Kubrick’s memorable title “How I learned to stop worrying and love the bomb” evokes a hapless Major Kong who, after fiddling with electrical wiring in the bay of his plane, mounts the nuclear bomb that will set off the doomsday device, riding it to his death while hooting and waving his cowboy hat. It is a fitting choice of title for an invitation to embrace technology with abandon. But there are real choices to be made. How reproducible do we want our research to be in an era when proprietary language models are becoming an unexamined default? How complicit do we want to be in polluting our information ecology with the fallout of synthetic text? How do we make principled decisions about where to focus our efforts and how to future-proof our research?

The quality and integrity of scientific research rest on a set of core values that most scientists share and uphold. A widely used code (NCCRI, 2018) mentions honesty (reporting processes and sources accurately), scrupulousness (being precise and thoughtful at every stage of the research), transparency (working according to open science principles), independence (being impartial and not beholden to commercial or political interests), and responsibility (aiming to avoid environmental and societal harms). Language models present an unprecedented opportunity to undermine all of these principles in one fell swoop. Ever since BERT, their training data has been a carefully guarded and legally fraught secret (Liesenfeld, Reference Liesenfeld2025; Samuelson, Reference Samuelson2023); their frictionless design blinds us to underlying processes and transformations (Drucker, Reference Drucker and O’Sullivan2024); their blackboxed nature endangers reproducible research (Spirling, Reference Spirling2023); subscription-based access control leads to user lock-in (Widder, Whittaker, & West, Reference Widder, Whittaker and Myers West2024); and the computational burdens of training and inference put exorbitant demands on energy and environment (Luccioni, Jernite, & Strubell, Reference Luccioni, Jernite and Strubell2024). Understanding human language is meant to be an endeavor to reveal the workings of what was once a black box, not an invitation to pay-per-query for access to a new one.

Dazzling new capabilities have a way of obscuring sociotechnological complications (Suchman, Reference Suchman and Teresa2019). Thousands of papers published in the period 2021–2023 relied on nominally free access to OpenAI’s text-davinci-002 language model. In late 2023, OpenAI suddenly phased out this model, undercutting at one blow the reproducibility of all this work (Liesenfeld, Lopez, & Dingemanse, Reference Liesenfeld, Lopez and Dingemanse2023). One affected study purported to show that “large language models are better than linguists”; its authors had to note that after text-davinci-002 was withdrawn, “the GPT 3 model that we were able to use as a replacement … is less sophisticated at understanding complex tasks” (Ambridge & Blything, Reference Ambridge and Blything2024). The result is a scientific edifice built on shifting sands, where “findings” have best-by dates haphazardly set by commercial parties. Blind appeal to technological advances won’t suffice: for new models offered as replacements, there is typically less clarity about training data, fine-tuning procedures, and system prompts imposed by model providers, all of which shape input–output transformations in ways that are unobservable and irreproducible (Han et al., Reference Han, Tan, Miao, Chen, Sun, Miaomiao, Xin, Chang and Yiliao2026; Zhou et al., Reference Zhou, Schellaert, Fernando Martínez-Plumed, Ferri and Hernández-Orallo2024). Open-source models might provide some respite, but the most widely used models are typically open weights, meaning that we lack the data and documentation needed to understand and meaningfully manipulate model behavior (Liesenfeld & Dingemanse, Reference Liesenfeld and Dingemanse2024). When such information surfaces, it often puts supposed abilities in a new light, much like the excitement over ChatGPT passing the bar exam withers under a cold hard look at the training data and measurement methods (Martínez, Reference Martínez2025). Court evidence revealed that Meta’s Llama, like many language models, has been trained not only on Internet-scraped data but also on the LibGen database of pirated books (Reisner, Reference Reisner2025), which includes thousands of works in linguistics. We would be in considerably less awe at a child effortlessly acquiring, e.g., the Russian case system if we found they had been somehow trained on virtually all linguistic descriptions of that system ever written.

Current language models are fascinating technological artifacts that deserve careful study. Classic work (Suchman, Reference Suchman2007; Weizenbaum, Reference Weizenbaum1976) and modern research (Askari et al., Reference Askari, Zarrieß, Alacam, Sieker, Charpentier, Choshen, Cotterell, Omer Gul, Hu, Liu, Jumelet, Linzen, Mueller, Ross, Sanjay Shah, Warstadt, Gotlieb Wilcox and Williams2025; Mayor, Bietti, & Bangerter, Reference Mayor, Bietti and Bangerter2025; Pütz & Esposito, Reference Pütz and Esposito2024; Smit, Smits, & Merrill, Reference Smit, Smits and Merrill2024) already provide constructive ways to think about the cognitive and linguistic capacities at play, going far beyond the properties of disembodied, decontextualized strings of linguistic tokens. While the target article acknowledges a “narrowness in focus” within generative linguistics (§3.3.1), we argue that language models do not expand this focus sufficiently to justify their uncritical adoption as de facto essential tools in understanding language. The consequences of casting language models as central to understanding human language are neither socially nor scientifically benign. Besides undermining scientific integrity, uncritical adoption risks inflating model capacities by taking illusions of understanding at face value (Messeri & Crockett, Reference Messeri and Crockett2024) and intensifies the hyperfocus on text over talk within cognitive science (Cuskley, Woods, & Flaherty, Reference Cuskley, Woods and Flaherty2024).

We reject the frame that linguists should “learn to stop worrying.” Concern about the reliability and validity of models and tools is at the core of effective scientific practice. Calls for the uncritical adoption of language models ring hollow when there is a growing array of well-informed research that makes visible how language models reify and reinforce narrow views of language (Birhane & McGann, Reference Birhane and McGann2024; Birhane, Reference Birhane2021; Rasenberg et al., Reference Rasenberg, Amha, van Koppen, van Miltenburg, de Rijk, Stommel and Dingemanse2023) and how they affect the language sciences as both a community of practice and a scientific endeavor (Guest et al., Reference Guest, Suarez, Müller, van Meerkerk, Oude Groote Beverborg, de Haan, Reyes Elizondo, Blokpoel, Scharfenberg, Kleinherenbrink, Camerino, Woensdregt, Monett, Brown, Avraamidou, Alenda-Demoutiez, Hermans and van Rooij2025; Jacobs, Reference Jacobs2023; Palmer, Smith, & Spirling, Reference Palmer, Smith and Spirling2024; van Rooij et al., Reference van Rooij, Guest, Adolfi, de Haan, Kolokolova and Rich2023).

Language models are not the first technology with revolutionary pretensions and they won’t be the last (Franklin, Reference Franklin1999; Illich, Reference Illich1973). A tech-first perspective tells us to stop worrying and embrace the inevitable. A values-first perspective asks how we can best uphold the values and standards that make our research well-rounded, robust, and reproducible. Robust research centers values over technology, moving from unexamined techno-positivism to a concern with doing the best science possible.