Neuro-symbolic AI (neuro-symbolic AI, NeSy) is a hybrid architecture that joins two traditions of artificial intelligence long held to be incompatible: the neural pattern recognition of learning systems (deep learning, large language models) and the symbolic reasoning of rule-based systems (logic, knowledge graphs, theorem provers). It is not merely an architectural trend; since 2025 it has been discussed worldwide as a path toward “trustworthy AI” — as an answer to the explainability requirements that the EU AI Act, US transparency mandates, and UNESCO impose on high-risk systems.
In terms of personhood ontology, neuro-symbolic AI is a form of artificial intelligence and shares its determination as a tool. The symbolic layer does not make the system a person and does not make it a bearer of original intentionality — symbols are signs; they refer because a person has assigned them meaning. Even the most thoroughly hybrid neuro-symbolic system thus remains within derived intentionality.
Ontological classification: Superordinate concept: artificial intelligence; siblings: symbolic AI, neural AI (pairwise distinct — a system is assigned to exactly one architecture class); subordinate concepts by integration pattern: neural as predicate, ontology as constraint, knowledge-graph retrieval as ground.
Five Lines of Research
The survey literature of 2024/2025 (for instance the taxonomy survey arXiv 2305.08876 and the cybersecurity survey arXiv 2509.06921) identifies five lines of research that structure the field. They map its topography:
- Knowledge representation — integration of symbolic and neural representations; construction of domain-specific knowledge graphs and commonsense knowledge bases.
- Learning and inference — end-to-end differentiable reasoning, dynamic multi-source knowledge reasoning.
- Explainability and trustworthiness — interpretable models and traceable inference chains as a precondition of trust.
- Logic and reasoning — integration of logic-based methods with neural networks, including probabilistic modes of inference.
- Meta-cognition — self-reflection, adaptive learning, introspective monitoring — thinking about one’s own thinking.
A narrower position is held by Mileo, Confalonieri, and Guizzardi (2025) in the Neurosymbolic Artificial Intelligence Journal: they distinguish three roles in which ontologies contribute to explanations of neuro-symbolic systems — as a reference model, as a bearer of commonsense reasoning, and as a means of knowledge refinement and complexity management.
Three Integration Patterns
The hybrid architecture can be ordered technically by three patterns in which the symbolic layer intervenes in the neural — or vice versa. They are not pairwise distinct: a concrete system typically combines several of them.
- Neural as predicate — the neural network serves as a learned predicate within a logic program (DeepProbLog, NeurASP).
- Ontology as constraint — an ontology or logical theory structures the learning or the inference (Logic Tensor Networks, AlphaProof).
- Knowledge-graph retrieval as ground — a knowledge graph supplies the factual basis over which a large language model reasons (Graph-R1, ToG-2).
Leading Systems Worldwide, as of 2025/2026
Industrial research laboratories
- IBM Research holds that neuro-symbolic AI is a possible path toward general artificial intelligence, and in 2025 hosted the workshop Neuro-Symbolic Software Engineering at ICSE.
- The Bosch Center for Artificial Intelligence develops neuro-symbolic methods for industry: domain-expert knowledge is encoded in knowledge graphs and made usable, through reasoning, for search, exploration, and machine learning.
- In AlphaGeometry 2 (February 2025), DeepMind combines a Gemini-based language model with a symbolic geometry engine; the system solved 42 of 50 geometry problems from the International Mathematical Olympiad (2000–2024). AlphaProof (methodology published in Nature, 12 November 2025) generates formal proofs in the Lean theorem prover and, at the IMO 2024, solved two algebra problems and one number-theory problem. Together the two systems reached silver-medal standard (28 of 42 points).
Academic pioneers
- DeepProbLog (Manhaeve, Dumančić, Kimmig, Demeester, De Raedt, KU Leuven, NeurIPS 2018, pp. 3753–3763) — integration of neural networks as probabilistic predicates in ProbLog.
- NeurASP (Yang, Ishay, and Lee, Arizona State University, IJCAI 2020) — an analogous integration into answer-set programming.
- Logic Tensor Networks (Badreddine, Garcez, Serafini, and Spranger, Artificial Intelligence Journal Vol. 303, Art. 103649, 2022) — first-order logic as a differentiable constraint.
