Large language models are not about natural language.
Clinical Snapshot
PICO Framework
| P — Population | Large Language Models (LLMs) as computational systems, compared conceptually to the human language faculty |
| I — Intervention | Large Language Models (probabilistic, data-driven string-processing systems) |
| C — Comparator | Human natural language, underpinned by an internal generative computational system (Universal Grammar / I-language framework) |
| O — Outcomes | Theoretical adequacy of LLMs as models of natural language; linguistic validity; capacity to distinguish real from impossible languages; data efficiency relative to human language acquisition |
Bottom Line
This theoretical commentary by Bolhuis and colleagues argues that Large Language Models (LLMs) are fundamentally mischaracterised when described as models of natural language. Drawing on Chomskyan generative linguistics, the authors contend that human language is an internal, recursive, hierarchical computational system that develops with minimal external input — properties that probabilistic, data-hungry LLMs do not replicate. The argument is conceptually coherent within the generative tradition and raises legitimate concerns about the misapplication of LLMs in linguistic science. However, clinicians and researchers should note that this is a theoretical opinion piece with no empirical data, no systematic review, and no engagement with the substantial contrary literature on LLM structural sensitivity. The absolute claim that LLMs are 'useless for linguistics' is contested and likely overstated. For Australian clinicians in speech pathology, neuropsychology, and health informatics, the paper serves as a useful theoretical caution against uncritical adoption of LLM outputs as valid proxies for linguistic competence. It does not, however, constitute sufficient evidence to guide clinical policy. Empirical validation studies remain essential before LLM-based language tools are adopted or rejected in clinical practice.
Key Findings
P Value: Not applicable — no statistical testing conducted
Effect Size: Not applicable — no empirical effect size reported
Primary Outcome: Theoretical argument that LLMs, as probabilistic models of externalized word strings (E-language), are fundamentally inadequate as models of natural human language (I-language), which is characterised by recursive, hierarchical, mind-internal generative computation
Nnt Or Sensitivity: Not applicable — theoretical commentary; no diagnostic, therapeutic, or prognostic metrics reported
Confidence Interval: Not applicable — no quantitative analysis performed
Clinical Application
The theoretical argument, if accepted, would counsel caution against using LLM outputs as proxies for genuine linguistic competence in clinical assessment. However, the paper does not provide implementation guidance, clinical thresholds, or practical recommendations for clinicians. Feasibility of applying these theoretical conclusions to clinical practice requires further empirical translation. Australia has an active and growing interest in LLM applications across healthcare, including the use of natural language processing tools in electronic health records, clinical decision support, and allied health (particularly speech pathology). The RACGP and Speech Pathology Australia have not yet issued formal guidance on LLM use in clinical linguistics. The TGA's evolving Software as a Medical Device (SaMD) framework is relevant if LLMs are deployed in diagnostic or therapeutic roles. The presence of a Macquarie University co-author (Crain) gives the paper direct Australian academic relevance. Australian speech pathologists and neuropsychologists should note the theoretical caution raised here when evaluating LLM-based language assessment tools, pending robust empirical validation studies. Clinicians, researchers, and health informaticists considering LLMs for applications in speech pathology, neuropsychological assessment, aphasia rehabilitation, clinical documentation, and neurolinguistic research
Abstract
Large Language Models are useless for linguistics, as they are probabilistic models that require a vast amount of data to analyze externalized strings of words. In contrast, human language is underpinned by a mind-internal computational system that recursively generates hierarchical thought structures. The language system grows with minimal external input and can readily distinguish between real language and impossible languages.
References
- 1.Bolhuis, J. J., Moro, A., Crain, S., & Fong, S. (2026). Large language models are not about natural language. Behavioral and Brain Sciences. https://doi.org/10.1017/S0140525X26104622
Related Research
PloS one
Pronunciation assessment in foreign language learning: Reliability and scoring bias in human-generative AI evaluation
31 July 2026
Biomedical physics & engineering express
Intelligent integrated system for dysphagia rehabilitation and assessment
27 June 2026
Journal of medical Internet research
Explainable and Interpretable AI for Voice and Speech Analysis in Clinical Care: Systematic Review
25 June 2026
This content is for educational purposes for healthcare professionals only and does not constitute clinical advice. Clinical decisions should be based on individual patient assessment, current guidelines, and appropriate specialist consultation. Editorial Standards · Privacy Policy · Terms of Service