Futrell and Mahowald (FM) argue that large language models (LLMs) have mastered the formal properties of human language. They point to work on the geometry of word embeddings and their ability to capture syntactic dependencies (Diego-Simón et al., Reference Diego-Simón, D’Ascoli, Chemla, Lakretz and King2024). However, dependency grammar is known to be not a viable candidate theory of natural language syntax, since it isolates word-word dependency graphs rather than hierarchical constituency structure (Marcolli, Chomsky, & Berwick, Reference Marcolli, Chomsky and Berwick2025) (dependency graphs are projections from this deeper algebraic structure). While FM (2025: 42) argue that recent work serves to “refute the claim that neural LMs can learn any language” (i.e., Kallini et al., Reference Kallini, Papadimitriou, Futrell, Mahowald and Potts2024), this work is confounded by the fact that Kallini and colleagues did not contrast count-based vs. constituency-based non-adjacency rules. Instead, they contrasted count-based non-adjacency with simple adjacency, which is not the appropriate resolution to isolate constituency-based rules that are not operationalized over linear distance (Hunter Reference Hunter2024). Other work casts doubt on the claim that LLM probabilities can distinguish between possible/impossible languages: Leivada et al. (Reference Leivada, Montero, Morosi, Moskvina, Serrano, Aguilar and Guenther2025a) show that LLM probabilities do not constitute reliable proxies for model-internal representations of syntactic knowledge, and Yang et al. (Reference Yang, Aoyama, Yao and Wilcox2025) show that perplexity scores from GPT-2-small do not distinguish between attested and unattested word orders.

Natural language syntax-semantics is not exhausted in formal and conceptual scope by distributional similarity or predictive probability. Many words are able to jointly host analog/continuous and digital representations (Jackendoff & Erk, Reference Jackendoff and Erk2025), alongside categorially incompatible types of inherent polysemy (Murphy, Reference Murphy2024a) (consider what kind of thing school is meant to be in “The progressive school next to the river was high in the NYT rankings and had just hired a new chancellor after being repainted”). In contrast to vector space semantic models operating within a single homogeneous metric space, natural language semantics is highly typed and heterogeneous: the cognitive architecture underlying it has to coordinate multiple representational formats (spatial structure, event structure, theory of mind, quantification, Boolean logic, causation, telicity, and evidentiality). These conceptual domains are not “dimensions” in the embedding sense. They host different mathematical structures with different neural substrates (Baggio, Reference Baggio2018; Murphy, Reference Murphy2025).

If FM are right that LLMs have learned non-trivial aspects of language, what have they learned when they affirm that “Eight eighteen resigned seventeen this eight twenty-one any ten warning” is a grammatical sentence (Leivada et al., Reference Leivada, Marcus, Günther and Murphy2025b)? The field of artificial intelligence has often leaned on metaphors to circumvent the black box nature of its agents (Mitchell, Reference Mitchell2024), with “learning” a clear example. The notion of emergence found in FM is another. Does any putative “mastery” of syntax entail that LLMs emulate the process of emergence as understood in biology (i.e., a behavior arising de novo as the (by-)product of a complex system)? Emergence is notoriously hard to define, but there is agreement that it entails novelty that is not predicted by lower-level generic components of the adaptive system (Krakauer, Krakauer, & Mitchell, Reference Krakauer, Krakauer and Mitchell2025). Human-engineered systems do not straightforwardly satisfy standard criteria for strong emergence, because they are purposely designed to emulate target behavior. In this context, anthropomorphizing labels such as learning, emerging, and reasoning are more accurately described as wishful mnemonics (McDermott, Reference McDermott1981) – a textbook case of machine yearning – than reasonable scientific claims documenting model capabilities.

LLMs can surely capture certain statistical regularities from the output of language (text, data), but they do not infer grammar the way humans do and fail to capture fundamental principles of linguistic structure (Murphy et al., Reference Murphy, Leivada, Dentella, Günther and Marcus2025). They do not reliably identify grammaticality (Dentella et al., Reference Dentella, Günther, Murphy, Marcus and Leivada2024; Dentella, Günther, & Leivada, Reference Dentella, Günther and Leivada2025a). Instead of performing higher-order reasoning, they have been shown to engage in “sophisticated pattern matching” (Shojaee et al., Reference Shojaee, Mirzadeh, Alizadeh, Horton, Bengio and Farajtabar2025). They have a strong linearity bias and do not capture deep syntactic structures (Diego-Simón et al., Reference Diego-Simón, Chemla, King and Lakretz2025). There is no evidence that LLMs capture cross-constructional generalizations about null complementizers (Momma, Richards, & Ferreira, Reference Momma, Richards and Ferreira2025). When tested to learn restrictions on filler-gap dependencies in English and Norwegian, the generalizations learned by language models diverge from those of humans (Kobzeva et al., Reference Kobzeva, Arehalli, Linzen and Kush2025). This does not support the claim that filler-gap dependencies and island constraints can be learned without domain-specific biases. LLMs have a preference for manipulating noun-related information over verb-related information (Dentella et al., Reference Dentella, Huang, Mansi, Grieve and Leivada2025b), thus distinguishing them from highly grammaticalized types of thought that emphasize dimensions of tense, aspect, and mood. Lastly, text-to-image models also fail to represent the meaning of prompts with relatively basic syntactic structures triggering compositional meanings (Leivada, Murphy, & Marcus, Reference Leivada, Murphy and Marcus2023; Murphy, de Villiers, & Morales, Reference Murphy, de Villiers and Morales2025).

Within cognitive neuroscience, LLMs have arguably been unable to help with causal-mechanistic questions concerning how the algebraic properties of syntax-semantics are neurally enforced (Marcus, Reference Marcus2001; Murphy, Reference Murphy2015, Reference Murphy2020, Reference Murphy2024b, Reference Murphy2025; Murphy et al., Reference Murphy, Woolnough, Rollo, Roccaforte, Segaert, Hagoort and Tandon2022). Many LLM-brain alignments seem to be “driven by fragile methodologies and overlooked confounds” (e.g., word rate and positional signals) (Hadidi et al., Reference Hadidi, Feghhi, Song, Blank and Kao2025). An fMRI study showed that transformers were significantly inferior at predicting neural signatures to models explicitly designed to encode the syntactic relations between words (Fodor, Murawski, & Suzuki, Reference Fodor, Murawski and Suzuki2025). Problematic causal logic is also common in the LLM literature (e.g., affirming the consequent) (Guest & Martin, Reference Guest and Martin2023), whereby researchers often confuse types of inference and misunderstand the evidentiary role that correlation plays (see also Meijer, Reference Meijer2021). A scientist or artificial model may be able to classify and predict very well – all without providing causal explanations of a system.

FM (2025: 3) warn against dismissing LLMs as approximation devices that can fit any data without revealing the true underlying system. But this analogy is apt. LLMs can mimic many linguistic phenomena because they are universal function approximators, and their capacity and data exposure are unbounded. LLM fluency should not be confused with competence (Slaats & Martin, Reference Slaats and Martin2025). We therefore stress that the type of functionalist approach to language discussed by FM sidelines the necessary contributions to theories of cognition from symbolic, algebraic structures.

FM invoke Kubrick’s Dr. Strangelove in their article’s title, referring to an emerging “love” for LLMs winning over irrational fear and anxiety. We feel that another Kubrick movie, 2001: A Space Odyssey, teaches an altogether simpler and more appropriate lesson: to hesitate before trusting HAL’s words.