In the 1960s, psycholinguists were excited about the possibility of a computational-level account of language processing where measures of “processing difficulty” were linked to aggregate measures of linguistic complexity. In that case, the difficulty measure (ratings or reading times for whole sentences) and the complexity measure (syntactic transformations) reflected the available (psycho)linguistic technologies of the time (Miller & McKean, Reference Miller and McKean1964). The so-called Derivational Theory of Complexity (DTC) is widely cited as a cautionary tale about What Not to Do in cognitive science, but there are some enduring results establishing expectations for relative processing difficulty among different syntactic constructions.

Fifty to sixty years later, we again find enthusiasm for a computational-level account that links overall measures of processing difficulty to aggregate measures of linguistic complexity. Although today’s technical advances reflect richer processing data (eye fixation durations, electrical potentials in the brain, or utterance latencies) and linguistic complexity (contextual probabilities estimated by LLMs), all over words (and subword units) instead of sentences, the core idea retains the half-century-old notion that a general-purpose measure to explain processing difficulty is a useful construct, now expressed as surprisal theory (Levy, Reference Levy2008). That said, it is important to learn from the mistakes of the DTC.

Just as in the 1960s, there are more cases where the general theory falls short, but there are some core insights that the field is reluctant to abandon. In the 2020s, these insights pertain to predictability effects in language processing (Staub, Reference Staub2025); unlikely, information in a given context is more difficult to process. Even simple LMs quantitatively describe these effects well, which has led researchers to claim language processing is associated with probabilistic inference (Smith & Levy, Reference Smith and Levy2013). LLMs have since characterized predictability effects across different languages (Wilcox et al., Reference Wilcox, Pimentel, Meister, Cotterell and Levy2023; Xu et al., Reference Xu, Chon, Liu and Futrell2023) and larger datasets (Shain et al., Reference Shain, Meister, Pimentel, Cotterell and Levy2024), but are largely based on the same arguments. A major issue in this debate is that for LMs, predictability effects are exclusively linked to probabilistic inference, while the same cannot necessarily be said for humans, thanks to insights from the last few decades of psycholinguistic research (Brothers & Kuperberg, Reference Brothers and Kuperberg2021; Szewczyk & Federmeier, Reference Szewczyk and Federmeier2022).

From the fall of the DTC to the rise of surprisal theory, psycholinguists discovered finer-grained phenomena and experimental measures, largely shifting their attention to algorithmic-level questions. This fueled decades of research, and so few considered psycholinguistics to be equipped to make computational-level contributions. Futrell & Mahowald do say language processing is highly incremental, but psycholinguists’ goal has been seen as characterizing the underlying mental computations from each step of incremental processing.

Contrary to the authors’ claims that psycholinguistics has been hampered by weak statistical models of language, the field has made progress through empirical investigations of phenomena testing the processing system in tightly controlled experimental settings. Some examples include studies on linguistic illusions (Paape, Reference Paape2024; Phillips, Wagers, & Lau, Reference Phillips, Wagers and Lau2011), consequences of violated predictions (Frisson, Harvey, & Staub, Reference Frisson, Harvey and Staub2017; Kuperberg, Brothers, & Wlotko, Reference Kuperberg, Brothers and Wlotko2020), time delays (Chow et al., Reference Chow, Lau, Wang and Phillips2018; Tabor & Hutchins, Reference Tabor and Hutchins2004), and production-comprehension contrasts (Kandel, Wyatt, & Phillips, Reference Kandel, Wyatt and Phillips2024; Lee & Phillips, Reference Lee and Phillips2025). Simpler, interpretable models, not LLMs, have proven useful in this context (Aurnhammer et al., Reference Aurnhammer, Delogu, Schulz, Brouwer and Crocker2021; Fitz & Chang, Reference Fitz and Chang2019; Nakamura et al., Reference Nakamura, Momma, Sakai and Phillips2024).

