2. Taxonomic rupture: Linnaean classification and Buffonian continuity

Modern data architecture, encompassing metadata schemas, ontologies, and hierarchical or object-oriented storage systems, can trace much of its conceptual DNA back to Linnaean taxonomy. Linnaeus’s idea of classifying organisms into a tree of categories – kingdom, class, order, genus, species – effectively laid the foundations for what we now recognize as hierarchical models, prevalent in contemporary databases and in Object-Oriented Programming (Booch et al. Reference Booch, Maksimchuk, Engle, Conallen, Houston, Young and Fuller2007). When assigning metadata to a record in a digital repository and positioning it within a hierarchy (e.g., document > scientific text > domain: biology > subdomain: plant taxonomy), we replicate the “box and sub-box” logic Linnaeus employed to organize his herbarium collections and manuscripts. Furthermore, metadata provides codified instructions that allow a “blind” interchange of information, also known as “procedural instructions” (Liu Reference Liu2004; Piez Reference Piez2001). These principles of functional categorization, designed to ensure the interchangeability and comparability of information, became the technical foundation for library indexing systems, data structures in high-level programming languages, and even the design of relational networks in the so-called Semantic Web (e.g., RDF, OWL) (Berners-Lee et al. Reference Berners-Lee, Hendler and Lassila2001).

However, while ontologies have become the bedrock of current digital infrastructure, they often mask a deeper problem – the imposition of singular knowledge systems that marginalize alternative ways of understanding the world. The apparent strength of this approach lies in its supposed universality; reducing complexity to consistent labels greatly facilitates queries, interoperability, and the adoption of global classification standards (Bowker and Star Reference Bowker and Star1999, Reference Bowker and Star2000; Koerner Reference Koerner1999).

In contrast to Linnaeus’s rigid and heavily typified hierarchical classification system, French naturalist Georges-Louis Leclerc, later known as the Comte de Buffon (1707 – 1788) – a contemporary of Linnaeus – believed nature was too complex and continuous to be divided into discrete groups. Buffon proposed that environmental factors could cause species to transform gradually, an important precursor to evolutionary theory that was more attuned to variability and geographical and cultural contexts, thus highlighting the technical limitations of strictly hierarchical approaches (Buffon Reference Buffon1752; Daston and Vidal Reference Daston and Vidal2004). Buffon’s more dynamic view of nature helps us understand how ongoing, complex interactions between organisms and their environment are critical concepts in ecological management, particularly regarding adaptation, speciation, and extinction.

However, in contemporary data analysis, for example, supervised machine learning is typically fed with training sets in which each instance must be assigned to a predefined class (Kotsiantis Reference Kotsiantis, Maglogiannis, Karpouzis, Wallace and Soldatos2007; Whang et al. Reference Whang, Roh, Song and Lee2023). This reproduces Linnaean philosophy: “find the correct drawer” based on a set of attributes, often overlooking the fluid transitions, hybridity, or anomalies in nature and society. Recent models – such as deep neural networks and transfer learning techniques – have begun introducing more flexible layers of representation, yet they remain reliant on meticulously labeled data structures that assume, in many cases, a universal classification principle (Jo and Gebru Reference Jo, Gebru, Hildebrandt and Castillo2020). Here lies the tension Buffon already identified: an actual phenomenon may fit neatly into two or more classes simultaneously, or it may shift category depending on context, and “hard” data systems rarely admit this dynamism. Instead of simply organizing what exists, hard data systems privilege certain ways of knowing while rendering others invisible.

Recent ethnobiological research demonstrates how Buffon’s dynamic perspective can be operationalized in concrete computational systems that diverge from Linnaean rigidity. For example, in Brazil, Albuquerque et al. (Albuquerque Reference Albuquerque, Albuquerque and De Medeiros2015, p. 4) describe the Ontological System for Biocultural Ethnoclassification (Sistema Ontológico para Etnoclasificación Biocultural [SOEB]), which encodes graph-based ontologies to reflect the overlapping ways communities classify plants beyond binomial nomenclature. Instead of forcing Ocimum campechianum (a local basil) into a single hierarchical taxon, SOEB models its ceremonial, culinary, and medicinal associations as simultaneously valid nodes linked by weighted edges. This approach captures cultural nuances and ecological relationships, illustrating how local knowledge can inform ontologies to embrace contextual depth rather than categorize according to fixed, Linnaean labels.

A parallel initiative in Mexico is detailed by Toledo and Barrera-Bassols (Reference Toledo and Barrera-Bassols2008, p. 94), who implemented the Sistema de Conocimiento Biocultural (SCB) to integrate Indigenous taxonomies into a multidimensional data framework. Technically, SCB uses fuzzy logic to store these traits, allowing classification “boundaries” to shift with seasonal or ceremonial context. As the authors note, this method “reflects the continuity and complexity” of Indigenous epistemologies far more faithfully than the single-axis structuring of Linnaean models (Toledo and Barrera-Bassols Reference Toledo and Barrera-Bassols2008, p. 101).

