Natural Syntax, Artificial Intelligence and Language Acquisition
Open accessThe DOI matches its Crossref record; 33 of the 33 references it cites were found in a registry or an index; 6 of the 60 sources its claims cite were checked for support.
Abstract
In recent work, various scholars have suggested that large language models can be construed as input-driven theories of language acquisition. In this paper, we propose a way to test this idea. As we will document, there is good reason to think that processing pressures override input at an early point in linguistic development, creating a temporary but sophisticated system of negation with no counterpart in caregiver speech. We go on to outline a (for now) thought experiment involving this phenomenon that could contribute to a deeper understanding both of human language and of the language models that seek to simulate it.
Overview
CoChat Summary
TL;DR
O'Grady argues that children's early preference for negative concord—a pattern absent from the standard English input they hear—reflects a processing-cost bias rather than input-driven learning, and proposes a thought experiment to test whether large language models can simulate this two-stage acquisition of negation. The paper positions "Natural Syntax" as a middle path between universal-grammar nativism and pure induction-based accounts of LLMs.
Key findings
- Three lines of evidence indicate an early child preference for negative concord even in standard English, which disallows it: Adam's spontaneous productions (e.g., "I didn't do nothing," file 63, age 3;5; "Because nobody didn't broke it," file 107, age 4;5), comprehension data, and cross-linguistic reports from Italian and Chinese.
- In the comprehension study cited, 15 of 20 children (aged 3 to 5) preferred the disallowed negative concord reading of test sentences, only 3 favored the double-negative reading, and 2 showed no preference; in contrast, only 2 of 15 adults favored the negative concord interpretation.
- Natural Syntax explains this as a processing-cost effect: negative concord requires no interaction between the two negatives (the pronoun keeps its default null-set denotation), whereas double negation requires an extra computation to yield the "something" interpretation.
- The proposed test asks whether a language model trained on developmentally plausible input reproduces the two-step trajectory (negative concord at ages 3–5, then correct double negation), using pretraining sets of 8–28 million words (age 4) and 24–84 million words (age 12), based on an estimate of 2–7 million words heard per year.
Methods
- Conceptual/theoretical paper: it reviews existing acquisition evidence on negation and develops the "Natural Syntax" framework, built on three principles—direct mapping, algorithmic orientation, and processing determinism.
- No new data were created; the empirical component is explicitly a thought experiment proposing future work with two model types (Type A, no processing-cost bias; Type B, with such a bias) trained on developmentally plausible corpora.
- Analysis is argumentative rather than statistical: the paper compares processing costs of negative concord versus double negation and derives predictions for model behavior.
Limitations
- Stated: the paper notes it leaves aside "the technical issue of precisely how a bias for processing cost should be incorporated into the language model," and the acknowledgements/funding sections state no external funding and no new data.
- Stated: the empirical test is a thought experiment only—"for now"—so no outcome is reported.
- Evident from the text: the child-language evidence is drawn from secondary reports (Bellugi's dissertation, Thornton et al., Moscati, Zhou et al.) rather than new experiments by the author, and the proposed pretraining word-count ranges are described as "a first approximation."
Model-written summary. Check it against the paper before relying on it.
Integrity signals
Registry
Cites
Statistics
Verification
Claims the paper makes
4 Valid · 4 Not valid · 35 Not checked · 17 Own results and interpretations
The cited sources support only part of this claim — read each source’s verdict below.
In the paper
“A striking feature of contemporary large language models (LLMs)-ChatGPT, GPT4, BERT and others-is their remarkable grammatical accuracy, which seems to surpass the abilities of even the most articulate of humans.”
§ Natural Syntax, Artificial Intelligence and Language Acquisition > 1. Introduction
What the sources say
- Partly supportedcites [1] Contreras Kallens, P.; Kristensen-McLachland, R.D.; Christiansen, M.H. Large language models demonstrate the potential of statistical learning in language. Cogn. Sci. 2023, 47, e13256. [CrossRef] [PubMed] sourceExperimental — accuracy not yet measured
Our copy strongly supports the grammatical-accuracy half of the claim: the passage at p. 3 of 6 states that 'the output of these models is, almost without exception, grammatical,' that on SubSimulatorGPT3 'one cannot help but notice that the content is grammatically correct,' and that with ChatGPT 'one is hard-pressed to find examples of ungrammatical sentences in English.' However, the comparative element attributed here — that this accuracy 'seems to surpass the abilities of even the most articulate of humans' — is not what the source says; it explicitly caps its claim at 'human-level grammatical language' / 'human-like grammatical language' and warns that 'it is easy to overhype LLMs,' noting limitations at the semantic and discourse levels. The named systems also differ: the paper discusses GPT-3, Gopher, OPT, BLOOM, ChatGPT and BERT, with no mention of GPT-4. (LLM, conf=0.78)
hedged- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Partly supported
Every source the paper cites for this claim was checked and supports it.
