AI-assisted knowledge generation is the production of knowledge candidates by an AI system, typically a large language model: draft concepts, classification proposals, webs of relations, rough drafts of definitions. The concept carries the decisive distinction between generation and validity: the Münchhausen trilemma does not bite at the point of producing but at the question of validity. A candidate is not yet knowledge — it becomes knowledge only when it passes examination by independent authorities.
What the system can do
Empirical research on machine knowledge modelling draws a consistent picture: language models generate relations between concepts with high precision, while their definitions turn out weaker than human ones — so the DRAGON-AI study (Toro et al. 2024), whose authors emphasize that subject-matter experts must curate the process; the NeOn-GPT pipeline (Fathallah et al. 2024), which has machine-generated knowledge models checked externally in a feedback loop, reaches the same conclusion. A side finding by Toro et al. is remarkable: the higher the professional competence of the examining person, the more readily she detects errors in machine-generated definitions — empirical evidence that the adequacy assessment requires expertise and cannot be automated away.
The passage to validity
Between candidate and valid knowledge lies examination against at least one verification authority: formal (machine checking of consistency and structure), personal (the adequacy judgment of the expert), and, where applicable, normative (the posited norm of the field). Whoever shortcuts this passage and lets the generating system itself decide on validity practises AI self-validation — and falls prey to the circle.
Assessment in terms of personhood ontology
Generation is not a personal act of cognition — it is categorially (disjointly) distinct from cognition. Cognition is the act in which the spirit touches reality and grasps a being as it is; precisely this touch — the correspondence with the thing — is what the generating system cannot enact by itself. It operates over signs whose meaning has been conferred on it by persons (derived intentionality). This is a determination of place, not a devaluation: as a generator of candidates the system is a legitimate, often highly productive tool — but the validity of its results is established elsewhere, in the personal judgment.
Counter-argument
A functionalist position holds the separation to be dogmatic: if a system models the connections of the world internally in such a way that its predictions reliably come true, that is a form of cognition — refusing the concept is mere definitional politics. The personalist reply: reliable modelling is the space of operation, not the space of understanding. That the results hold true for us does not make them cognitions of the system — the signs mean nothing to it, and the correspondence with reality is established not by it but by the examining person. Pragmatic success, moreover, is itself only an external signal — and a weak one: a model can “work” for a long time and nevertheless be wrong; without an adequacy assessment, nobody knows.
Ontological classification: distinct from: cognition (a personal act); related: AI self-validation (the impermissible shortcut), verification authority (the path to validity); produced by: artificial intelligence.
Sources: Generated by querying the Personhood ontology. Research as of 14 July 2026 (research report The Münchhausen Problem in LLM-Assisted Ontology Engineering).
Further sources:
- Toro, Sabrina et al. (2024): Dynamic Retrieval Augmented Generation of Ontologies using Artificial Intelligence (DRAGON-AI). Journal of Biomedical Semantics 15, Art. 19.
- Fathallah, Nadeen; Das, Arunav; De Giorgis, Stefano; Poltronieri, Andrea; Haase, Peter; Kovriguina, Liubov (2024): NeOn-GPT: A Large Language Model-Powered Pipeline for Ontology Learning. The Semantic Web: ESWC 2024 Satellite Events, Hersonissos, May 2024. Cham: Springer.
- Huang, Jie et al. (2024): Large Language Models Cannot Self-Correct Reasoning Yet. ICLR 2024.
- Searle, John R. (1980): Minds, Brains, and Programs. Behavioral and Brain Sciences 3, pp. 417–457.
- Spaemann, Robert (1996): Personen. Versuche über den Unterschied zwischen „etwas” und „jemand”. Stuttgart: Klett-Cotta. (English: Persons: The Difference between ‘Someone’ and ‘Something’. Oxford: Oxford University Press 2006.)
See also
- Münchhausen Trilemma
- AI Self-Validation
- Verification Authority
- Cognition
- Judgment
- Truth
- Artificial Intelligence
- Derived Intentionality
Generated by querying the Personhood ontology.