The tension between Linnaean classification and Buffonian continuity reveals a fundamental challenge which is explicit in hard data systems that rarely acknowledge Buffonian dynamism. While Linnaeus’s rigid classification system offers computational efficiency and standardization that commands databases and algorithms, it arbitrarily imposes definitions – and thus restrictions – on phenomena that are inherently fluid, contextual, and interconnected. Buffon’s more dynamic, continuous vision of natural relationships better captures the holistic continuity of nature and culture, where categories blend, overlap, and transform based on environmental contexts, acknowledging diverse ways of knowing that might better reflect the interconnected reality of living systems.

This has particularly significant implications in Indigenous taxonomies, which often conceptualize plants, language, and natural resources through a large relational semiotic system – a sort of grand system of communication within an ecological context (Bridle Reference Bridle2022). When Indigenous knowledge systems encounter Western scientific classification, vital knowledge such as language, medicinal properties, seasonal variations, and sustainable harvesting practices progressively get lost. Such taxonomic ruptures are not merely formal or computational concerns but have tangible consequences for biodiversity conservation, bioprospecting, intellectual property rights, and the exploitation of natural resources, particularly in Indigenous territories, although they spread with a ripple effect throughout greater society as a whole. For example, Cámara-Leret and Bascompte (Reference Cámara-Leret and Bascompte2021) point out that declines in language diversity directly impact plant diversity and the production of food and medicines, and their research shows that language loss is an even more critical factor in the disappearance of medicinal knowledge than biodiversity loss, as 75% of useful plants in three regions of the world (including the Amazon) have names only in Indigenous languages (Cámara-Leret and Bascompte Reference Cámara-Leret and Bascompte2021; Zariquiey Reference Zariquiey and Mariátegui2025).

3. AI’s inheritance of Linnaean classification systems

Contemporary AI metadata systems, especially in supervised learning and many deep-learning approaches, largely inherit Linnaean taxonomy’s logic by viewing data annotation as a singular, stable labeling process. In practice, this means that each semiotic notation rendered as data in a technological platform – an image, text, or audio file – must belong to a discrete class or set of classes established in advance.

Building on Jaton’s careful examination of algorithmic constitution, we can see how this taxonomic imperative materializes through what he calls the “ground-truthing” process – the creation of authoritative datasets that serve as the bedrock for algorithmic training (Jaton Reference Jaton2020). These ground truths do not merely represent reality; they actively constitute it through practices of selection, cleaning, and labeling that render messy phenomena into computationally tractable forms. Just as in Linnaean binomial nomenclature, every species is rigidly placed within a genus (Koerner Reference Koerner1999). Selecting appropriate labels is a crucial step, as an AI model’s performance depends on the consistent, large-scale application of these metadata.

However, much like Linnaean taxonomy, this method is reductive: it requires deciding from the outset which categories exist, thus legitimizing the worldview that establishes them – one generally aligned with Western frameworks – without accommodating the diverse perspectives or uses that may converge around the same object or phenomenon (Milan and Treré Reference Milan and Treré2019). Therefore, data is never neutral, as it embodies design decisions, path dependencies, and organizational aims (Alaimo and Kallinikos Reference Alaimo and Kallinikos2024). As a result, each increasing datafication promotes the homogenization of knowledge and knowledge-making (Yoo et al. Reference Yoo, Henfridsson and Lyytinen2010). James Bridle elegantly calls this the reinforcement of the “one world fallacy” – a misplaced objectivity that is the result of the Enlightenment of the 17th and 18th centuries and the scientific revolutions that followed which aimed to eliminate inconsistencies and idiosyncrasies, promoting the “belief that the world has a single, coherent narrative and that there exists a one-size-fits-all framework for interpreting it” (Bridle Reference Bridle2022, p. 68).

Within AI architectures that employ data annotation pipelines, the insistence on univocal categories aims to streamline training for classification or object-detection models. This practice is directly linked to the Linnaean hierarchical structure: we ascend and descend a taxonomy (or tree diagram) until locating the correct “leaf node” for each example. While this organization yields computational simplicity, as each instance can be unambiguously positioned, accelerating the convergence of learning algorithms, the trade-off is significant. Just as Linnaean taxonomy overlooked the ecological and cultural subtleties Buffon highlighted (Buffon Reference Buffon1752), these metadata systems disregard contextual factors (regional, symbolic, or situational) that might alter a datum’s interpretation or value.

