The term 'ontology' performs two distinct roles across two disciplines. In philosophy, ontology asks 'What exists?': the most influential twentieth-century criterion for ontological commitment ties 'what exists' to the values assigned to variables in formal languages—a theory must assign variables to entities for it to be true (as summarized in the Stanford Encyclopedia of Philosophy entry 'Logic and Ontology'). Philosophy even debates a higher-order question called meta-ontology: What, exactly, is ontology trying to answer? Philosophical inventories have never been finalized—but engineering cannot wait. In information science, an ontology is an engineering deliverable: publishable, versionable, and machine-readable. In 1993, a paper from Stanford borrowed this term from philosophy and gave it an engineering definition—launching an ongoing, unresolved debate: What should the world’s top-level inventory look like, and who holds the authority to write it? A 2007 comparative study surveyed seven proposals; an international standard adopted one in 2021; today, large language models are reshaping the very nature of the problem.
## 1993: Engineers Borrowed the Term
That 1993 paper appeared in *Knowledge Acquisition*, Volume 5, Issue 2, pages 199–220. Its abstract opens with a clear purpose: to support the sharing and reuse of formalized knowledge among AI systems, a common vocabulary is needed. The definition follows: 'An ontology is a specification of a shared conceptualization—the vocabulary (definitions of classes, relations, functions, and other objects) used to represent a domain.' The introduction offers a shorter original phrasing: 'An ontology is an explicit specification of a conceptualization.' Here, 'conceptualization' refers to an abstract, simplified view of the world held for a particular representational purpose—a usage drawn from a 1987 textbook cited in the paper. The paper does not conceal the term’s philosophical origins: it states plainly, 'The term is borrowed from philosophy,' where ontology is the systematic account of 'being.' Its technical contribution is Ontolingua, a translation mechanism that converts predicate-logic definitions into various specialized representations, enabling ontologies to be portable.
This is where the meaning shifts: philosophical ontology accounts for 'being'; engineering ontology specifies 'vocabulary'—the object described moves from the world itself to conventions governing how we speak. The consequences of this drift only became visible later: if it is merely a convention, then inevitably a forum must emerge to decide which convention counts.
## The Top-Level Inventory: Seven Proposals
Upper ontologies (also called top-level ontologies) govern the most general categories: objects, processes, properties, and relations—the highest layer above all domain-specific ontologies. In 2007, three researchers from the University of Genoa and the Technical University of Valencia compared seven upper ontologies using standardized software-engineering criteria: BFO, Cyc, DOLCE, GFO, PROTON, SUMO, and a proposal named after its individual author. The paper candidly notes that all three authors come from computer science—and thus find engineering criteria more intuitive than philosophical ones.
The scale differences in the comparison table span several orders of magnitude: BFO contains 36 classes, linked solely by 'is_a' relations; DOLCE includes approximately 100 concepts and 100 axioms; SUMO comprises 20,000 terms and 60,000 axioms (including those from domain ontologies), and the paper calls it one of the largest publicly available formal ontologies at the time; Cyc, a commercial project, contains roughly 300,000 concepts, 3,000,000 assertions, and 15,000 relation types. Though bearing the same name, their scales differ as drastically as a pocket dictionary versus an encyclopedia.
Scale is superficial; stance is foundational. A specialized comparison cited in the 2007 study summarizes the divergence between DOLCE and BFO in one sentence: the former is fundamentally conceptualist in methodology, the latter fundamentally realist. As the 2007 study puts it, DOLCE aims to capture the ontological categories underlying natural language and human commonsense; maintained by the Applied Ontology Laboratory of ISTC-CNR, it debuted stably in 2002–2003, is written in first-order logic, and has undergone formal consistency proofs. BFO, by contrast, insists on being a 'small upper ontology', deliberately excluding terms from specific sciences such as physics, chemistry, and biology; the project launched in 2002 under funding from the Volkswagen Foundation’s 'Forms of Life' initiative. A telling counterpoint: the same comparative study notes substantial structural agreement between the two—both distinguish continuants from occurrents, and both adopt an eternalist stance—making methodological divergence the most fundamental divide. A third path is SUMO: launched in December 2000 by Teknowledge Corporation, it served as the starting document for the IEEE Standard Upper Ontology Working Group and included manual mappings of all WordNet 2.1 lemmas. The comparative paper envisions upper ontologies as a 'lingua franca' for integrating heterogeneous knowledge across software agents: even without exposing their full ontologies, two systems could align via a shared top level. Over thirty years later, this vision has neither fully failed nor fully succeeded—the next act unfolds in an unexpected venue.
