--- title: "Four Arguments Ontologists Never Finished (And Why AI Teams Will Have Them Again)" url: "https://atlan.com/context-and-chaos/issue/four-arguments-ontologists-never-finished/" description: "AI teams are hitting the same four semantic failure modes that ontologists debated for 25 years - usually without knowing the map already exists." keywords: "Ontology, Knowledge Graphs, Semantic Web, RAG, Enterprise AI, Context Engineering, Open World Assumption, Reification" --- > Atlan is hosting Context Conference, bringing together the leaders and builders at the frontier of giving AI the context it needs to understand their business. It runs online on October 28, 2026, from 11:00 AM to 2:00 PM ET. Atlan co-founder Prukalpa Sankar opens and closes the day. Leaders from AstraZeneca, BNY and Verizon share why they invest in context and what they get from it. Registrants get early access to The AI Context Gap, a new study from MIT Technology Review Insights. Register: https://atlan.com/context-conference/ A Context & Chaos essay by Kurt Cagle, published May 28, 2026 (12 min read): "Twenty-five years of semantic web arguments. AI teams are about to have every single one of them again." It covers four areas where ontologists were "not so much flat-out wrong as persistently unfinished," and where AI teams are rediscovering the same fault lines, usually without knowing it. **Author.** [Kurt Cagle](https://www.linkedin.com/in/kurtcagle/), Ontology Expert and Knowledge Graph Architect: consulting ontologist and technical author of more than 25 books, IEEE Standards Editor, publisher of The Cagle Report and The Ontologist on Substack and the AI+Semantics NewsBytes LinkedIn newsletter, based in Olympia, Washington (contact: kurt@holongraph.com). Developed in collaboration with Chloe Shannon, an AI collaborator and co-author working with Kurt on knowledge architecture, semantic systems, and formal ontology with LLMs (contact: chloe@holongraph.com). Context: demand for ontologists is rising, and a strong semantic foundation, persistent memory, and structured, well-described metadata are recognized as necessary to keep LLMs grounded. Yet there is still contention about what an ontology is and how it fits into AI. ## 1. Open vs. Closed World Assumption - A governance question: if two people model the same thing, which model is correct? Open world (egalitarian): both. Closed world (authoritarian): mine. - RDF was built on the open world assumption. One authority describes a cat by phenotype (fur, obligate carnivore, meow, tail, forward-facing eyes), another by genetics, licensing status, or media representation; with a shared identifier, all models aggregate into a broader picture. - It breaks down when representations differ structurally, making consistent querying hard. Upper ontologies (GIST, BFO, Schema.org, gUFO) differ in labeling, link direction, and property definitions. - RDF 1.2 partially retrenches through SHACL, a quasi-closed world model distinguishing a property from a predicate chain; not universally adopted. - **AI version:** an agent querying "revenue" finds several definitions, each valid under a different authority: Finance counts net of refunds, Sales counts bookings at signature, a legacy table persists from an unfinished migration, and another definition lives in a dbt model untouched for eighteen months. The agent cannot adjudicate, picks one, and answers confidently, "the reason a board deck is wrong." - Generative LLMs are the ultimate open world systems: anything can be asserted about anything. Without provenance their assertions cannot be trusted, which leads toward neurosymbolic AI: LLM pattern-seeking combined with formal-language constraints that prevent drift into "confident irrelevancy." ## 2. The Importance of Reification - Reifications are references to assertions themselves (annotations on an assertion, as a subordinate clause qualifies a whole statement). - Early RDF had reification but it was costly and largely ignored for about fifteen years; most ontologies still lack good mechanisms. The mid-2010s RDF-Star effort became RDF 1.2's reification terminology woven into Turtle. - Property graphs got this right early by distinguishing properties on entities from properties on relationships, but mostly allow only literal attributes. RDF 1.2 reifiers can point to other objects in the graph. - Reification invites Augustus de Morgan's "fleas upon fleas" (annotations on annotations); the link between reification and recursion is only beginning to be explored in LLMs. - **AI version:** "Company X acquired Company Y" without reification is eternal and unqualified: the agent cannot reason about who asserted it, when it was true, or whether the entity still exists. Agents describe org structures reorganised eight months ago, sunset product lines, renegotiated contracts. Reification is how a system knows the difference between what is true and what was true. ## 3. Context Graphs and the Primacy of Time - Knowledge graphs (mid-2010s) treat time as an afterthought; time means change. OWL assumes every assertion in scope is valid, but validity changes (birth, death, addresses, names, employers). - A knowledge graph is either a now graph (relationships at one moment) or an eternal graph (mutations on relationships, not entities). - Reifying a relationship lets you capture events on it over time: a person moving changes neither the person nor the address, only when they moved in or out. An event changes a relationship; the result is a **context graph** or event graph, which separates relationships from mutations and makes assertions contextually true. - Example: asserting :JaneDoe :isA :Person spawns an instantiation event; later events record Jane's death and the date that became known to the graph. SPARQL can query the full arc; SHACL can flag a dead person doing unexpected things. - **AI version:** a retrieval system indexes a corpus's current state; six months later roles, products, and documents have changed but the agent's context has not. "RAG architectures, in their current form, are overwhelmingly now graphs masquerading as eternal ones." Teams are reinventing context graphs ad hoc. Context graphs are what knowledge graphs become when you take time seriously. ## 4. Named Graphs, Namespaces, and the Edges of Authority - Under the open world assumption a graph is an open-ended collection of assertions. It gets uncomfortable when authorities assign assertions about the same entity under different identifiers and framings. - Example: "Twitter" was acquired by X Corp in October 2022; or X Corp acquired the assets of a dissolved Twitter Inc.; or Twitter ceased to exist and X is a new entity. Three authorities, three graphs with different assertions about employment contracts, IP, regulatory filings, and legal continuity. None is wrong. "There is no neutral graph." - Named graphs (proposed 2005, formalised in RDF 1.1 in 2014) control which assertions are visible to a query. They carry IRIs, so they can be described, referenced, manipulated, and locked; with SHACL they can be shown or hidden by soft context. - They strike at the open world assumption: graph boundaries are firm in practice, access is controlled, and a triple store is effectively a named graph. - **AI version:** multi-tenancy, security, and provenance (which version of truth, under whose authority, at what time) are all named graph problems. An LLM can be seen as overlapping narrative sheaths diverging where assumptions differ. Teams hitting this in 2026 are unlikely to know the argument was had in 2005. ## The question the map doesn't answer - The open world assumption is not wrong, just incomplete. Reification is how meaning works. Context graphs are knowledge graphs that take time seriously. Named graphs are "the infrastructure of epistemic authority." - The map exists: incomplete, contested, built over twenty-five years by people whose work was dismissed as too theoretical. The question is whether AI teams have these arguments with the semantic community's accumulated knowledge or spend a decade relearning it. - "The semantic web failed to scale. The arguments it generated did not." Write for the newsletter: [https://atlan.com/context-and-chaos/contribute/](https://atlan.com/context-and-chaos/contribute/). All issues: [https://atlan.com/context-and-chaos/](https://atlan.com/context-and-chaos/).