Large language models have learned to use language.
Clinical Snapshot
PICO Framework
| P — Population | Large language models (LLMs) as subjects of linguistic and cognitive scientific inquiry; implicitly, human language users as the comparative reference class |
| I — Intervention | Conceptual reframing: acknowledging that LLMs have genuinely 'learned to use language' as a theoretical and empirical proposition |
| C — Comparator | Traditional frameworks for evaluating language knowledge (e.g., Turing Test paradigms, nativist/innatist accounts, classical psycholinguistic benchmarks) |
| O — Outcomes | Theoretical and empirical advances in language science; revised criteria for assessing linguistic competence in artificial and biological systems; implications for post-Turing test era research methodology |
Bottom Line
This single-author theoretical commentary in Behavioral and Brain Sciences argues that large language models have genuinely learned to use language — not merely simulate it — and that acknowledging this should catalyse a paradigm shift in language science. The paper is intellectually provocative and engages meaningfully with foundational debates about linguistic competence, the Turing Test, and how we evaluate language knowledge in artificial systems. However, it presents no empirical data, no systematic evidence synthesis, and no operationalised definitions of its central constructs. By CEBM standards, it occupies the lowest tier of the evidence hierarchy (expert opinion). For clinicians in speech-language pathology, neuropsychology, or health informatics, the paper raises important conceptual questions about how LLM capabilities should be understood and evaluated, but provides no actionable clinical guidance. The absence of empirical grounding, engagement with LLM failure modes, and harm-benefit analysis limits its immediate clinical utility. It is best read as a stimulus for research agenda-setting rather than a basis for practice change. Australian clinicians should await empirical studies with validated outcome measures before drawing clinical inferences from this theoretical position.
Key Findings
P Value: Not applicable — no statistical testing performed
Effect Size: Not applicable — no empirical effect sizes reported
Primary Outcome: Theoretical proposition that LLMs have genuinely learned to use language, warranting a fundamental reconceptualisation of language science methodology and evaluation criteria
Nnt Or Sensitivity: Not applicable — theoretical commentary; no diagnostic, therapeutic, or prognostic metrics calculable
Confidence Interval: Not applicable — no quantitative analyses conducted
Clinical Application
The theoretical reframing proposed has no direct clinical implementation pathway without supporting empirical validation. Clinicians cannot act on a philosophical repositioning of LLM language competence without evidence that this translates to reliable, safe, and equitable clinical performance. Significant infrastructure, regulatory, and training barriers exist. In Australia, speech-language pathology and neuropsychological assessment of language disorders is governed by Speech Pathology Australia and AHPRA-registered psychologists. The TGA does not currently regulate most LLM-based language tools as therapeutic devices unless they meet the definition of a medical device under the Therapeutic Goods Act 1989. The RACGP has no specific guidelines on LLM use in language assessment. Medicare does not fund LLM-assisted language therapy. Australian clinicians should treat this paper as a theoretical provocation rather than practice-changing evidence, and await empirical validation before modifying assessment or therapeutic approaches. Indirectly relevant to clinicians working with language disorders (speech-language pathology, neuropsychology, psychiatry), clinical educators, and health informaticists considering LLM integration into clinical communication assessment or therapy support tools
Abstract
Acknowledging that large language models have learned to use language can open doors to breakthrough language science. Achieving these breakthroughs may require abandoning some long-held ideas about how language knowledge is evaluated and reckoning with the difficult fact that we have entered a post-Turing test era.
References
- 1.Lupyan, G. (2026). Large language models have learned to use language. Behavioral and Brain Sciences. https://doi.org/10.1017/S0140525X26104804
Related Research
Neuropsychologia
Reading between the lines: Combining pause dynamics and semantic coherence for automated assessment of thought disorder.
29 July 2026
Science advances
Curiosity-driven development of action and language in robots through self-exploration
25 July 2026
Behavior research methods
The English spaced compound word database
12 July 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