## Standards Bodies Enter the Scene
In 2021, something unprecedented in the history of philosophy occurred: an upper ontology became an international standard. According to publicly available documentation, ISO/IEC 21838-2:2021 bears the title *Information technology — Top-level ontologies (TLO) — Part 2: Basic Formal Ontology (BFO)*. Part 1 specifies the requirements an ontology must meet to qualify as a 'top-level ontology'; Part 2 formally adopts BFO as a standard top-level ontology.
The timeline, per the Wikipedia entry: BFO 1.0 was released in 2001, version 2.0 in 2015, BFO 2020 in 2020, and standardization in 2021. The official README, however, states initiation in 2002—discrepancies exist and are recorded as-is. Scale figures also diverge: the README (a 2.0-era document) reports over 130 projects using BFO; the Wikipedia entry cites over 650 projects built upon it, predominantly in biomedicine. Standards determine 'which one to use', not 'which one is correct'.
## What Success Looks Like
Truly enduring ontologies do not reside at the pyramid’s apex but thrive within domains. In May 2000, *Nature Genetics* published a paper authored by the 'Gene Ontology Consortium': genome sequencing revealed that genes specifying core biological functions are widely shared across all eukaryotes, and knowledge about one organism often transfers to another. The Consortium aimed to provide a shared vocabulary for such transfer: a dynamic, controlled vocabulary updated alongside knowledge accumulation—specifically, three independent ontologies covering biological processes, molecular functions, and cellular components.
Seven years later, *Nature Biotechnology* published a paper by the OBO Foundry, opening with an awkward truth: the very success of this approach triggered an explosion of ontologies—and that proliferation itself became a new barrier to data integration. The response was not to build a larger ontology, but to establish governance rules: existing ontologies (including the Gene Ontology) underwent harmonization reforms; new ontologies were to be developed collaboratively and continuously, aiming for interoperability and logical well-formedness across an ontology family.
From these two papers emerges an observation (this article’s judgment): winning solutions are rarely the most philosophically elegant—they are those that solve problems of division of labor and maintenance. The Gene Ontology solved 'each database inventing its own terms'; OBO solved 'multiple ontologies talking past each other'—neither required answering philosophical questions.
## Large Models Arrive—The Problem Transforms
Starting in 2023, ontologies acquired a new variable: large language models. A paper accepted at ISWC 2023 tasked nine model families with three ontology learning tasks—term typing, taxonomy discovery, and non-taxonomic relation extraction—in zero-shot settings, using materials from WordNet, GeoNames, and UMLS. It asked: Can machines learn inventories directly from text?
Two 2024 papers pushed further. One pursued end-to-end generation: fine-tuned models produced entire ontology taxonomies directly—not decomposed into subtasks—with a method named OLLM; evaluation metrics were newly invented, using deep learning techniques to measure semantic and structural similarity against reference answers. Another probed whether models truly learn domain knowledge: using parallel corpora of normal English versus nonsense words, it found that off-the-shelf models do not consistently reason about semantic relations, relying instead heavily on word meanings and their framing.
A second route uses ontologies in reverse. A 2023 extraction method called SPIRES takes user-defined knowledge schemas and prompts models to recursively extract matches, assigning identifiers from existing ontologies and vocabularies to all matched elements. Its abstract reports accuracy comparable to mid-tier existing relation extraction methods, with the selling point being zero training data for novel tasks. Here, ontologies no longer constitute knowledge to be learned—they serve as output-constraining templates: every statement the model makes must map onto a publicly available inventory.
OntoLearner, posted in July 2026, scales up evaluation: claiming to be the first unified framework of its kind (under review at *Nature Communications*, status recorded verbatim), it releases 180 machine-readable ontologies across 22 domains and evaluates 22 retrieval models and 12 large language models. Its conclusion points to structure: failure patterns grow with ontology complexity—not with model size or architectural sophistication—indicating the primary bottleneck lies in a structural mismatch between how models encode knowledge and how ontologies organize it.
At the close of the last century, the question was 'Which inventory is correct?'; today, it has become 'How do we use inventories to govern models?'
## Three Judgments
First, the upper-ontology dispute is not academic chatter (this article’s synthesis). ISO’s adoption of BFO as Part 2 of a multipart standard confirms that the question of who holds authority over top-level categories is real—and has moved from journals into standards bodies. Stakeholders hold divergent conceptions of 'correctness': conceptualists aim to describe how we talk about the world; realists aim to describe the world itself; inclusivists bet on comprehensiveness. The disagreement reflects differing stances—not a binary right-or-wrong. The standard’s value lies not in ending debate, but in providing engineering practice with a default choice that is revocable, replaceable, and contestable.
Second, the shift—from 'an inventory of what the world is' to 'an agreed-upon specification of conceptualization'—is not degradation, but a condition of survival (this article’s judgment). The 1993 paper explicitly states the term is borrowed from philosophy: engineering demands deliverables that are versionable and portable; philosophy requires an eternally unfinished question. Only through semantic drift can the term coexist in both disciplines.
