Epistemics of the Virtual (EpiVir)

• EpiVir is a research field that examines how virtual, mixed, and AI-mediated environments structure knowledge through dynamic cognitive and material interactions.
• It highlights the role of simulations, urban data, and agentic virtual objects in co-producing insights and reshaping epistemic authority.
• Research in EpiVir employs formal models and empirical studies to analyze epistemic cues, group knowledge, and the effects of AI on knowledge validation.

The epistemics of the virtual, or “EpiVir” (Editor’s term), can be understood as the study of how knowledge, belief, evidence, authority, agency, and interpretation are constituted in virtual, mixed, simulated, and AI-mediated environments. Across the relevant literature, the virtual is not treated uniformly as an alternate field. Instead, it is described, variously, as a “dynamic cognitive and sensitive interaction with reality,” as a layer of urban information and knowledge generated through “myriads of micro-histories,” and as a domain in which simulations function as “computable executable hypotheses” (Bodon, 2024, Iaconesi et al., 2015, Tolk et al., 2013). This suggests that EpiVir is less a doctrine about unreality than a family of inquiries into how mediated environments produce epistemic effects: what can be known there, by whom, under what conditions, with what forms of validation, and through which infrastructures.

1. Conceptual foundations of the virtual

A major strand of the literature rejects the view that virtuality is a substance, a parallel world, or a merely modal category such as possibility. In a Peircean and pragmatic formulation, virtuality is an “operating capacity” (virtus) that produces simulations of real or fictional contexts in order to experiment with their effects. On this account, VR and AR are “symbolic assemblies”: combinations of signs produced by computation and completed by semiosis, whose interpretation can generate real effects in perception, action, and understanding (Bodon, 2024). The virtual is therefore not “fake”; it is functionally real insofar as it reorganizes senses and understanding.

Urban theory extends this de-substantialized view by arguing that the distinction between virtual and physical increasingly collapses under conditions of wireless sensing, smart dust, and participatory sensing. The “Third Infoscape” is defined as “the information and knowledge generated through the myriads of micro-histories, through the progressive, emergent and polyphonic sedimentation of the expressions of the daily lives of city dwellers.” In this formulation, urban meaning is not a separate digital overlay but a distributed informational ecology in which data, memory, affect, and algorithmic mediation are entangled (Iaconesi et al., 2015). The associated claim that “information mutat[es] into landscape” further shifts the ontology of the virtual from representation to spatially operative mediation.

A complementary formulation appears in work on “virtual materiality.” There, materiality in virtual environments is defined as environments being composed of objects that can “actively influence user experience.” The key move is to distinguish this from surface texture in graphics and instead align it with sociomateriality, actor-network theory, and intra-action: virtual objects are treated as agentic participants in cognition, affect, and reflection (Arya et al., 22 Apr 2025). Taken together, these approaches imply that the virtual is best understood not as an immaterial supplement to reality, but as a regime of effective mediation in which signs, sensors, objects, and interfaces participate in the production of knowledge.

2. Knowledge formation through presence, materiality, and agency

Empirical VR research treats virtual environments as epistemically active rather than neutral containers. In a preliminary VR reflection intervention, participants moved through a nature-based environment while answering repeated “clean” questions about thesis or major-project planning. The analysis identified Association, Memory, Emotion, Metaphor, Moral/value reflection, Social reflection, and Embodied cues as environmental mechanisms, with metaphor, emotion, and memory the most prominent. Concrete properties such as size, shape, arrangement, lighting, scenery, motion, elevation, immersion, presence of environmental objects, and interaction affordances were found to shape reflection, often producing “unplanned influences” and helping participants move toward Change or Action (Arya et al., 22 Apr 2025). The epistemic implication is that virtual objects and layouts can co-produce insight.

Social VR studies show that these epistemic effects are cue-sensitive. When the realism of co-located virtual others and self-avatars was manipulated in a $2 \times 3$ mixed design, realistic groups of virtual others were judged more human-like and increased co-presence and the impression of interaction possibilities, while realistic personalized self-avatars improved virtual body ownership and self-identification. By contrast, incongruence between a stylized self-avatar and realistic virtual others diminished self-location and self-identification. The paper interprets realism and congruence as epistemic cues by which participants judge whether entities could “realistically or possibly exist,” whether they are socially available, and whether the virtual social scene “hang[s] together” with expectations (Mal et al., 2024). Virtual knowing here is relational: not just object recognition, but inference about social affordance and bodily fit.

Research on virtual co-embodiment sharpens this point by showing that agency in virtual environments is not a transparent readout of causal control. In a proof-of-concept where two users shared the same virtual avatar from a first-person perspective, participants were generally good at estimating their real level of control, yet they significantly overestimated their sense of agency when they could anticipate the avatar’s motion. The main explicit measure of agency, Feeling of Control, increased linearly with control weight, but shared targets and constrained trajectories made users feel in control even at very low actual control (Fribourg et al., 2019). This suggests that epistemic access to one’s own action in virtual settings is inferential and distributed, shaped by prediction, task structure, and high-level expectations rather than by exclusive motor authorship alone.

