It’s impossible to resist pointing out that in Stanley Kubrick’s 1964 film Dr. Strangelove, which inspired Futrell and Mahowald’s article title, the central theme involved people not thinking things through carefully enough, leading to a global apocalypse. I’m not worried about quite that level of catastrophe, mind you, but I do think the article risks further feeding into a pervasive and ultimately harmful misperception about the relationship between language models (LMs) and human language processing – namely, the too-readily overextended idea of LMs as “model systems” alluded to in the article’s abstract.

What do we mean when we say “model system”? A model, at any level of explanation, is an embodiment of theory that, in addition to capturing generalizations and licensing verifiable inferences and predictions – requirements of any theory – also contains structured correspondences between entities and relationships in the model and corresponding entities and relationships in the real-world system being modeled (Frigg & Nguyen, Reference Frigg, Nguyen, Magnani and Bertolotti2017). Notwithstanding its imperfections, Marr (Reference Marr1982) remains a widely accepted scaffolding (Peebles & Cooper, Reference Peebles and Cooper2015), so let’s consider the status of and prospects for LMs as models of language processing at each of Marr’s three levels.

First, the implementation level, which we need not spend much time on. The authors argue for a hypothesis-generation role, but they are also clearly aware that functional parallels are not implementation-level models. Indeed, the connectionism underlying today’s LMs was founded on the abandonment of implementation-level biological correspondences in favor of function optimization. In Fridman (Reference Fridman2021), Jay McClelland recalls Geoff Hinton successfully arguing, “The problem with the way you guys have been approaching this is that you’ve been looking for inspiration from biology…That’s the wrong way to go about it.” Consequently, the dominant LM research today demonstrates little substantive engagement with the numerous relevant advances made in implementation-level understanding of the brain since connectionism’s loose 20^th-century inspirations (e.g., Barrett & Satpute, Reference Barrett and Satpute2013; Hawkins & Ahmad, Reference Hawkins and Ahmad2016; Hickok & Poeppel, Reference Hickok and Poeppel2007; Wright, Hedrick, & Komiyama, Reference Wright, Hedrick and Komiyama2025; for higher-level reviews see Pulvermüller et al., Reference Pulvermüller, Tomasello, Henningsen-Schomers and Wennekers2021; Pulvermüller, Reference Pulvermüller2023 and Durstewitz, Averbeck, & Koppe, Reference Durstewitz, Averbeck and Koppe2025).

Algorithmic/representational level. The authors’ argument here hinges on convergence, that systems solving the same hard problem are driven toward similar mechanisms (Cao & Yamins, Reference Cao and Yamins2024). But unlike visual object recognition, a key part of what makes “language” hard is that it decomposes into very distinct subproblems, not merely competing constraints within a single problem. In addition, mechanistic exceptions to convergence are easily found in biology (e.g., sharks and dolphins facing ecologically similar prey-capture problems evolved markedly different mechanisms, relying on electroreception and echolocation, respectively) and in computation (e.g., iterative and recursive computations of the Fibonacci sequence). In medical research, model-organism choices are supported by existing, solidly established mechanistic evidence; e.g., mice and ferrets are preferred in work on tumorigenesis and influenza, respectively – not vice versa – because mice share oncogene pathways with humans and ferrets share lung physiology (Belser, Katz, & Tumpey, Reference Belser, Katz and Tumpey2011; Bedell, Jenkins, & Copeland, Reference Bedell, Jenkins and Copeland1997).

Consider the following scenario. Two aliens from another galaxy, whom I’ll dub Abbot and Costello after Villeneuve (Reference Villeneuve2016), show up on Earth with their own billions of years of evolution behind them. In a year, they absorb nearly everything ever written, and become phenomenal conversationalists; moreover, they generously offer to let human scientists experiment and probe their brain-boxes to our hearts’ content. Here’s the question: why in the world would one expect to find mechanistic correspondences between what’s happening in Abbot and Costello’s boxes and what is happening in the human cognitive apparatus? Nonetheless, for LMs, the “argument from amazingness” is pervasive – if they are doing something so very human so very well, surely they must be operating under the hood in human-like ways?

As I’ve highlighted, the argument is unreliable at best, and, moreover, it just is not necessary in order for LMs to contribute value to computational cognitive modeling. Giulianelli, Opedal, and Cotterell (Reference Giulianelli, Opedal and Cotterell2024), for example, use large language models to generate a sample of continuations in order to operationalize a generalization of surprisal that captures distinctions between anticipatory and responsive online processes in comprehension. This use depends only on the generation of plausible text, LMs’ undisputed strength (for better or worse; see Resnik, Reference Resnik2025), without requiring claims about their cognitive plausibility. Similarly, LMs can provide useful proxies (with caveats; see Oh & Schuler, Reference Oh and Schuler2023) for human comprehenders’ next-word probabilities when operationalizing models where such probabilities play a role, e.g., Hahn et al.’s (Reference Hahn, Futrell, Levy and Gibson2022) integration of expectation- and memory-based theories of processing difficulty. LMs are a promising source of world knowledge for hierarchical predictive coding models, which are neurobiologically plausible and have recently been applied to language comprehension (Nour Eddine et al., Reference Nour Eddine, Brothers, Wang, Spratling and Kuperberg2024; Ohams et al., Reference Ohams, Nair, Bhattasali and Resnik2026).

Computational theory level. Ironically, many who “love the language models” seem to exhibit the same physics envy that you find in Chomskian theory, idealizing a uniform, elegant computation-level paradigm for a wide range of phenomena. But evolution is a scruffy tinkerer, not a neat software engineer (Jacob, Reference Jacob1977). It doesn’t re-design from the bottom up when it produces new functionality. For LMs, just as for Chomsky, the drive to explain “du visible compliqué par de l’invisible simple” (Perrin, Reference Perrin1913) creates a problematic gap: even if these architectures can approximate diverse linguistic phenomena, and even if correlations are found between one black box and the other, ultimately there remains the burden of mapping any such computational theory to a biologically plausible algorithmic account of real-time processes distributed over diverse, specialized components. And there are plenty of other, less problematic computational-theory games in town (e.g., Eliasmith, Reference Eliasmith2013; Suárez et al., Reference Suárez, Mihalik, Milisav, Marshall, Li, Vértes, Lajoie and Misic2024; Sprevak & Smith, Reference Sprevak and Smith2023).

At all levels, we do need a pluralistic approach, and LMs can be tremendously productive as tools in support of developing explanatory models. But LMs themselves are poor candidates as “model systems” at any of Marr’s three levels. The distinction matters, for how we prioritize research questions, interpret results, and communicate findings to neighboring fields. For a whole variety of reasons, large language models are the subject of enormous hype. Let’s not undermine some of the more nuanced arguments in this paper by overstating the case.