Futrell and Mahowald attempt to reconcile two extreme views on the utility of language models (LMs) for studying linguistic theories, human language learning, and processing. However, the presented viewpoints constitute a false dichotomy, relying on several assumptions which do not necessarily stand true and positioning investigations into the syntactic structure of well-resourced written languages as essential to linguistic knowledge. While there is no denying the influential status of generative linguistics, broadly stating that “linguists” should “love language models” does not do justice to the present diversity of objectives and approaches in the field.
Specifically, Futrell and Mahowald constitute a re-enforcement of narrow and exclusionary perspectives on the significance of syntactic structure to the nature of human language. Moreover, the titular reference to the film Dr. Strangelove (Kubrick, Reference Kubrick1964) raises the specter of highly dangerous and destructive technology while lending an air of inevitability to discussion of its applications. Despite setting expectations that it will engage with conceptions of risk, Futrell and Mahowald overall neglect any such analysis. It instead advances a crackpot realism, wherein the concerns of a powerful elite are portrayed as “realistic” in a sense which is technocratic and detached from broader human consequences (Mills, Reference Mills2000).
Arguing that the reality of linguistic structure underlies the success of LMs, Futrell and Mahowald declare that “compressibility is structure.” However, we must consider how fundamentally lossy this compression is. The LMs (LMs) under discussion rely on text, and the limited “linguistic layers” transparently encoded therein. Contrastively, children receive richly diverse audiovisual information and variation in phonetic and prosodic forms, grounded in a social world. The conflicting positions described in Futrell and Mahowald share an imagined connection between linguistic structure and function which casually excludes sociolinguistic variation and its functional correlates. The result is that both camps effectively champion a rarefied form of synthetic data, largely excluding sociolinguistic behavior from the conception of what language is. In the context of understanding human language acquisition, this is a notable oversight, as social and phonetic attunement to familiar varieties is acquired early (Kinzler, Dupoux, & Spelke, Reference Kinzler, Dupoux and Spelke2007) and may substantially guide processes such as word learning (Tripp, Feldman, & Idsardi, Reference Tripp, Feldman and Idsardi2021).
The assumption that model input data belong to a single variety positions standard language ideology as compatible with both approaches to describing language, despite the centrality of social and pragmatic functions to evaluating LM success. The ideology is more broadly visible in the habitual conflation of language ability and intelligence found in discourse on LMs (Mitchell & Krakauer, Reference Mitchell and Krakauer2023) and in the putative connection between linguistic ability and a “language of thought” (Berwick & Chomsky, Reference Berwick and Chomsky2016, p. 84). As language scientists, we must confront, not evade, this ideology and the way it animates both approaches.
Despite calls to diversify the languages under study, the role of English as a central source of insight is not critically addressed. Reliance on the most well-resourced language is a matter of convenience which undermines the scientific rigor of studies probing LM abilities. Most world languages do not have such ample training data, and in fact many lack an adequate number of living speakers to fruitfully compare LMs to the learning and processing of these languages (Liu, Richardson, Hatcher, & Prud’hommeaux, Reference Liu, Richardson, Hatcher, Prud’hommeaux, Muresan, Nakov and Villavicencio2022). Consequently, indigenous and endangered languages are again pushed to the periphery of linguistic knowledge.
Likewise, universal sociolinguistic abilities such as the acquisition of multiple grammars, selection of speech registers, and interpretation of social identity through language behavior are all disregarded as outside the “core” linguistic function of incremental predictive processing, positioning the monolingual generative capacity of Internet-scale LMs as more critical to our understanding of language than any living speaker of the world’s many endangered languages. This evaluation aligns with the profit motive behind LLM development.
Futrell and Mahowald associate danger with the neglect of LMs as a source of insight, insisting that we may disregard them “at our peril.” LMs, as understood in this frame, effectively erase ways of languaging which are not readily captured in large repositories of text strings. Multimodal signals, admixtures of codes, and the use of sign languages are all ignored, implicitly treated as complexity which is justifiably lost in message “compression.”