3. Formal models of virtual knowledge, evidence, and simulation

Several formal programs give EpiVir a precise logical and semantic structure. In synchronous distributed systems, replacing the standard epistemic operator $K_i$ with the node-based $K_{(i,t)}$ makes time explicit within epistemic nesting and yields a richer language for coordination. Generalized mutual knowledge is defined by $E_A\varphi := \bigwedge_{a\in A}K_a\varphi$, and generalized common knowledge by $C_A\varphi := \bigwedge_{k\ge 1}(E_A)^k\varphi$. In this setting, epistemic necessity is decoupled from temporal necessity: one can represent claims such as $K_{(i,t)}K_{(j,t+3)}\varphi$, and generalized common knowledge corresponds not to simultaneity as such but to “tight coordination,” of which simultaneity is only one instance (Ben-Zvi et al., 2012). The virtual here is not immersive media but a formally articulated space of possible epistemic states indexed by time.

Topological evidence models extend this logic from coordination to group evidence and belief. A multi-agent topo-e-model is given as

$$M=(X,\Pi_1,\ldots,\Pi_n,\tau_1,\ldots,\tau_n,\llbracket\cdot\rrbracket),$$

where partitions $\Pi_i$ represent hard evidence and topologies $\tau_i$ represent soft evidence. Group hard evidence and group soft evidence are constructed by joining these individual structures, and group belief and group knowledge are defined relative to the pooled evidence. The paper’s central claim is that “virtual group knowledge” is what a group could know after sharing all its evidence, not what it already jointly knows. This yields a non-$S5$, fallibilist conception in which group knowledge need not be monotonic in the group argument, because additional shared evidence can defeat prior commitments (Baltag et al., 29 Aug 2025). The accompanying dynamic evidence-sharing operators are shown to be co-expressive with the static logics.

Simulation theory offers a parallel formalization at the level of scientific method. Simulations are treated not as images of reality but as “computable executable hypotheses.” Modeling is abstraction, simulation is execution, and validation becomes analogous to hypothesis testing and theory building. Because computer-based simulation is constrained by computability, simulation-based knowledge is bounded by algorithmic decidability, complexity, and the consistency of the theories embedded in the model. The paper uses model theory, Robinson Consistency, and Łoś’s theorem to analyze equivalence, consistency, and federation across simulation systems, thereby locating virtual experimentation within a mathematically explicit epistemology rather than within mere engineering pragmatics (Tolk et al., 2013).

4. Infrastructures, platforms, and spatial organizations of virtual knowing

Urban and infrastructural analyses emphasize that virtual knowledge is spatially and institutionally organized. In the Third Infoscape, the city is no longer determined primarily by distance and time but by “densities and presences.” Profiles, maps, biometrics, reviews, social networks, and proximity data become part of an informational landscape in which algorithmic aggregation and relational salience reorganize what counts as being “there” in the city (Iaconesi et al., 2015). The paper’s ecological metaphors—roots, cracks, ruins, weeds, overgrowth—present the virtual as a distributed reserve of traces rather than a discrete digital layer.

Museum and archive research frames this organization as discursive place-making. Drawing on heterotopias, one framework treats digital environments as “discursive playgrounds” in which multimedia becomes the “language of information,” archival practices become a form of self-formation, and place production becomes an epistemic layer. Implemented across PC, VR, and MR, the system uses metadata, sorting and grouping algorithms, and procedural generation methods such as Binary Space Partitioning, Cellular Automata, Growth Algorithm, and Procedural Room Generation to turn storage into exhibit, exhibit into exhibition, and exhibition into discourse (Korkut et al., 2023). This suggests that virtual environments do epistemic work not only by presenting content but by spatially reordering it.

Platform studies generalize this point to social systems. Epistemic networks, or epinets, model platforms as generators of higher-order knowledge: who knows what, who knows who knows what, how trust is structured, and how covertness is maintained. Different platforms generate different epistemic neighborhoods. Point-to-multipoint systems such as Twitter create distributed knowledge, while automatic read receipts in WhatsApp or WeChat more quickly produce mutual knowledge; Slack private channels can instantiate security neighborhoods and covert conduits (Moldoveanu et al., 2021). The platform is therefore an epistemic architecture, not just a communication channel.

A broader diagnostic framework, “Situated Epistemic Infrastructures” (SEI), argues that LLM-era knowledge systems are entering a “post-coherence” condition. In such environments, credibility is mediated across institutional, computational, and temporal arrangements rather than anchored in a single domain or community. SEI analyzes this mediation through Infrastructure Typology, Power Signatures, Symbolic Compression, and Breakdown Dynamics, and explicitly shifts inquiry from classification to coordination (Kelly, 7 Aug 2025). Virtual knowledge environments, on this view, are hybrid material-symbolic arrangements in which authority is provisional, compressed into portable markers, and exposed most clearly when infrastructures fail.

