of research themselves. I suspect that, as with many of the frontiers this seminar proposes to explore, the distinction is less clear than it seems, and that this ambiguity is exactly the kind of fertile ground from which the most interesting questions can emerge.

Projects:

• SacredTexts AI: https://sacred-texts-ai.vercel.app
• SacredSounds Radio Hub: https://sacredsounds.francescomariano.art

Behind-the-scenes section of SacredTexts AI:
https://sacred-texts-ai.vercel.app/behind.html

Author: Francesco Mariano, Multimedia Artist & Lecturer
Academic context: RESILIENCE Superseminar, Reggio Emilia, March 9-11, 2026
Development: Francesco Mariano with assistance from Claude (Anthropic) and Gemini (Google)
Technology stack: Vanilla JavaScript (ES6+) with no build step, p5.js 1.9.0, Tone.js 14.8.49, Google Gemini AI, IndexedDB, Service Worker (PWA)