This Linnaean-style classificatory logic is evident in protocols and standards like RDF (Resource Description Framework) and OWL (Web Ontology Language), where relationships and properties are defined in highly typified, hierarchical ways (Berners-Lee et al. Reference Berners-Lee, Hendler and Lassila2001). Although these tools are certainly an advance for semantic interoperability, they, as mentioned, also tend to promote a model in which each entity “is” something defined by a central vocabulary, neglecting the possibility that one single object may simultaneously belong to multiple, culturally situated taxonomies (Aikenhead and Michell Reference Aikenhead and Michell2011)

4. The struggle of indigenous taxonomies in a world of Linnaean dominance

Labeling data under a Linnaean classification goes beyond platforms and technologies, functioning as an expression of power with far-reaching cultural and ecological ramifications. As mentioned, when a dataset is “correctly labeled” according to the dominant scientific paradigm, it can obscure entire systems of values and knowledge, resulting in a narrow or biased view of reality. Western industrial farming provides us with a good example of this. The tension between amaranth, a sacred grain for many Indigenous cultures, and the expansion of transgenic soybean monocultures in South America is not simply agricultural but a manifestation of deep ecological and cultural struggle. Agro-industrial corporations have long sought to erase amaranth by reducing it to the status of a “weed,” yet scientific studies demonstrate its remarkable resistance to glyphosate, the cornerstone herbicide of industrial soybean production. In Argentina, a population of Amaranthus palmeri was documented to resist glyphosate not through the common target-site mutation, but via mechanisms of reduced herbicide absorption and translocation (Palma-Bautista et al. Reference Palma-Bautista, Osuna, Rojano-Delgado, Lobato-Lourido, Torra, Travlos, Tataridas, Chachalis and De Prado2019). Similarly, glyphosate-resistant biotypes of Amaranthus hybridus have been found in soybean fields in Argentina, linked to multiple mutations in the EPSPS gene (Heap Reference Heap2005). Beyond the biochemical level, Beilin and Suryanarayanan’s article The War between Amaranth and Soy: Interspecies Resistance to Transgenic Soy Agriculture in Argentina (Beilin and Suryanarayanan Reference Beilin and Suryanarayanan2017) interprets these mutations as part of a broader, multispecies resistance in which amaranth embodies ecological defiance against the homogenizing violence of monocultures and agrochemical dependence, Thus, amaranth’s persistence is not merely a biological accident but a symbolic and political act of survival, contesting the colonial logics embedded in industrial agriculture. As Agrawal (Reference Agrawal1995, p. 2) explains, the discourse of development often portrays local knowledge as “inefficient” or “primitive,” underestimating its potential and reinforcing a hierarchy of epistemologies. This tension underscores how relying on a single classificatory lens can perpetuate the marginalization of alternative worldviews that do not fit neatly into standardized taxonomies.

Much of this impulse to classify and normalize hinged on the cameralist approaches popular in Linnaeus’s Sweden, where knowledge was marshaled to serve state interests and a new kind of empiricism based on statistics (Peters Reference Peters1988). As Koerner (Reference Koerner1999) explains, Linnaeus’s strategic vision went beyond botanical curiosity, as he imported profitable crops, cultivated them locally, and located indigenous substitutes for expensive foreign goods, thereby reducing reliance on external markets. This cameralist project dovetailed with the expansion of imperial powers and required comprehensive taxonomies to identify, catalog, and exploit “useful” species while often dismissing or marginalizing those that did not fit state or colonial profit schemes.

From that standpoint, the large-scale exportation of Linnaean classification – facilitated by colonial administrations in the 18th and 19th centuries – operates in a fashion similar to the modern standardization of metadata. Both reinforce monocultures (whether botanical or informational) by branding traditional or local knowledge as irrelevant or secondary (Koerner Reference Koerner1999). Going back to the example of what is considered a “weed,” we see how the assumed neutrality of a classification tree intersects with the power to impose categories, rendering invisible the multiple uses and meanings a plant may hold, expanded further to include dataset or even entire cultural context (Agrawal Reference Agrawal1995). This legacy continues in today’s supervised classification algorithms and hierarchical data structures: while highly efficient, they frequently narrow our perception of reality to only what the system can measure, categorize, and monetize.

Hence the urgency to design data architectures in which binomial logic is not the sole framework, and where aspects such as ambiguity, cosmology, symbolism, local evolution, and cross-category overlap can find meaningful technical support rather than mechanically reproducing the hegemony of a single rationality that assumes – due to specific but narrow practical considerations – the superiority of one technology over another (Rodríguez-Alegría Reference Rodríguez-Alegría2014). It is not that Western classifications are inherently superior in their ability to represent reality or organize knowledge into tools and AI systems, but rather their dominance stems from asymmetric power relations that determine who controls the technological infrastructure, who funds research and development, and who shapes global technical standards. Ultimately, large corporations use their economic resources to develop, patent, and distribute systems that enable market dominance.