5. AI-mediated virtual epistemics

AI-centered work extends EpiVir into dialogical and generative systems. “Epistemoverse” proposes a “metaverse of knowledge” in which AI-reincarnated philosophers engage in iterative discourse, retrieval, and question-asking. The paper’s evidence comes from dialogues among clones of Aristotle, Nietzsche, Machiavelli, and Sun Tzu in the Syntropic Counterpoints installations. Knowledge is operationalized as concept networks whose nodes are extracted concepts and whose edges are either lexical overlap or “maieutic links” generated by another agent’s question. Connectivity is summarized by average degree centrality,

$K_i$0

and the maieutic condition yields denser networks and fewer isolated nodes, with graphs reported to converge after about 60 chunks (Nikolić et al., 13 Dec 2025). The virtual environment here is an epistemic space for preservation, reinterpretation, and extension of intellectual heritage through interaction.

At the same time, work on epistemic modality in LLMs shows that generative fluency does not guarantee reliable epistemic expression. Using controlled stories rather than broad QA benchmarks, one study tests whether LLMs choose appropriate modal auxiliaries and attitude verbs under manipulated evidence and commitment conditions. Experiment 1 uses 150 stories across base, 1-shot, and counterfactual formats to probe must/have to versus may/might; Experiment 2 uses 30 ToMi-derived stories expanded into eight statements each to probe know, believe, and doubt. Across eight open-weight instruction-tuned models, larger models perform better, but necessity is easier than possibility, fact-based statements are much easier than belief-based statements, paired accuracy is more difficult than item accuracy, and joint accuracy is often 0 or near 0 (Li et al., 2 Jun 2025). The paper concludes that model-generated uncertainty expressions are limited and not robust, so verbalized uncertainty is not a transparent readout of epistemic competence.

A further layer concerns evaluation. A corpus study of 553 journal articles on AI published in 2024 maps epistemic virtues and values such as true, accurate, transparent, coherent, creative, responsible, predictable, and generative. The proposed framework uses virtue epistemology to ask what counts as “good knowledge” when AI reconfigures knowledge work into hybrid human-machine assemblages. Creativity is treated as a pivotal case, but the paper’s distinctive move is to widen the evaluative horizon toward “generativity,” conceived as more scalable, distributed, and inclusive than older ideals of singular creativity (Liu, 2 Jul 2026). This suggests that AI-mediated EpiVir is not only about what virtual systems know, but also about how their outputs and processes should be normatively assessed.

6. Debates, limitations, and normative directions

The literature is not unified on the ontology of the virtual. The dominant tendency is pragmatic, semiotic, material, or infrastructural: virtuality is an operation, a symbolic assembly, an active environment, or a mode of mediation (Bodon, 2024). A more speculative strand argues that the physical universe itself may be a “virtual reality created by information processing,” distinguishing a calculable universe, a calculating universe, and a calculated universe. In that hypothesis, relativity and quantum theory are interpreted as outputs of informational processing, the big bang is read as a “boot-up,” and incomputability is proposed as a possible falsifier. The paper is explicit, however, that its case is abductive and analogical, not a direct proof, and that it offers no concrete experimental program distinguishing virtual-reality theory from objective realism (0801.0337). This marks a genuine controversy: whether virtuality is primarily a mode of interaction within reality or a candidate ontology of reality itself.

Empirical work also remains limited in scope. The virtual materiality study is explicitly preliminary: it has a small sample size of 21 participants, studies only thesis or major-project reflection, uses only nature environments, provides no non-virtual baseline, and supports correlation rather than causation (Arya et al., 22 Apr 2025). The avatar realism experiment notes that its virtual others were behaviorally restricted to maintain consistency, which likely lowered perceived interaction possibilities (Mal et al., 2024). The co-embodiment study is a proof-of-concept with 24 male participants and a tightly controlled setup (Fribourg et al., 2019). These limitations do not negate the epistemic claims, but they constrain generalization across tasks, populations, and virtual forms.

Normatively, several papers converge on the need for legibility, transparency, and participatory governance. The Third Infoscape argues that meaningful urban information requires legible relational graphs, transparent generation and processing, and the ability to intervene, remix, and propagate information, culminating in the concept of the “Ubiquitous Commons” (Iaconesi et al., 2015). SEI, by contrast, emphasizes “anticipatory and adaptive” epistemic stewardship under post-coherence conditions, with special attention to symbolic drift, algorithmic opacity, and breakdown as a diagnostic resource (Kelly, 7 Aug 2025). The AI virtue literature widens this into a whole-society problem of evaluating hybrid knowledge systems by truth, accuracy, creativity, responsibility, and generativity rather than by inherited assumptions about stable human-only knowledge work (Liu, 2 Jul 2026). A plausible implication is that EpiVir increasingly concerns not only epistemic states within virtual environments, but also the design of the environments, infrastructures, and evaluative vocabularies through which those states become authoritative.