Third, large models have not rendered ontologies obsolete—they have shifted their value center (this article’s judgment). In extraction pipelines, ontologies supply identifiers and constrain output shape (SPIRES); in end-to-end generation, evaluation demands newly engineered semantic and structural metrics (OLLM); large-scale evaluation locates bottlenecks in structural mismatches—not model scale (OntoLearner, under review). Value is shifting from 'teaching machines to understand the world' toward 'ensuring machine outputs are traceable to authoritative sources.' Evidence remains thin: most cited material is at the abstract level; one source remains unreviewed—hence stated here as a conjecture pending verification.
## What Different Readers Can Take Away
For knowledge graph and data platform practitioners: when selecting a top-level framework, first ask which philosophical stance it adopts, then assess ecosystem scale. The Gene Ontology and OBO experiences show that maintenance mechanisms and division-of-labor principles matter more for longevity than axiom elegance. An ontology that cannot answer 'Who maintains it? Whose authority resolves conflicts?' will not endure—even if its categories are impeccably drawn.
For engineers building extraction and retrieval-augmented systems: the most robust use of ontologies is not expecting models to learn them, but using them to constrain outputs—schema-first, with identifiers serving as verifiable anchors. Zero-shot extraction requires no training data, but achieves only mid-tier accuracy relative to existing methods (as reported in the paper)—do not market it as replacing human review. Before evaluating generated ontologies, first examine how the evaluation metrics themselves were constructed.
For readers in philosophy and scientific research: the ISO event is a ready-made case study in meta-ontology—the question of who gets to decide 'what exists' now has an engineering answer, delivered not by philosophical argument but by practical need. Publicly documented application distributions suggest it grew organically from real-world data-integration demands.
For general readers: the next time you see AI categorizing things with apparent fluency, remember those categories stem from pre-drawn inventories. That 'pre-' spans over thirty years—and the topmost inventory remains unsettled. What you see is usually just one faction’s answer in an ongoing dispute.
Scope and Boundaries: The 1993 and 2007 papers were read in full; citations follow cached originals. Philosophical views rely on the Stanford Encyclopedia of Philosophy entries. The Gene Ontology and OBO papers, plus five arXiv papers, are cited at the abstract level only; content absent from abstracts (e.g., OBO’s specific principles list) is omitted; publication and peer-review status are recorded verbatim from abstract pages. ISO’s official website was accessed twice and denied; the standard number 21838-2 and its title rely on the Wikipedia entry’s secondary reporting—the entry’s claim of 'over 650 projects' appears alongside the README’s (2.0-era document) 'over 130' without reconciliation. SUMO’s official site was also denied access; its description draws solely on the 2007 comparative study. Two English quotations in the main text originate from the cached 1993 paper. No personal names appear in the text; papers are referenced by year and publication venue. 'Cyc' is recorded as a commercial product name per the comparison table, without constituting product evaluation. None of the paper-based sources include conflict-of-interest disclosures; no assertion is made regarding their presence or absence. Translation conventions: 'ontology' or 'ontological', 'upper ontology', 'conceptualization', 'term typing', 'taxonomy discovery', 'non-taxonomic relation extraction'—all appear with English originals at first occurrence.
Source: Knowledge Acquisition 5(2):199-220, 'A Translation Approach to Portable Ontology Specifications', 1993 (https://tomgruber.org/writing/ontolingua-kaj-1993.pdf); WOA 2007, 'A Comparison of Upper Ontologies'; DOLCE official page (ISTC-CNR Laboratory for Applied Ontology); BFO official README (documentation from the 2.0 era); Wikipedia entry on Basic Formal Ontology (secondary source); Nature Genetics 25(1):25-9, 'Gene ontology: tool for the unification of biology', May 2000, PMID 10802651; Nature Biotechnology 25(11):1251-5, 'The OBO Foundry: coordinated evolution of ontologies to support biomedical data integration', November 2007, PMID 17989687; arXiv 2307.16648, 'LLMs4OL: Large Language Models for Ontology Learning', July 31, 2023, accepted by ISWC 2023; arXiv 2410.23584, 'End-to-End Ontology Learning with Large Language Models', October 31, 2024; arXiv 2407.19998, 'Do LLMs Really Adapt to Domains? An Ontology Learning Perspective', July 29, 2024, accepted by ISWC 2024; arXiv 2304.02711, 'Structured prompt interrogation and recursive extraction of semantics (SPIRES)', April 5, 2023; arXiv 2607.01977, 'OntoLearner: A Modular Python Library for Ontology Learning with Large Language Models', July 2, 2026, under review at Nature Communications; Stanford Encyclopedia of Philosophy, 'Logic and Ontology', first published in 2004, substantially revised in 2023