In this way, linguistic structure is presumed discoverable through probing systems designed specifically to produce prestige language forms. Further, these efforts are framed with urgency. While the engineering of improved language technology clearly stands to benefit those who control proprietary software systems, broader human consequences of reliance on LMs entail the continued exclusion and subordination of marginalized subjectivities. Language technology disproportionately discriminates against and misrepresents minoritized language users (Joshi, Santy, Budhiraja, Bali, & Choudhury, Reference Joshi, Santy, Budhiraja, Bali, Choudhury, Jurafsky, Chai, Schluter and Tetreault2020). Foregrounding models of written standard English reproduces this existing dynamic, enshrining a specific variety as a uniquely useful and iconic representation of language form, function, and human intelligence.
For language science to equitably serve humankind, it is necessary to view both generativist and connectionist linguistic traditions with skepticism, opposing their focus on majoritized paradigms (e.g., monolingualism and literacy) at the expense of acknowledging the true diversity of natural human language behavior. This commentary seeks to contextualize LM utility within a broader discourse regarding answers to two questions: 1) who should be considered stakeholders in investigations of linguistic structure? and 2) what are our ethical obligations to those stakeholders? The reality of LMs is that a narrow sample of language varieties associated with globally dominant social groups have been treated as offering pressing insight into the human condition, while minoritized language systems are relegated to the role of confirmatory testbeds, if they are considered at all. Treating politically minoritized, disabled, racialized, and gendered persons as stakeholders in these investigations requires us to regard LMs which disregard and distort these subjectivities (Wang, Morgenstern, & Dickerson, Reference Wang, Morgenstern and Dickerson2024) as obviously inadequate and potentially dangerous.
In seeking a way forward which appropriately leverages LMs for linguistic insight, we must be explicit regarding the utility of written text and oppose the characterization of sociolinguistic diversity as harmlessly excluded from technocratic distillations of linguistic knowledge.
You have
Access
Futrell and Mahowald attempt to reconcile two extreme views on the utility of language models (LMs) for studying linguistic theories, human language learning, and processing. However, the presented viewpoints constitute a false dichotomy, relying on several assumptions which do not necessarily stand true and positioning investigations into the syntactic structure of well-resourced written languages as essential to linguistic knowledge. While there is no denying the influential status of generative linguistics, broadly stating that “linguists” should “love language models” does not do justice to the present diversity of objectives and approaches in the field.
Specifically, Futrell and Mahowald constitute a re-enforcement of narrow and exclusionary perspectives on the significance of syntactic structure to the nature of human language. Moreover, the titular reference to the film Dr. Strangelove (Kubrick, Reference Kubrick1964) raises the specter of highly dangerous and destructive technology while lending an air of inevitability to discussion of its applications. Despite setting expectations that it will engage with conceptions of risk, Futrell and Mahowald overall neglect any such analysis. It instead advances a crackpot realism, wherein the concerns of a powerful elite are portrayed as “realistic” in a sense which is technocratic and detached from broader human consequences (Mills, Reference Mills2000).
Arguing that the reality of linguistic structure underlies the success of LMs, Futrell and Mahowald declare that “compressibility is structure.” However, we must consider how fundamentally lossy this compression is. The LMs (LMs) under discussion rely on text, and the limited “linguistic layers” transparently encoded therein. Contrastively, children receive richly diverse audiovisual information and variation in phonetic and prosodic forms, grounded in a social world. The conflicting positions described in Futrell and Mahowald share an imagined connection between linguistic structure and function which casually excludes sociolinguistic variation and its functional correlates. The result is that both camps effectively champion a rarefied form of synthetic data, largely excluding sociolinguistic behavior from the conception of what language is. In the context of understanding human language acquisition, this is a notable oversight, as social and phonetic attunement to familiar varieties is acquired early (Kinzler, Dupoux, & Spelke, Reference Kinzler, Dupoux and Spelke2007) and may substantially guide processes such as word learning (Tripp, Feldman, & Idsardi, Reference Tripp, Feldman and Idsardi2021).