Consequently, the resulting issues extend beyond efficiency or predictive accuracy. Numerous AI ethics studies indicate that these metadata systems, grounded in Linnaean rigidity, perpetuate colonial or capitalist biases when applied in fields like precision agriculture, environmental policymaking, and even pattern recognition in social data (Birhane Reference Birhane2021; Mohamed et al. Reference Mohamed, Png and Isaac2020). Failing to include categories reflecting Indigenous ideas of “companion plants” or “rituals embedded in cultivation” systematically omits local knowledge that might radically reshape an AI model’s recommendations about fertilizers or pesticides (Kimmerer, Reference Kimmerer2013; Ludwig and El-Hani Reference Ludwig and El-Hani2020). As Ludwig and El-Hani (Reference Ludwig and El-Hani2020, p. 12) argue, “The epistemic frameworks of Indigenous communities often represent relationships between organisms in ways that are inadequately captured by Western taxonomic systems,” leading to AI systems that reproduce epistemological hierarchies while claiming objective neutrality. A similar situation arises when a labeling system, initially devised for purely operational purposes, crystallizes into a single source of “truth” within a large-scale technological infrastructure.

In response, some recent developments in AI and database design attempt to incorporate label multiplicity, multi-layered annotation, or the coexistence of different metadata schemas for the same dataset. Exploratory structures based on reticular or heterogeneous graphs are closer to Buffon’s complexity, allowing an entity to belong to several categories at once, depending on viewpoint or analytical goals (Knublauch and Kontokostas Reference Knublauch and Kontokostas2017). Nonetheless, these approaches remain comparatively marginal in the face of Linnaean tradition, which enforces categorical uniformity. As a result, systems in agriculture, climate science, healthcare, or computational sociology typically ignore any variable that does not “fit” neatly.

Far from resolving bias, AI then reinforces a colonial and industrial reading of reality, portraying it as the most “efficient” or “practical” approach while essentially imposing an arbitrary reduction of complexity (Escobar Reference Escobar2018). As Ricaurte explains – following the work of Peruvian sociologist Aníbal Quijano–, this data-centric rationality should be understood as “an expression of the coloniality of power, manifested as the violent imposition of ways of being, thinking, and feeling that leads to the expulsion of human beings from the social order, denies the existence of alternative worlds and epistemologies, and threatens life on Earth.” (Ricaurte Reference Ricaurte2019, p. 351) The dominant epistemology assumes that it rationally reflects reality and generates the most valuable knowledge for greater efficiency, but at the same time, these assumptions reinforce capital concentration and colonial power relations. This epistemological violence becomes particularly evident when contrasted with Indigenous knowledge systems that explicitly acknowledge territorial and relational foundations of knowledge. As Lewis et al. (Reference Lewis, Arista, Pechawis, Kite, Arista, Costanza-Chock, Ghazavi, Kite, Klusmeier, Lewis, Pechawis, Sawyer, Zhang and Zhang2021) observe, Indigenous epistemologies recognize that relationality emerges from specific places and contexts, with particular worldviews arising from the dynamic forces operating within specific territories. In these systems, knowledge functions as guidance for navigating relationships within particular landscapes, where language, cosmology, and ceremony serve as territorial and relational means of sharing wisdom.

A meaningful contrast emerges when we compare the strictly hierarchical Linnaean classification systems with similar approaches to Buffon’s, such as certain Indigenous ontological frameworks from Indigenous cultures that have been nurtured for thousands of years. Whereas Linnaean models focus on morphological attributes to assign organisms to predetermined niches, Indigenous ontological frameworks based on cosmologies emphasize variability, dynamic interrelations, and continuity between environment and living beings (Aikenhead and Michell Reference Aikenhead and Michell2011). Moreover, the Indigenous ontological frameworks carry ancestral knowledge in the form of medicine, ecology, and spirituality, presenting ontological schemes in which the boundaries between human, animal, and divine realms are not rigidly compartmentalized (Leon-Portilla Reference Leon-Portilla1990).