In the paper
“The key point that we want to make is that the output of [large language models] is, almost without exception, grammatical.”
§ Natural Syntax, Artificial Intelligence and Language Acquisition > 1. Introduction
What the sources say
- Supportedcites [1] Contreras Kallens, P.; Kristensen-McLachland, R.D.; Christiansen, M.H. Large language models demonstrate the potential of statistical learning in language. Cogn. Sci. 2023, 47, e13256. [CrossRef] [PubMed] source
Passage [1] (body text, p. 3 of 6) states almost verbatim: 'The key point that we want to make is that the output of these models is, almost without exception, grammatical,' with 'these models' being the transformer-based LLMs named in passage [3] (GPT-3, Gopher, OPT, BLOOM). The same point is reinforced elsewhere in our copy (passage [6]: 'the overwhelmingly grammatical language produced by LLMs'; passage [4] Abstract: 'human-like grammatical language'). The citing sentence is a direct quotation with the bracketed specification of the referent, fully supported. (LLM, conf=0.95)
plain- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Supported
The cited sources support only part of this claim — read each source’s verdict below.
In the paper
“. . . the avoidance [and minimization] of cognitive demand is a fundamental principle of cognition . . .[15] (p. 2), citing [16]”
§ Natural Syntax, Artificial Intelligence and Language Acquisition > Processing Determinism
What the sources say
- Supportedcites [15] Christie, S.T.; Schrater, P. Cognitive cost as dynamic allocation of energetic resources. Front. Neurosci. 2015, 9, 289. [CrossRef] [PubMed] source
The Introduction states that Kool et al. (2010) propose that avoidance of cognitive demand is a fundamental principle of cognition. It also describes a tendency to seek the least cognitively expensive way to achieve a goal, supporting the added wording about minimization. The cited paper reports this principle rather than presenting it as its own original finding, consistent with the citing sentence’s “citing [16].” (LLM, conf=0.91)
- Partly supportedcites [16] Kool, W.; McGuire, J.; Rosen, Z.; Botvinick, M. Decision making and the avoidance of cognitive demand. J. Exp. Psychol. Gen. 2010, 139, 665-682. [CrossRef] sourceExperimental — accuracy not yet measured
The Abstract reports a bias toward less cognitively demanding actions in six experiments, and the General Discussion says the results support a ‘law of least mental effort.’ The paper frames this as evidence about behavioral decision-making, not as establishing demand avoidance as a fundamental principle of cognition generally. Our copy provides section labels but no page numbers. (LLM, conf=0.83)
plain- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Rests on an assumption not shown here.
- Supported
None of the sources cited for this claim have been checked yet.
In the paper
“The mapping between form and meaning is regulated by algorithms that operate in real time in the course of speaking and understanding.”
§ Natural Syntax, Artificial Intelligence and Language Acquisition > Algorithmic Orientation
What the sources say
- Not checked yetcites [11] O'Grady, W. Natural Syntax: An Emergentist Primer, 3rd ed.; 2022; Available online: http://ling.hawaii.edu/william-ogrady (accessed on 12 August 2022).
- Could not verifycites [12] O'Grady, W. Working memory and natural syntax. In The Cambridge Handbook of Working Memory and Language; sourceExperimental — accuracy not yet measured
Our copy of the handbook holds the chapter 'Working memory and high-level text comprehension processes', and every passage retrieved (sections 2-5, pp. 6-20) concerns comprehension monitoring, updating information, predictive processing (N400/PNP) and cognitive-control accounts in L1/L2 bilinguals — e.g., the Baker (1985) evaluation/regulation distinction in §2 (p. 6) and the situation-model revision results in §5 (p. 20). No passage states or implies that form-meaning mapping is regulated by algorithms operating in real time during speaking and understanding; searches for 'mapping between form and meaning', 'algorithms', 'lexical retrieval form to meaning in production and comprehension', and 'real-time online processing speaking' returned only semantically adjacent text-comprehension material. The claim's point is therefore not addressed by our copy rather than contradicted by it. (LLM, conf=0.68)
plainstates a cause- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
In the paper
“[LLMs] are trained only on text prediction.”
