Futrell & Mahowald (F&M) want readers to see them as reasonable centrists, but do so by casting other positions as polarized extremes via misdirection and strawmen. Their centrism is an illusion and comes at the cost of saying anything substantive or advancing the conversation. By the end, the sheep becomes a wolf: Claudette is inviting Norm for dinner, and Norm is on the menu. With an invitation like that, no wonder linguists would rather continue to focus on the human capacity for language in all its glorious variations.

F&M frame radical positions as a middle ground by strawmanning the debate. For example, regarding learnability constraints, they state “There is no logical necessity for a restrictive formalism to explain language learning.” and “The key mistake was … that inductive bias has to come from restrictions on the learner’s hypothesis space.” They highlight Wilson’s (Reference Wilson2025) position paper describing several deep learning phenomena using the PAC-Bayes framework via soft’ biases. In reality, these are a set of hidden “hard” biases: a fixed computable hypothesis class, a fixed description-length prior tied to the architecture, I.I.D. sampling assumptions, bounded-loss assumptions, task-dependent representational primitives (by fallaciously appealing to Li and Vitanyi’s (Reference Li and Vitányi2008) “additive-constant” invariance theorem), and other capacity constraints. While description-length priors are coherent, Wolpert (Reference Wolpert2021) shows a bias only justifies a particular hypothesis class to the extent it matches the (unknown) distribution over tasks (see James et al., Reference James, Witten, Hastie, Tibshirani and Taylor2023 and Curth, Jeffares, & van der Schaar, Reference Curth, Jeffares and van der Schaar2023 on task-dependent complexity and double-descent). Without explicitly formalizing that alignment, moving from “real-world language data is inherently compressible” to endorsing biases like those above is meaningless. Whether hypotheses are restricted, admissible data presentations are restricted, or learning resources in time and space are unrestricted, there are fundamental trade-offs in solving learning problems (Angluin, Reference Angluin1988; Goyal & Bengio, Reference Goyal and Bengio2022, a.o.).

F&M’s own candidate constraints like sensitivity and information locality are also model-dependent (Hahn & Rofin, Reference Hahn and Rofin2024; Someya et al., Reference Someya, Svete, DuSell, O’Donnell, Giulianelli and Cotterell2025), task-dependent (Hahn, Jurafsky, & Futrell, Reference Hahn, Jurafsky and Futrell2021), and ultimately “hard.” Even Wilson and colleagues agree that “it is largely the design of the architecture that makes generalizable solutions more easily accessible” (Goldblum et al., Reference Goldblum, Finzi, Rowan and Wilson2024). Like all “domain-general” biases, F&M’s candidates turn out to be just “domain-specific” biases of a different sort (Rawski & Heinz, Reference Rawski and Heinz2019). In other words, the domains and biases F&M want are real, and the others are not. F&M can only achieve centrism from such an extreme position by misassigning people to trivially false claims, like that Rawski and Baumont (Reference Rawski and Baumont2023)’s 1-page paper says Large Language Models (LLMs) “can learn to approximate anything” when it does not. Strawmanning debates by mischaracterizing one perspective to position oneself as a middle ground, while ignoring the substance of the issues, is disingenuous and ultimately leads nowhere.

F&M cover these realities with doublespeak and fallacious arguments. They hype LLM success, concede the successes are disagreed upon and may always emerge from shallow heuristics (Liu et al., Reference Liu, Ash, Goel, Krishnamurthy and Zhang2022; McCoy, Pavlick, & Linzen, Reference McCoy, Pavlick and Linzen2019; Petty et al., Reference Petty, Hu, Wang, Ravfogel, Merrill and Linzen2025; Vafa et al., Reference Vafa, Chang, Rambachan and Mullainathan2025; Vázquez Martínez et al., Reference Vázquez Martínez, Heuser, Yang and Kodner2026, a.o.), backtrack to “this is not the right place to adjudicate these issues,” and then reassert that LLMs have “learned ‘the real thing’” and are “not purely relying on heuristics or memorization or cheap tricks.” Even their interpretability “thought experiment” about an idealized “c-command neuron” is a revamp of standard cognitive neuroscience techniques with the same reification fallacies and circularity (Buzsaki, Reference Buzsáki2019). Again they admit “LMs just don’t represent anything in such human-interpretable ways,” yet insist they are “linguistically interesting.” This is not a serious discussion.

Beyond the banal observation that “there are increasingly researchers using language models to ask questions that are informed by and which can inform linguistic theory,” what is F&M’s actual point? Simultaneously insisting on the Eliminativist fallacy (Smolensky, Reference Smolensky1988) that restrictive symbolic formalisms are “dogma,” while claiming “we are not eliminativists about grammatical structure,” is incoherent. Taking a Limitivist view (Smolensky, Reference Smolensky1988) where symbolic formalisms are an “approximate explanation” and “graded, probabilistic, and fuzzy” subsymbolic primitives are cognitively real, they bizarrely omit 30 years of neural networks in generative linguistics via Optimality Theory and Harmonic Grammar (Prince & Smolensky, Reference Prince and Smolensky1997; Smolensky & Legendre, Reference Smolensky and Legendre2006). Advocating for “domain-general” learning biases, which end up being domain-specific, paints them into a Dennettian corner to conclude the “troll’s truism” (Schackel, Reference Shackel2005) that “linguistic structure is real.” Little else remains, certainly nothing new.

Here we see the tragedy of the centrist, ruled by “the zeitgeist” in F&M’s terms, at the cost of ideas or content. Like most self-described centrists, however, their strategies serve to cover another goal. F&M seem to want linguistics to effectively become a niche within psychology (Núñez et al., Reference Núñez, Allen, Gao, Miller Rigoli, Relaford-Doyle and Semenuks2019), where replicability and theory crises (Eronen & Bringmann, Reference Eronen and Bringmann2021) mean “theories rise and decline, come and go, more as a function of baffled boredom than anything else; and the enterprise shows a disturbing absence of that cumulative character that is so impressive” (Meehl, Reference Meehl1978). Linguistic theory is so robust and successful at phenomena discovery that LLM performance benchmarks (Gauthier et al., Reference Gauthier, Hu, Wilcox, Qian and Levy2020, a.o.) and evaluations (Hale & Stanojević, Reference Hale and Stanojević2024, a.o.) crucially rely on them. When neural models are used to trivialize the substance of linguistic theory, as F&M do with competence and performance (Berwick, Reference Berwick2018; Firestone, Reference Firestone2020; West et al., Reference West, Lu, Dziri, Brahman, Li, Hwang, Jiang, Fisher, Ravichander, Chando, Newman, Koh, Ettimner and Choi2024), linguists have no reason to engage. Replacing theory-driven science with model-driven science, and phenomena with effects-chasing, is a profound retreat from the core insights which gained linguistics membership in cognitive science in the first place.

F&M wonder why many linguists remain uninterested in LLMs, which have not yet discovered virtually any new phenomena nor explained anything of theoretical significance. Yes, LLMs are a technological advance and engineering achievement. Nonetheless, language models, in whatever guise, will continue to be a scientific embarrassment to language scientists as long as they do not offer robust phenomena discovery, counterfactual support, explanatory depth, and other scientific insight into the fundamental properties of the human capacity for language.