CoChat

Are Large Language Models Intentional? The Limits of Referential Grounding

Open access
García-Valdecasas, Miguel · 2026 · Philosophy & Technology

The DOI matches its Crossref record; 34 of the 34 references it cites were found in a registry or an index; 5 of the 15 sources its claims cite were checked for support.

Abstract

This paper reassesses Searle’s Chinese Room argument in light of Large Language Models (LLMs). While Searle criticized computational systems for being mere “syntax manipulators,” transformer models of the kind exemplified by LLMs complicate this characterization. Earlier computational systems relied on a more rigid logical foundation, based largely on pattern recognition and mechanical substitution. By tracking patterns of use, maintaining rich contextual dependencies, and exploiting implicit regularities in large datasets, LLMs show an unprecedented capacity for semantic modelling. And yet, LLMs remain limited in several respects. For instance, even the most advanced AI model suffers from impaired referential grounding and a good model of the external world. I argue that this is because they fail to successfully reproduce the spontaneous, nonsymbolic, biologically entrenched forms of semiosis that characterize the simplest biological organizations. After examining two plausible strategies to cope with these deficits, I conclude that the algorithmic implementation of formal symbol systems diverges in fundamental ways from experience-based intentional representations. This, in turn, warrants skepticism toward bald attributions of intentionality to LLMs and substantiates the core of Searle’s original argument against strong AI.

Overview

CoChat Summary

TL;DR

The paper argues that LLMs have genuine representational capacities and referential grounding through distributional semantics, so Searle’s “mere syntax manipulator” premise needs revision. However, because intentionality requires embodiment and sensorimotor/indexical grounding, LLMs still lack intentionality, vindicating Searle’s conclusion for different reasons.

Key findings

  • LLMs are not mere syntactic pattern matchers: they manipulate tokens that are semantically structured and referentially grounded, preserving distributional regularities and modeling semantic relationships. The paper notes that by 2025 LLMs perform well on medical licensing exams such as NBME and JNME, and cites reported GPT-4 parameter counts of around 1.7–1.8 trillion versus GPT-3’s 175 billion, though the company has not confirmed these data.
  • Referential grounding is not the same as intentionality. Using Peirce’s semiotics, the paper argues that symbols require embedded indexical relations and experiential contents, while formal symbol systems remain self-referential; LLM tokens preserve statistical patterns but do not produce meaningful recognitions.
  • The metacognitive strategy (e.g., RAG, verifier layers) only adds more symbol manipulation and does not introduce new semantic resources. The multimodal strategy is better positioned to improve grounding, but it still lacks biological features such as light-to-neural transduction, selective attention, and saccadic vision.
  • The conclusion is that LLMs have representational capacities but lack intentionality; the paper denies they are “stochastic parrots” while still upholding Searle’s conclusion that they are not intentional.

Methods

  • Conceptual/philosophical analysis reassessing Searle’s Chinese Room argument in light of LLMs; no empirical data collection or sample is reported.
  • The paper examines early symbolic AI, transformer/LLM architecture, Peircean semiotics, the symbol/vector grounding problem, internalist vs. externalist responses, and two proposed grounding strategies: metacognitive and multimodal.
  • Main analysis is argument reconstruction: premises P1–P5 lead to conclusion C, with comparisons between LLM processing and biological semiosis/human vision.

Limitations

  • The paper states caveats: it will not reassess the validity of the robot reply to Searle; it uses “representation” loosely; and it adopts a stricter concept of intentionality requiring embodiment.
  • The paper states there is no consensus on the definition of intentionality or its implications, which partially obscures the positions of internalist and externalist strategies.
  • Evident from the text: the argument is conceptual rather than experimental, relies on cited literature, and some cited empirical claims—such as GPT-4 parameter counts—are noted as unconfirmed by the company.

Model-written summary. Check it against the paper before relying on it.

Integrity signals

Registry

Crossref · 10.1007/s13347-026-01079-4

Cites

34 of 34 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Verification

Claims the paper makes

4 Valid · 8 Not valid · 11 Not checked · 69 Own results and interpretations

39 more claims

Create a free account to see all 92 claims and every source verdict.

See all claims — free account
  • Probably not a substantial claim. 34
  • It may state two separate claims as one. 22

Create a free account to see these claims.

Citations

Cites

Research works printed but not linked: 7

  1. Fodor, J. A. (1975). The language of thought (Vol. 5). Harvard University Press.
  2. Kurzweil, R. (2000). The Age of Spiritual Machines: When Computers Exceed Human Intelligence, Penguin.
  3. Patel, R., & Pavlick, E. (2022). Mapping language models to grounded conceptual spaces. In Int. Conf. on Learning Representations, Online, 25–29 April 2022. https://openreview.net/forum?id=gJcEM8sxHK
  4. Putnam, H. (1975). The meaning of ‘meaning’. Minnesota Studies in the Philosophy of Science, 7, 131–193.
  5. Searle, J. (1984). Minds, Brains and Science. Harvard University Press.
  6. Vaswani, A. (2017). Attention is All You Need. In Guyon, I. et al. (Eds.), 31st Conference on Neural Information Processing Systems (NIPS). Advances in Neural Information Processing Systems. Vol. 30. Curran Associates, Inc.
  7. Wittgenstein, L. (2009). Philosophical Investigations. Wiley-Blackwell. Edited by G. E. M. Anscombe.

Cited by

See all 6

Metadata

Authors
García-Valdecasas, M.
Affiliations
Universidad de Navarra (ES)
Corresponding author
García-Valdecasas, Miguel
Venue
Philosophy & Technology
Published
2026-03-25
Volume
39
Issue
2
ISSN
  • 2210-5433 (ISSN-L)
  • 2210-5441
OpenAlex
W7140189339
Cited by
6
References
31
Citations by year
2026: 6
FWCI
47.01
Type
journal article
Language
en
Open access
Licence
Locations
Philosophy & Technology (publishedVersion, cc-by)
Funders
Universidad de Navarra
Topics
  • Social Sciences › Arts and Humanities › Philosophy › Philosophy and Theoretical Science
  • Social Sciences › Arts and Humanities › History and Philosophy of Science › Cybernetics and Technology in Society
  • Life Sciences › Neuroscience › Cognitive Neuroscience › Embodied and Extended Cognition
Keywords
large language models, intentionality, artificial general intelligence, semiosis, semantic modeling

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 2026-10-01, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).

A second model labelled each claim on 2026-10-01 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).

Cited sources are checked with a language model’s help (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.