The assumption that model input data belong to a single variety positions standard language ideology as compatible with both approaches to describing language, despite the centrality of social and pragmatic functions to evaluating LM success. The ideology is more broadly visible in the habitual conflation of language ability and intelligence found in discourse on LMs (Mitchell & Krakauer, Reference Mitchell and Krakauer2023) and in the putative connection between linguistic ability and a “language of thought” (Berwick & Chomsky, Reference Berwick and Chomsky2016, p. 84). As language scientists, we must confront, not evade, this ideology and the way it animates both approaches.
Despite calls to diversify the languages under study, the role of English as a central source of insight is not critically addressed. Reliance on the most well-resourced language is a matter of convenience which undermines the scientific rigor of studies probing LM abilities. Most world languages do not have such ample training data, and in fact many lack an adequate number of living speakers to fruitfully compare LMs to the learning and processing of these languages (Liu, Richardson, Hatcher, & Prud’hommeaux, Reference Liu, Richardson, Hatcher, Prud’hommeaux, Muresan, Nakov and Villavicencio2022). Consequently, indigenous and endangered languages are again pushed to the periphery of linguistic knowledge.
Likewise, universal sociolinguistic abilities such as the acquisition of multiple grammars, selection of speech registers, and interpretation of social identity through language behavior are all disregarded as outside the “core” linguistic function of incremental predictive processing, positioning the monolingual generative capacity of Internet-scale LMs as more critical to our understanding of language than any living speaker of the world’s many endangered languages. This evaluation aligns with the profit motive behind LLM development.
Futrell and Mahowald associate danger with the neglect of LMs as a source of insight, insisting that we may disregard them “at our peril.” LMs, as understood in this frame, effectively erase ways of languaging which are not readily captured in large repositories of text strings. Multimodal signals, admixtures of codes, and the use of sign languages are all ignored, implicitly treated as complexity which is justifiably lost in message “compression.”
In this way, linguistic structure is presumed discoverable through probing systems designed specifically to produce prestige language forms. Further, these efforts are framed with urgency. While the engineering of improved language technology clearly stands to benefit those who control proprietary software systems, broader human consequences of reliance on LMs entail the continued exclusion and subordination of marginalized subjectivities. Language technology disproportionately discriminates against and misrepresents minoritized language users (Joshi, Santy, Budhiraja, Bali, & Choudhury, Reference Joshi, Santy, Budhiraja, Bali, Choudhury, Jurafsky, Chai, Schluter and Tetreault2020). Foregrounding models of written standard English reproduces this existing dynamic, enshrining a specific variety as a uniquely useful and iconic representation of language form, function, and human intelligence.
For language science to equitably serve humankind, it is necessary to view both generativist and connectionist linguistic traditions with skepticism, opposing their focus on majoritized paradigms (e.g., monolingualism and literacy) at the expense of acknowledging the true diversity of natural human language behavior. This commentary seeks to contextualize LM utility within a broader discourse regarding answers to two questions: 1) who should be considered stakeholders in investigations of linguistic structure? and 2) what are our ethical obligations to those stakeholders? The reality of LMs is that a narrow sample of language varieties associated with globally dominant social groups have been treated as offering pressing insight into the human condition, while minoritized language systems are relegated to the role of confirmatory testbeds, if they are considered at all. Treating politically minoritized, disabled, racialized, and gendered persons as stakeholders in these investigations requires us to regard LMs which disregard and distort these subjectivities (Wang, Morgenstern, & Dickerson, Reference Wang, Morgenstern and Dickerson2024) as obviously inadequate and potentially dangerous.
In seeking a way forward which appropriately leverages LMs for linguistic insight, we must be explicit regarding the utility of written text and oppose the characterization of sociolinguistic diversity as harmlessly excluded from technocratic distillations of linguistic knowledge.
Acknowledgments
I thank Zoey Liu for her feedback on the initial proposal submission.
Financial support
This research received no specific grant from any funding agency, commercial or not-for-profit sectors.
Competing interests
The author declares none.