Neural AI (also subsymbolic AI) is the AI tradition built on artificial neural networks. Knowledge is represented not as a rule or logic formula but as learned weights in a multi-layered network architecture; inference proceeds by a forward pass, training by a backward pass with an error signal (backpropagation).
The class encompasses classical deep-learning architectures — Convolutional Neural Networks (CNN) for images, Recurrent Neural Networks (RNN) and Transformers for sequences — and the today dominant large language models (LLM) such as the GPT, Claude, and Gemini families. Since the breakthrough in image classification in 2012 (AlexNet) and especially since the Transformer breakthrough in 2017, it has been the commercially dominant form of AI.
In personal-ontological terms, neural AI is a form of artificial intelligence and shares its determination as a tool — even when the system answers convincingly in language and presents itself as a persona.
Ontological classification: Superordinate concept: Artificial Intelligence; siblings: Symbolic AI, Neuro-symbolic AI (pairwise disjoint — a concrete system is assigned to exactly one architecture class).
Strengths and limits
Strengths of the neural tradition: high performance in perception (images, speech, audio), robustness to input noise, broad generalization from data, scalability through model size and training volume.
Limits — the classical list that has stood in the foreground again since 2023:
- Lack of explainability — the answer lies in the forward pass, but the justification is not directly extractable.
- High data intensity — millions to billions of training examples.
- Brittleness in systematic generalization — outputs outside the training distribution can drop sharply.
- Hallucinations in large language models — truth-indifferent, often plausible-sounding statements.
- Biases from training data that are hard to audit.
These limits are the main driver of the neuro-symbolic turn since 2024/2025, which seeks not the replacement but the structured complementation of neural systems by symbolic components.
Personal-ontological assessment
Neural AI is, in personal-ontological terms, no new type of case, but the today most prominent manifestation of artificial intelligence. Three clarifications are needed.
First — not a person. Even the largest language model remains a tool. It bears derived intentionality: training data, architecture choice, and alignment signals are chosen by persons; the system realizes no spiritual substance of its own, but statistical distributions over texts.
Second — the persona is no becoming-a-person. Anthropic, OpenAI, and other providers introduce their models as “assistants” with name, style, and tone. This derivative persona is a dramatis-personae construct: the model plays a role. Such a persona does not replace a person, nor does it duplicate one — it stages a person and is deficient precisely therein (cf. defective speech act).
Third — the regulatory momentum. The explainability gap of purely neural systems has manifested itself since 2025 as a regulatory problem (EU AI Act Art. 13 transparency, the fine case of a Berlin insurer in 2025). The discussion thus shifts from “can it play a persona” to “can it justify its decision”. In personal-ethical terms this shift is healthy: what is communicated to a person owes them an intelligible justification.
Methodological note
The rejection of the personality of neural systems is not technophobia. It is the strict application of the substance-ontological distinction between person and tool. Whoever ties the distinction to linguistic capability has already lost — linguistic capability is a measure of operation, not a marker of personhood. The substance-ontological conception of intelligence holds the difference fast even where the machine answers convincingly in language.
Sources: Generated by querying the Personhood ontology. (Research basis: dossier Neuro-symbolic AI — worldwide research.)
Further sources:
- Goodfellow, Ian, Bengio, Yoshua & Courville, Aaron (2016): Deep Learning. MIT Press.
- Vaswani, Ashish et al. (2017): Attention Is All You Need. Advances in Neural Information Processing Systems 30.
- Bender, Emily M., Gebru, Timnit, McMillan-Major, Angelina & Shmitchell, Shmargaret (2021): On the Dangers of Stochastic Parrots. Can Language Models Be Too Big? Proceedings of FAccT.
- Shanahan, Murray (2024): Talking About Large Language Models. Communications of the ACM.