Roberto Zariquiey has shown that, in contrast to relatively simplistic Linnaean classification, the Kakataibo people of the Peruvian Amazon reveal that Indigenous taxonomies can demonstrate remarkable complexity and sophistication (Zariquiey Reference Zariquiey2018, Reference Zariquiey and Mariátegui2025). Their classification system divides animals into intricate subcategories based on habitat, locomotion and relationship to humans, creating a multidimensional taxonomy that reflects deep ecological knowledge. For example, while Western biology recognizes one species of tapir (Tapirus terrestris), the Kakataibo distinguish six different types, demonstrating how their language encodes environmental observations that transcend the granularity of Western scientific classification (Zariquiey Reference Zariquiey and Mariátegui2025). Unlike Linnaeus’s system, which was developed in part to facilitate the economic exploitation of natural resources, the Kakataibo taxonomy demonstrates that this community has not only detailed knowledge but also a humanistic vision of nature. From a technical perspective, this suggests alternatives to linear data structures and discrete labels, opening the door to ontological models that capture multiple layers of meaning and contextual relations.

Another example is found in the Meso-American worldview, where maize was not simply a “crop” or a “nutritional resource” but a plant intertwined with creation myths, kinship networks to deities, and ritual cycles. In the Popol Vuh, the sacred text of the K’iche’ Maya, humans were made from maize dough after previous creation attempts using mud and wood failed (Bazzett Reference Bazzett2018). This requires highly relational ontologies that document transversal connections and describe an entity (in this case, a plant) under different roles in the same database. Graph-based semantic systems could implement such an approach, allowing edges with distinct weights and properties, for example, a node “maize” connecting to “sowing ritual” in one manner and “cultivation practice” in another. This plasticity contrasts with Linnaean taxonomy, which would simply place maize in the “gramineae” category without acknowledging the historic and symbolic tapestry attached to it (Leon-Portilla Reference Leon-Portilla1990; Tsing Reference Tsing2015).

These examples show that in the worldview of many Indigenous cultures, the meaning of an object, living being, or natural event is far from fixed; instead, it takes shape according to the elements present at any given time. A hylomorphic perspective helps to clarify this dynamic, wherein hyle or “matter” (i.e., concrete entities) and morphe or “form” (i.e., the principles that endow them with being and meaning) are not separate, definitive entities but rather become a particular combination of matter and form, articulated within contextual and symbolic networks of similitudes and correspondences (Aristotle 2007; Austin Reference Austin1988; Peters Reference Peters1988). For instance, if a particular bird arrives at dawn and during a specific time of year, an Indigenous community might interpret the event as signaling fertility cycles, protective deities, or omens associated with harvest seasons. The bird is thus not merely a creature bearing a name and biological attributes; it becomes a messenger, a symbol, or a ritual indicator that redefines itself whenever cultural or environmental circumstances converge.

Taking this further, a concrete example of a technical alternative might revolve around graph-oriented databases (e.g., Neo4j, JanusGraph) or RDF/OWL infrastructures enhanced by contextual metadata (Knublauch and Kontokostas Reference Knublauch and Kontokostas2017). Instead of a single monolithic (Linnaean) schema, one could define a core ontology with the possibility of attaching additional “vocabularies” or “sub-ontologies” that correspond to different cosmologies, cultures, or domains of knowledge. Following Buffon’s idea, entities and relationships might be timestamped or georeferenced (e.g., maize in the highlands vs. maize in coastal regions) to reflect spatiotemporal variation. Collaborative annotations by local communities could add attributes and properties not accounted for in the central ontology. This would yield a basic framework (analogous to Linnaean genus and species) while dynamically enabling diverse layers of meaning (ritual uses, traditional farming practices, and historical narratives) to coexist in the same data structure (Daston and Vidal Reference Daston and Vidal2004; Tsing Reference Tsing2015).

Similarly, the Calendar Project emerged in 2002 through Waman Wasi, an institution that mediates support from the plural Indigenous biodiversity visions around forest conservation in San Martin, Peru. Working with the Kichwa Lamas, Awajún, and Shawi communities, the project addresses a fundamental challenge: how to preserve and revitalize Indigenous knowledge systems that have been marginalized by colonial modernity while creating bridges between different ways of understanding and caring for the territory (Leon Reference Leon2023). Unlike Linnaean taxonomies, the Calendar Project embraces a relational approach where humans, non-humans, activities, seasons, and spiritual dimensions are interconnected through dynamic cycles.

This project involves a set of interconnected tools, with the first involving physically walking through territories with Indigenous forest keepers. This practice acknowledges that knowledge exists not in abstract categorization but in embodied movement through living spaces. Thereafter, an organizational tool called the Matrix of Knowledge synthesizes the cyclic knowledge gathered during these walks. The matrix is adaptable, representing the collective nurturing of an articulated, dynamic territory. It consists of six vertical columns (month, season, house, chacra/farm, forest, and water) and twelve horizontal rows for months, creating a framework that connects temporal cycles with spatial domains and activities (Leon Reference Leon2023). Finally, a visual representation called a “community calendar” serves as a time-shift technology, showing the community’s action plan and time commitments. Unlike classification systems which are static, the calendar expires and is collectively updated, reflecting the dynamic nature of ecological relationships.