§ Natural Syntax, Artificial Intelligence and Language Acquisition > 1. Introduction
What the sources say
- Not checked yetcites [2] Piantadosi, S. Modern Language Models Refute Chomsky's Approach to Language. Unpublished ms. Department of Psychology, University of California, Berkeley, CA, USA. 2023. Available online: http://colala.berkeley.edu/ (accessed on 1 May 2023).
plainstates a cause- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
25 more claims
Create a free account to see all 60 claims and to verify the 54 references still open.
- Probably not a substantial claim. 21
- It may state two separate claims as one. 14
Claims other papers make about this one
Each is checked against this paper’s own text: does it say what the citing paper says it does?
The source backs part of the claim, not all of it — or backs it for a different population, condition or comparison.
In the citing paper
“As non-autonomous systems, generative models are exclusively input-driven [101,102]. Their operations are triggered by user-specified prompts and they lack internal states, persistent memory, or goal-following mechanisms [103-105].”
From AI Agents vs. Agentic AI: A Conceptual taxonomy, applications and challenges (2026) · cites this paper as [102]
Experimental — accuracy not yet measured
Self-check
Percentages, sums and reported differences re-computed from the paper’s own numbers.
Each contradiction between two passages listed here was confirmed by an independent check.
Rated minor: it would not touch a result — a value no result rests on, a p-value that stays on the same side of the significance level, or a conflict that is clerical, resolved elsewhere in the paper, or outside its own findings.
In the paper
§ Natural Syntax, Artificial Intelligence and Language Acquisition > Negative-Concord Interpretation
“(A similar idea has been put forward by [25] (pp.352-354) in their study of negation in Mandarin.)”
§ Natural Syntax, Artificial Intelligence and Language Acquisition > References
“25. Moscati, V. Children (and some adults) overgeneralize negative concord: The case of fragment answers to negative questions in Italian. Univ. Pa. Work. Pap. Linguist. 2020, 26, 169-178.”
Citations
Cites
Efficiency and Complexity in Grammars
Hawkins, J. A. (2004) · doi:10.1093/acprof:oso/9780199252695.001.0001 · Cited by 1,743
The cognitive revolution: a historical perspective
Miller, G. A. (2003) · Trends in Cognitive Sciences · doi:10.1016/s1364-6613(03)00029-9 · Cited by 1,279
Decision making and the avoidance of cognitive demand
Kool, W., McGuire, J. T., Rosen, Z. B., & Botvinick, M. M. (2010) · Journal of Experimental Psychology: General · doi:10.1037/a0020198 · Cited by 1,141
Dual language exposure and early bilingual development
HOFF, E., CORE, C., PLACE, S., RUMICHE, R., SEÑOR, M., & PARRA, M. (2012) · Journal of Child Language · doi:10.1017/s0305000910000759 · Cited by 1,007
Chang, F., Dell, G. S., & Bock, K. (2006) · Psychological Review · doi:10.1037/0033-295x.113.2.234 · Cited by 984
Sentential negation and negative concord
Zeijlstra, H. (2004) · Cited by 763
A dissertation on natural phonology
Stampe, D. L. (1979) · Americanae (AECID Library) · Cited by 513
How Efficiency Shapes Human Language
Gibson, E., Futrell, R., Piantadosi, S. P., Dautriche, I., Mahowald, K., Bergen, L., & Levy, R. (2019) · Trends in Cognitive Sciences · doi:10.1016/j.tics.2019.02.003 · Cited by 485
Mapping the Early Language Environment Using All-Day Recordings and Automated Analysis
Gilkerson, J., Richards, J. A., Warren, S. F., Montgomery, J. K., Greenwood, C. R., Kimbrough Oller, D., Hansen, J. H. L., & Paul, T. D. (2017) · American Journal of Speech-Language Pathology · doi:10.1044/2016_ajslp-15-0169 · Cited by 472
Cross-Linguistic Variation and Efficiency
Hawkins, J. A. (2014) · doi:10.1093/acprof:oso/9780199664993.001.0001 · Cited by 405
Empirical re-assessment of stimulus poverty arguments
Legate, J. A., & Yang, C. D. (2002) · The Linguistic Review · doi:10.1515/tlir.19.1-2.151 · Cited by 280
O'Grady, W. (2005) · doi:10.4324/9781410612571 · Cited by 251
Self-citationFischer, O., Kemenade, A. V., Koopman, W., & Wurff, W. V. D. (2001) · Cambridge University Press eBooks · doi:10.1017/cbo9780511612312 · Cited by 226
How Can We Accelerate Progress Towards Human-like Linguistic Generalization?