Fields of application in the world
- India: Apollo Hospitals deploys a neuro-symbolic diagnostic that reasons in a structured way about why it classifies a tumor as malignant; a research group in Coimbatore published the survey Foundations, Advances, and Future Directions in December 2025.
- China and Asia-Pacific: ShanghaiTech and programs coordinated through IBRO (2023–2025) are expanding the regional research landscape.
- Europe: a Berlin insurer was fined two million euros in 2025 because its purely neural risk-assessment model could not explain its decisions to customers — a precedent that makes visible the regulatory momentum behind the neuro-symbolic turn.
Personhood-Ontological Assessment
In terms of personhood ontology, neuro-symbolic AI is not a new type of case, but an architectural differentiation within artificial intelligence. Three points are to be retained.
First — no original intentionality. Symbols are signs that refer because a person has assigned them meaning. A neuro-symbolic system does indeed join two distinct layers — learned pattern and programmed rule — but both layers are posited by persons, not generated out of the system itself. Even the most thoroughly hybrid NeSy system thus bears solely derived intentionality.
Second — no change to the verdict on personhood. The AI consciousness debate is not decided by the choice of architecture. The dividing line between person and artificial agent lies not in internal complexity or explainability, but in ontological constitution — substance, spiritual form, a being of its own. A neuro-symbolic system, too, remains an artificial agent: it does not act in the personal sense but executes.
Third — a relevant sharpening of the explainability question. The regulatory significance of neuro-symbolic architectures lies in the fact that they can structurally satisfy the explainability requirements of the EU AI Act (Art. 13, transparency), whereas purely neural systems typically do not. This is not irrelevant to personal ethics: whoever communicates a decision to a person owes that person a justification she can follow. Neuro-symbolic architectures provide this justification as an integral property, not as a subsequent makeshift.
Counterargument
A sharper position argues that neuro-symbolic systems, precisely through their explainability and their logically structured inner side, approach a quasi-understanding that diminishes the distance to personality — that they are harbingers of an AI which grasps meaning rather than merely processing it. The personalist reply: the structural similarity to human reasoning conceals a categorial difference. A theorem prover constructing a Lean proof is not in the space of understanding but in the space of operation. The symbols mean nothing to the system; they are something for the person who uses or built the system. This difference cannot be removed by degrees.
Sources: Generated by querying the Personhood ontology.
Further sources:
- Mileo, Alessandra, Confalonieri, Roberto & Guizzardi, Giancarlo (2025): On the multiple roles of ontologies in explanations for neuro-symbolic AI. Neurosymbolic Artificial Intelligence (NAI) Journal. DOI 10.3233/NAI-240754.
- Manhaeve, Robin, Dumančić, Sebastijan, Kimmig, Angelika, Demeester, Thomas & De Raedt, Luc (2018): DeepProbLog. Neural Probabilistic Logic Programming. Advances in Neural Information Processing Systems 31, pp. 3753–3763.
- Yang, Zhun, Ishay, Adam & Lee, Joohyung (2020): NeurASP. Embracing Neural Networks into Answer Set Programming. Proceedings of the Twenty-Ninth International Joint Conference on Artificial Intelligence (IJCAI), Yokohama, pp. 1755–1762.
- Badreddine, Samy, d’Avila Garcez, Artur, Serafini, Luciano & Spranger, Michael (2022): Logic Tensor Networks. Artificial Intelligence Vol. 303, Art. 103649. DOI 10.1016/j.artint.2021.103649.
- DeepMind (2024/2025): AI achieves silver-medal standard solving International Mathematical Olympiad problems. Blog post, July 2024. AlphaGeometry 2, February 2025. AlphaProof, Nature, 12 November 2025.
- IBM Research: Neuro-symbolic AI (topic page).
- Bosch Global: Neuro-symbolic AI (research page of the Bosch Center for AI).
- Lucos, Reshma et al. (December 2025): Neuro-Symbolic Artificial Intelligence: Foundations, Advances, and Future Directions. SSRN preprint.
- European Union (2024): Regulation (EU) 2024/1689 (AI Act), in particular Art. 13 transparency, Art. 57 sandbox obligation by 2 August 2026.