Researchers find that AI models express uncertainty in 84% of cases using neutrosophic logic, revealing limitations of traditional scalar representations across five major vendors.

Leyva-Vázquez and Smarandache's 2025 neutrosophic framework showed 'hyper-truth' (Truth+Indeterminacy+Falsity > 1.0) in 35% of LLM evaluations when dimensions are unconstrained

New study replicates findings across models from Anthropic, Meta, DeepSeek, Alibaba, and Mistral, finding hyper-truth in 84% of unconstrained evaluations

Scalar T/I/F representations cannot distinguish between different epistemic states—paradox, ignorance, and contingency all collapse to identical outputs when models adopt an 'Absorption' position (T=0, I=1, F=0)

Tensor-based approaches are proposed to recover the epistemic distinctions that neutrosophic scalars fail to express