Contemporary computer science is not unaware of the limitations of rigid, single-label systems, and several technical strategies have attempted to move beyond strict Linnaean hierarchies. Multi-label classification frameworks, for example, allow a single instance to be tagged with multiple categories, while fuzzy ontologies assign partial degrees of membership to capture gradual or overlapping attributes. Graph-based approaches, such as RDF/OWL or knowledge graphs implemented in Neo4j, further emphasize relationality by structuring data as nodes and edges rather than fixed compartments. Contextual ontologies and modular ontology design also gesture towards situational meaning, enabling an entity to shift definition depending on its domain of use. These strategies represent important steps away from univocal categorization, acknowledging at least partially that the world resists being neatly divided into single, universal drawers.

Yet despite these advances, their scope remains technically and epistemically constrained. Multi-label systems treat multiplicity as a matter of adding tags but fail to capture the cultural weight or symbolic priority of those labels. Graph databases encode connections but presuppose centralized vocabularies, leaving little space for contested or contradictory meanings to co-exist as legitimate data. Fuzzy logic introduces mathematical degrees of membership, but the assignment of weights is typically external, decided by system designers rather than negotiated or validated by communities themselves. Likewise, contextual ontologies reduce meaning to predefined rules without addressing temporality, situated governance, or the possibility of semantic conflict. In all these cases, the effort to move beyond Linnaean rigidity remains subordinated to computational efficiency and standardization, rather than embracing relational, plural, and dynamic frameworks of meaning. This is precisely the gap in which a rhizomatic hylomorphism can intervene, not by discarding these techniques but by expanding them to include cultural validation, local governance, and the registration of multiplicity – including contradiction – as data in its own right.

5. Indigenous taxonomies for reimagining AI-driven agriculture

From these examples, if we apply these ideas into a practical case of an AI-driven precision agriculture scenario, the system could, instead of labeling any unplanned plant in a soybean field as “weed,” incorporate cultural variables (whether the plant has medicinal use), ecological dimensions (whether it encourages pollination of local species), and symbolic elements (whether it is tied to community rituals). The AI would then analyze these heterogeneous annotations to propose different actions based on broader objectives rather than merely maximizing soybean yield. Inspired by Buffon and Indigenous taxonomies, this reticular architecture would allow more nuanced decisions beyond a binary “remove or do not remove” approach. Ecological and cultural interdependencies would thus gain real traction. In practical terms, this approach calls for new indexing methods in machine learning algorithms, for example, embeddings that correlate words or semantic nodes along multiple dimensions (ritual, ecological, productive) rather than a single axis of morphological or functional similarity.

Likewise, from a co-design perspective, such a system would facilitate the coexistence of Western and non-Western criteria, expanding the worldview without forcing an immediate unification of categories.

Linnaean taxonomy operates primarily as an ontological system, asserting that certain categories represent the fundamental structure of reality. However, this ontological framework emerged from specific epistemological assumptions – 17th and 18th-century European approaches to observation, classification, and scientific authority that privileged morphological similarity over ecological relationships, cultural significance, or contextual meaning. When AI systems inherit these ontological structures through training data and metadata schemas, they naturalize what are actually epistemological choices. The system treats categories as objective features of the world rather than as one particular way of organizing knowledge among many possible alternatives. This conflation of ontology with epistemology becomes particularly problematic when these systems encounter phenomena that do not fit neatly into its predetermined categories or when they operate in cultural contexts where different epistemological frameworks might yield more appropriate ontological structures.

While Linnaean logic might classify any species into standardized compartments, a multi-layered graph model could simultaneously acknowledge or incorporate “medicinal” or “sacred” attributes. This demands not only a shift in data structures but also rethinking curation, interfaces, and the algorithms that parse these enriched metadata. Most crucially, it also demands a novel approach to comprehending plants and incorporating such definitions as cultural heritage. The goal is to construct a more polycentric data model that can integrate plural realities without collapsing them into a single hierarchical representation (Bowker and Star Reference Bowker and Star2000; Escobar Reference Escobar2018).

Indigenous notions of interdependence among beings, environments, and divine agencies invite a method of information modeling that recognizes permeable ontological frontiers. In practice, this means building technical systems in which each element is not a mere isolated object – stamped once, as with binomial nomenclature – but rather a node whose properties may shift according to community, physical surroundings, or historical moment. Instead of envisaging a “unified” closed ontology, one might create an ontological ecosystem capable of expansion and even internal conflict – where a single node could clash with another scheme if its category differs, and the system would record that dissonance in the dataset. This falls in line with Buffon’s portrayal of nature as unstable and subject to multiple interpretations while also aligning with many Indigenous logics that refuse to reduce a being to an externally imposed category (Aikenhead and Michell Reference Aikenhead and Michell2011).

