As linguists working on language evolution, we share F&M’s enthusiasm about the potential for Language Models (LMs) to help answer questions about language learning and the human language faculty. Studying the ability of artificial systems like LMs to acquire important features of natural languages is a potentially powerful tool for identifying critical cognitive capacities that contribute to language learning in humans. We see parallels with the long history of work in comparative psychology, uncovering capacities in humans that may be shared with non-human animals (see Arnon et al., Reference Arnon, Carmel, Claidière, Fitch, Goldin-Meadow, Kirby, Okanoya, Raviv, Wolters and Fisher2025; Gibson, Reference Gibson, Gibson and Tallerman2011; Petkov & ten Cate, Reference Petkov and ten Cate2020). However, current LMs far outstrip any non-human animal in their ability to approximate human language. As F&M point out, to the extent that LMs assign structures to sentences that align with human intuition and generalise in the same way that humans do, they provide an existence proof that linguistic input alone is sufficient to achieve descriptive adequacy for human language.

But as F&M also point out, a central goal of linguistics is explanatory adequacy, i.e., explaining how language is learned and why some logically possible types of language are rare or unattested. F&M, following Chomsky, assume that learnability and language universals are straightforwardly causally linked: constraints on language variation arise from inductive biases, and inductive biases active during learning will also lead LMs to generate only human-like languages. We would like to challenge this assumption. Learning alone cannot explain language structure, because learning is only one of the constraints acting on the cultural process by which languages are transmitted.

F&M argue that both humans and LMs “encode and decode information in linguistic strings incrementally and predictively…To the extent that human language is structured in a way that supports this kind of incremental prediction…it is also aligned with the…inductive biases of language models.” In other words, shared functions can lead to convergent, shared solutions. F&M makes the analogy between bird and airplane wings: both are shaped by the function of flight. However, this is at best only one highly proximal function that shapes these systems. A theory of bird flight based on modern airplane design would fail to account for the many differences between these two methods of flying through the air. To account for the differences between birds and planes, we need to consider less proximate functions than just aerodynamic lift, such as survival and reproduction in the case of the bird, and transporting passengers at low cost in the case of the plane. Similarly, LMs and human language are shaped by the function of incremental prediction, but this captures only one narrow aspect of communicative function. After all, human language is not purely a product of learning to incrementally predict. It is also shaped by the need to activate and transform conceptual representations, including complex thoughts, grounded in world knowledge and reflecting features of human perception and action, into language so that they can be shared with others (e.g., Culbertson et al., Reference Culbertson, Schouwstra and Kirby2020). In other words, LMs are trained only to incrementally predict language, whereas for humans, that task is incidental, or only necessary so far as it facilitates communication. Any theory of language that does not explicitly incorporate the communicative function of language will thus not achieve explanatory adequacy; it will fail to account for constraints on human language arising from this function.

To the extent that LMs, tasked with next word prediction, produce human-like language, they may do this because they are trained on data generated by humans with real communicative goals, not because they themselves have communicative goals. In the absence of a communicative task, even human learners will produce language that is highly predictable but not communicatively functional. To explore this, we use laboratory experiments in which human participants learn and reproduce a miniature language, thereby transmitting it over multiple simulated generations. If these languages are not used for communication, then the languages that evolve in this cultural process are shaped purely by pressures for learnability, ultimately leading to simple (i.e., easy-to-learn) systems which are useless for communication (e.g. Kirby, Cornish, & Smith, Reference Kirby, Cornish and Smith2008). Having participants also use the miniature languages for communication provides a crucial counteracting pressure, preventing this kind of language collapse. Only when both learning and communication are part of the cultural process do we see structured systems emerge with language-like properties (e.g., Kirby et al., Reference Kirby, Tamariz, Cornish and Smith2015; Motamedi et al., Reference Motamedi, Schouwstra, Smith, Culbertson and Kirby2019; Smith & Culbertson, Reference Smith and Culbertson2025). Similarly, if LMs are trained on text generated by other LMs, languages transmitted in this way degenerate as they are passed from model to model; later models produce probable sentences too frequently and generate meaningless sequences that no human would ever produce (Shumailov et al., Reference Shumailov, Shumaylov, Zhao, Papernot, Anderson and Gal2024). F&M do mention another body of promising modelling work, the so-called emergent communication paradigm (e.g., Chaabouni et al., Reference Chaabouni, Kharitonov, Dupoux and Baroni2021; Havrylov & Titov, Reference Havrylov and Titov2017; Lazaridou, Peysakhovich, & Baroni, Reference Lazaridou, Peysakhovich and Baroni2017). This does explicitly model communication, usually between pairs of simulated agents, but uses different model architectures and learning algorithms, not based on incremental prediction, which do not currently come close to achieving descriptive adequacy for human language.

Both iterated learning experiments and artificial language learning experiments, in which participants must communicate and extrapolate beyond their input, have been particularly helpful in building explanatory theories of human language (e.g., Culbertson, Reference Culbertson2023; Smith, Reference Smith2022). To the extent that languages emerging from similar experiments with LMs differ from the features we see in human populations, then the particularities of human cognition, human inductive biases, human communicative goals, and/or human non-linguistic experience actually matter for explaining language structure. This is an exciting open question, and we look forward to more research that uses methods from language evolution to probe the explanatory adequacy of LMs. Particularly promising for us is to manipulate the input and training objectives to incorporate more of what we have suggested might currently be missing: world/conceptual knowledge and representations, and the task of mapping between these representations and language, in the context of communication.