Linzen, T. (2020) · Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics · doi:10.18653/v1/2020.acl-main.465 · Cited by 148
Two negations for the price of one
Thornton, R., Notley, A., Moscati, V., & Crain, S. (2016) · Glossa: a journal of general linguistics · doi:10.5334/gjgl.4 · Cited by 144
The opportunity cost of time modulates cognitive effort
Otto, A. R., & Daw, N. D. (2019) · Neuropsychologia · doi:10.1016/j.neuropsychologia.2018.05.006 · Cited by 138
Large language models demonstrate the potential of statistical learning in language
Contreras Kallens, P., Kristensen‐McLachlan, R. D., & Christiansen, M. H. (2023) · Cognitive Science · doi:10.1111/cogs.13256 · Cited by 130
The role of language processing in language acquisition
Phillips, C., & Ehrenhofer, L. (2015) · Linguistic Approaches to Bilingualism · doi:10.1075/lab.5.4.01phi · Cited by 123
Prediction in Second Language Processing and Learning
(2021) · Bilingual Processing and Acquisition · doi:10.1075/bpa.12 · Cited by 104
Cognitive cost as dynamic allocation of energetic resources
Christie, S. T., & Schrater, P. (2015) · Frontiers in Neuroscience · doi:10.3389/fnins.2015.00289 · Cited by 99
When Do You Need Billions of Words of Pretraining Data?
Zhang, Y., Warstadt, A., Li, X., & Bowman, S. R. (2021) · Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers) · doi:10.18653/v1/2021.acl-long.90 · Cited by 85
Findings of the BabyLM Challenge: Sample-Efficient Pretraining on Developmentally Plausible Corpora
Warstadt, A., Mueller, A., Choshen, L., Wilcox, E., Zhuang, C., Ciro, J., Mosquera, R., Paranjabe, B., Williams, A., Linzen, T., et al. (2023) · Proceedings of the BabyLM Challenge at the 27th Conference on Computational Natural Language Learning · doi:10.18653/v1/2023.conll-babylm.1 · Cited by 74
Children’s knowledge of double negative structures in Mandarin Chinese
Zhou, P., Crain, S., & Thornton, R. (2014) · Journal of East Asian Linguistics · doi:10.1007/s10831-013-9118-y · Cited by 68
Quantity doesn’t buy quality syntax with neural language models
van Schijndel, M., Mueller, A., & Linzen, T. (2019) · Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP) · doi:10.18653/v1/d19-1592 · Cited by 65
Hosseini, E. A., Schrimpf, M., Zhang, Y., Bowman, S., Zaslavsky, N., & Fedorenko, E. (2024) · Neurobiology of Language · doi:10.1162/nol_a_00137 · Cited by 63
(2022) · doi:10.1101/2022.10.04.510681 · Cited by 62
Modern language models refute Chomsky's approach to language
Piantadosi, S. T. (2024) · Zenodo (CERN European Organization for Nuclear Research) · doi:10.5281/zenodo.12665932 · Cited by 61
The Cambridge Handbook of Working Memory and Language
(2022) · doi:10.1017/9781108955638 · Cited by 43
Bioprograms and the innateness hypothesis
Bates, E. A. (1984) · Behavioral and Brain Sciences · doi:10.1017/s0140525x00073799 · Cited by 33
What is Natural in Natural Morphology (NM)?