Embracing Indigenous cosmologies thus demands more than a data-structure overhaul; it requires a methodological shift in how information is collected, curated, and interpreted alongside new user-interface paradigms and different approaches to algorithmic design. The potential reward is a technical system that not only tolerates but actively represents and manages cultural and ecological diversity, minimizing the imposition of static categories and enabling the emergence of multiple narratives (Bowker and Star Reference Bowker and Star2000; Escobar Reference Escobar2018). This step is more than a mere improvement; it is an epistemological and technical evolution in which the Linnaean heritage is finally reconciled with alternative ways of understanding and classifying the world.

As mentioned, the fluidity of alternative classification stands in stark contrast to the rigid classification systems prevalent in Western modernity. From Linnaean taxonomy to mainstream digital ontologies, the prevailing tendency is to fix an object’s identity under a single, definitive label, detached from any symbolic or contextual transformations. A typical computational model, for example, categorizes a bird via metadata such as “species X, family Y,” focusing on its possible usefulness in intensive agricultural production (e.g., is it a pest, or does it help control pests?). Returning to the earlier example of how Indigenous meanings form part of how a bird is defined or conceived of in the culture – for example, the idea that the bird’s presence at a certain hour can change how it is interpreted collectively, thus shifting the bird’s ontological status – rarely appears in conventional databases or classification schemas. This example illustrates why standardized categories fail to capture the multiple and interwoven meanings that emerge from complex socio-ecological contexts. Just as transformative symbolism assigns the bird a different function depending on timing and cultural factors, hylomorphism suggests that “form” and “matter” should not be seen as separate, static entities but rather as realities in constant interaction.

In a non-Western framework, the being of an object is defined by its relationships and by the web of meanings it constructs with other entities or natural processes (Austin Reference Austin1988; Lévi-Strauss Reference Lévi-Strauss1962). Thinking along these lines poses major challenges for any rigid classification, as, for example, what is considered an avian “messenger” today could become a “protector” tomorrow, or in another setting, a bad omen.

In the specific domain of industrialized agriculture, the disparity between a hylomorphic perspective and Western productivist logic becomes starkly visible. Intensive production systems drastically simplify both the diversity and the value assigned to each element: monocropping homogenizes environmental conditions to maximize yield for a single resource, relegating anything outside that resource – another plant, an insect, or a fungus – to a category of “nuisance” or “waste.” This reductionism compels reality to fit within a schema: a list of inputs (genetically modified seeds, fertilizers, pesticides) and outputs (harvests of profitable species) that disregards the multiplicity of relations with the environment. The apparent efficiency of industrial agriculture thus relies on an ontological model diametrically opposed to the ecological and symbolic interdependence inherent in a hylomorphic worldview. It gives no thought to a bird arriving at dawn to signal auspicious omens, or, by its feeding habits, possibly altering soil fertility at a particular point in the lunar calendar (Altieri and Nicholls Reference Altieri and Nicholls2008; Leon Reference Leon2023).

Such extreme utilitarianism, codified in digital classification systems, produces tangible repercussions. By labeling everything purely from a “yield” or “pest-control” perspective, any species failing to boost productivity is, by definition, deemed a threat or obstacle. This in turn homogenizes biomes, accelerates biodiversity loss, and severs the cultural practices that link agriculture with ritual or social dimensions (Shiva Reference Shiva2016). As Vandana Shiva explains, “The narrow measure of yield fails to measure the biodiversity of plants, animals, and insects that maintain ecological balance” (Shiva Reference Shiva2016, p. 46). In technical terms, AI-driven decision systems reinforce such classifications when they process large-scale datasets that omit each element’s symbolic variability. An algorithm that calculates a crop’s “value” solely by market price or yield volume will never account for the possibility that a particular bird or plant, one spiritually significant to a local community, can foster communal cohesion or play a crucial role in soil regeneration.

6. Conclusion: toward a rhizomatic hylomorphic approach in the age of AI

As we have illustrated, Linnaean taxonomy’s rigid hierarchical structures have fundamentally shaped modern data architectures, including AI systems, thus reinforcing power dynamics that marginalize alternative knowledge frameworks and fail to capture the richness of human and non-human relationships. The philosophical tragedy here, as we have already stated, is that modern computational systems have embraced “ontology” while seemingly abandoning critical engagement with “epistemology.” By defining what exists (ontos) rather than examining knowledge (epistēmē), these systems naturalize particular knowledge frameworks as truth.