That said, an important theoretical contribution of LLMs is to highlight the limits of predictability as a satisfying explanation for finer-grained aspects of language processing that unfold in time. LLM probabilities have limits as a functional explanation of human neural responses because they conflate semantic association with contextual expectations (Krieger et al., Reference Krieger, Brouwer, Aurnhammer and Crocker2025). On the behavioral side, they insufficiently capture qualitative patterns in argument role reversals (Lee, Nair, & Feldman, Reference Lee, Nair and Feldman2024) and other types of linguistic illusions (Zhang et al., Reference Zhang, Gibson and Davis2023), and quantitative patterns in syntactic processing (Huang et al., Reference Huang, Arehalli, Kugemoto, Muxica, Prasad, Dillon and Linzen2024).

These recent findings mark a shift from computational-level probabilities back towards algorithmic-level processes. To this end, Futrell & Mahowald suggest imposing cognitively realistic resource constraints on LLMs’ predictability estimates, and Giulianelli, Wallbridge, and Fernández (Reference Giulianelli, Wallbridge and Fernández2023, Reference Giulianelli, Wallbridge, Cotterell and Fernández2024) derive LLM-based difficulty measures for different levels of linguistic representation. We suggest another direction for progress: building on interactivity across levels of representation and incorporating predictability-based factors in process models. The field is moving in this direction already; LLMs complement insights about the time course of syntactic processing (Slaats, Meyer, & Martin, Reference Slaats, Meyer and Martin2024; Timkey et al., Reference Timkey, Huang, Oh, Prasad, Arehalli, Linzen and Dillon2025), can be combined with syntactic and semantic measures (De Santo, Reference De Santo2025; Jacobs et al., Reference Jacobs, Hubbard, Grobol and Federmeier2025; Stanojević et al., Reference Stanojević, Brennan, Dunagan, Steedman and Hale2023), and support algorithmic-level accounts of inference (Vigly et al., Reference Vigly, Qian, Sonderegger and O’Donnell2025).

Process models can also be motivated by implementational-level ideas such as predictive coding (Clark, Reference Clark2013; Friston, Reference Friston2009; Rao & Ballard, Reference Rao and Ballard1999), grounded in theories of neurobiological function. Even if Futrell & Mahowald cite this work, LLMs’ successes are at best rhetorically connected with predictive coding due to their limited biological plausibility (Goldstein et al., Reference Goldstein, Zada, Buchnik, Schain, Price and Aubrey2022; Ryskin & Nieuwland, Reference Ryskin and Nieuwland2023; Schrimpf et al., Reference Schrimpf, Blank, Tuckute, Kauf, Hosseini, Kanwisher, Tenenbauam and Fedorenko2021).

Building on successfully implemented predictive coding models in other cognitive domains, their recent extension to linguistic data is promising. Not only do they elegantly explain disparate effects in controlled experimental settings (Nour Eddine et al., Reference Nour Eddine, Brothers, Wang, Spratling and Kuperberg2024), they also explain naturalistic reading data, with predictive power comparable to LMs with similar size (Ohams et al., Reference Ohams, Nair, Bhattasali and Resnik2026). Crucially, both these advances were derived from connecting simpler cognitive models (Elman, Reference Elman1991; Rumelhart & McClelland, Reference McClelland and Rumelhart1981) to biologically plausible methods of computation and provide interpretable measures that are directly tied to interactive cognitive processes, unlike LLMs.

To summarize, LMs have “revived” an older psycholinguistic approach, now providing far more accurate estimates of aggregate processing difficulty and linguistic complexity. Probabilistic models of linguistic data are useful, but progress in psycholinguistics will also require mechanistically characterizing the mental computations at the algorithmic level, in order to translate estimates of which words are (un)predictable from above into neural processes of how they are processed from below. Thus, psycholinguists can find different ways of loving what LMs can do for them, either as models of the core mechanism in language processing or as components of more articulated algorithms, but their worries are justified.