Dressler, W. U. (1999) · Prague Linguistic Circle papers · doi:10.1075/plcp.3.11dre · Cited by 21
Algebraic Structures in Natural Language
Lappin, S., & Bernardy, J.-P. (2022) · doi:10.1201/9781003205388 · Cited by 21
Moscati, V. (2020) · Scholarly Commons (University of Pennsylvania) · Cited by 7
On the Syntax of English Variable Negative Concord
Robinson, M., & Thoms, G. (2021) · ScholarlyCommons (University of Pennsylvania) · doi:10.25405/ncl.34462194 · Cited by 6
Research works printed but not linked: 5
- Kaan, E., Grüter, T., Eds.; John Benjamins: Amsterdam, The Netherlands, 2021; pp. 1-24.
- Bellugi, U. The Acquisition of the System of Negation in Children's Speech. Ph.D. Thesis, Harvard University, Cambridge, MA, USA, 1967.
- Ambridge, B. A computational simulation of children's language acquisition. In Proceedings of the Third Conference on Language, Data and Knowledge (LDK 2021), Zaragoza, Spain, 1-3 September 2021; pp. 4:1-4:3.
- Miller, G. The Society Theory. In Artificial Intelligence: An MIT Perspective;
- MacWhinney, B., Ed.; Erlbazum: Mahwah, NJ, USA, 1999; pp. 1-27.
Cited by
AI Agents vs. Agentic AI: A Conceptual taxonomy, applications and challenges
Sapkota, R., Roumeliotis, K. I., & Karkee, M. (2026) · Information Fusion · doi:10.1016/j.inffus.2025.103599 · Cited by 671
“• Reactivity: As non-autonomous systems, generative models are exclusively input-driven [101,102].”
AI Agents vs. Agentic AI: A Conceptual Taxonomy, Applications and Challenges
Sapkota, R., Roumeliotis, K. I., & Karkee, M. (2025) · arXiv · doi:10.48550/arxiv.2505.10468 · Cited by 295
“• Reactivity: As non-autonomous systems, generative models are exclusively input-driven [101,102].”
AI Agents vs. Agentic AI: A Conceptual Taxonomy, Applications and Challenges
Sapkota, R., Roumeliotis, K. I., & Karkee, M. (2025) · SuperIntelligence - Robotics - Safety & Alignment · doi:10.70777/si.v2i3.15161 · Cited by 186
Generative AI as game changer: Implications for language education
Chapelle, C. A. (2025) · System · doi:10.1016/j.system.2025.103672 · Cited by 107
Open generative AI changes a lot, but not everything
Chapelle, C. A. (2024) · Modern Language Journal · doi:10.1111/modl.12927 · Cited by 37
Metadata
- Authors
- O’Grady, W., & Lee, M.
- Affiliations
- Department of Linguistics, University of Hawai‘i at Manoa, Honolulu, HI 96822, USA
- Hanyang University (KR)
- ORCID
- Lee, Miseon
- Venue
- Information
- Publisher
- MDPI AG
- Published
- 2023-07-20
- Volume
- 14
- Issue
- 7
- Pages
- 418
- ISSN
- 2078-2489 (ISSN-L)
- OpenAlex
- W4385069730
- Cited by
- 5
- References
- 32
- Citations by year
- 2025: 3 · 2024: 2
- FWCI
- 2.43
- Type
- journal article
- Language
- en
- Open access
- gold
- doi.org
- Licence
- Locations
- Information (publishedVersion, cc-by)
- DOAJ (DOAJ: Directory of Open Access Journals) (submittedVersion, cc-by-sa)
- Topics
- Social Sciences › Psychology › Developmental and Educational Psychology › Language Development and Disorders
- Social Sciences › Arts and Humanities › Language and Linguistics › Syntax, Semantics, Linguistic Variation
- Social Sciences › Psychology › Developmental and Educational Psychology › Child and Animal Learning Development
- Keywords
- language acquisition, large language models, negation, processing pressure, child-directed speech, artificial intelligence, language development
About these checks
Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.
The claim verdicts this report counts are experimental: their accuracy is not yet measured.
A language model drew the claims from the paper on 2026-10-05, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).
A second model labelled each claim on 2026-10-05 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).
Cited sources are checked with a language model’s help (qwen3.8-omni-flash and gpt-6-sol).
Statistics and arithmetic are recomputed, not judged.
Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.
What counts as a claim ›How we measure confidence ›How we check this ›
Are you an author, or do you see a problem with this page? Request a review
Data sources
Bibliographic data: Semantic Scholar, OpenAlex, Crossref.
Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).
Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.