In stark contrast, the complex, multidimensional taxonomies developed by Indigenous communities – such as the Kakataibo’s intricate animal classifications or the K’iche’ Maya’s sacred maize cosmology – encode ecological and cultural significance through fluid, relational dimensions of knowledge. Embracing these alternatives is not merely a technical adjustment but a profound epistemological shift that challenges colonial knowledge structures, questioning whether AI systems trained on Western (one-world) epistemological foundations can truly “know” the world beyond statistical pattern recognition or can engage with Indigenous ways of knowing that embrace animate relationships with non-human entities and embodied knowledge. As Aníbal Quijano (Reference Quijano2007) argues, decolonization demands dismantling the Western universalist paradigm to foster intercultural dialogues rooted in “alternative rationalities and respect for difference.”

What we propose is a multidimensional model for AI – one that transcends the rigid “tree” of Linnaean logic and is capable of recognizing multiple, equally valid ontological frameworks rather than imposing a single reality model derived from colonial taxonomic traditions, embracing a fresh perspective to what Deleuze and Guattari term a “rhizomatic framework” (Deleuze and Guattari Reference Deleuze and Guattari1980). Though emerging from Western philosophical traditions, Deleuze and Guattari’s concept of the rhizome offers a powerful framework for reconceptualizing our approach to knowledge systems. Unlike the Linnaean hierarchical taxonomies that impose fixed categories that dominate Western thinking, the rhizome operates as a non-hierarchical and decentralized network where entities (plants, rituals, ecosystems) exist in dynamic, overlapping dimensions. For instance, federated systems, combined with liquid ontologies (e.g., Mexico’s SCB or Brazil’s SOEB), allow metadata to shift contextually – such as maize being simultaneously “crop,” “deity kin,” and “soil regenerator” across overlapping cultural schemas.

Archives and memory institutions must evolve beyond static repositories into living networks. Inspired by Indigenous relationality, these institutions could adopt graph-based architectures (e.g., Neo4j) to map connections between entities, rituals, and ecological cycles. For example, the aforementioned Waman Wasi Calendar Project in Peru demonstrates how temporal and spatial relationships – encoded in matrices of months, seasons, and activities – can be visualized as dynamic, interconnected nodes rather than fixed categories. Here, federated learning becomes a tool for epistemic justice, enabling communities to contribute context-rich data without surrendering control to corporate or state infrastructures.

Critical imagination must also confront the agency of non-human actors. Indigenous taxonomies reveal that plants and animals are not passive objects but participants in what Deleuze and Guattari call assemblages – complex networks of relationships where meaning emerges through interaction (Deleuze and Guattari Reference Deleuze and Guattari1980). A bird arriving at dawn, for instance, might signal seasonal changes, embody spiritual significance, or influence agricultural practices. AI systems designed to recognize these assemblages could move beyond binary labels (“pest” or “pollinator”) to map relational contexts, such as a plant’s role in soil regeneration or communal rituals. This requires algorithms that prioritize ambiguity and multiplicity, embracing contradictions as inherent to ecological and cultural systems.

Aristotle’s Hylomorphism provides a lens to reinterpret this dynamism. However, technical systems often reduce form to static categories (e.g., labeling maize solely as Zea mays). By contrast, a “rhizomatic hylomorphism” represents a philosophical intervention which recognizes matter and form as fluid, co-constitutive processes. For AI, this means designing models that do not classify but instead cartographize – generating living maps where a plant’s “vitality” is defined by its relationships (medicinal, symbolic, ecological) rather than predetermined labels. Federated learning and graph databases could operationalize this by allowing communities to layer contextual metadata (e.g., geotagged ceremonial uses) onto shared ontologies.

The challenge lies in resisting industrial AI’s obsession with standardization. While digitalization thrives on striated space – compartmentalizing data into grids for efficiency – Indigenous systems flourish in smooth spaces of ambiguity and multiplicity. Bridging this divide demands a political shift to value epistemic pluriversality over capitalist extraction. As Arturo Escobar (Reference Escobar2018) explains, this is a struggle for ontological sovereignty, where federated architectures and Indigenous-led design become tools of resistance against data colonialism and moreover would give local communities the right to determine not just what they know, but how they know.

This approach favors Indigenous epistemologies that emphasize relationship, context, and multiplicity over fixed categorization. In this vision, archives are not static repositories but gardens – spaces where multiple knowledges take root, hybridize, and adapt. Just as the Waman Wasi’s Calendar Project nurtures cyclical relationships between forests and communities, AI must evolve from a tool of extraction to a medium of reciprocal care. Only then can technology honor what Gregory Bateson (Reference Bateson2000) termed the “ecology of mind” – the understanding that intelligence resides not in isolated algorithms but in the vibrant, entangled web of life itself.