Futrell & Mahowald (F&M) offer a compelling middle path in current debates about the relevance of language models (LMs) to linguistics. Against those who treat LMs as mere engineering tools with no bearing on linguistic theory, they argue that any system that solves the hard problem of language prediction and comprehension is likely to converge on a similar underlying structure. Against those who take the success of LMs to obviate linguistic theory, they insist that linguistic structure remains indispensable as a “real pattern” in Dennett’s sense: a higher-level abstraction that compresses, predicts, and supports counterfactual reasoning about both human and LM behaviour, even if it is not hard-wired. They further present mechanistic interpretability as a way to bridge such abstract structural descriptions and messy implementation details. I agree with this overall picture, but think it underplays what that bridge can carry. Not only do LMs discover real patterns in language, but they also implement specific algorithms for solving linguistic tasks. This algorithmic perspective strengthens F&M’s conciliatory project.

On Dennett’s account, a pattern is “real” if a description that uses it to explain a system’s behaviour is significantly more compressive and predictively useful than any available lower-level description. F&M plausibly treat grammatical constructs such as “subject,” “agreement,” or “filler–gap dependency” as real patterns in this sense: they are higher-level descriptions that let us summarise and forecast linguistic behaviour far more efficiently than tracking neural activations would. But “real pattern” is an intentionally liberal notion: many stable statistical regularities will qualify. By contrast, linguistic theories are traditionally construed as hypotheses about algorithms. A grammar does not just encode which strings are licit; it specifies how structured representations are built and transformed. If we translate this into generic talk of “real patterns,” we risk losing sight of what is distinctive about algorithmic explanation: it isolates the stepwise procedure that any adequate implementation must, in some sense, realise.

Mechanistic interpretability, as F&M emphasise, lets us uncover how LMs process information internally. They highlight work that identifies “circuits,” or subgraphs of neural networks, that appear to be causally responsible for computing particular syntactic features such as number agreement. Describing how the components of these circuits interact, while omitting many lower-level implementation details, yields what we might call mechanism sketches for linguistic behaviour (Piccinini & Craver, Reference Piccinini and Craver2011). Mechanism sketches of this kind are compressive and predictive. But algorithms are not mechanism sketches: they are formal, abstract, and multiply realisable procedures that stand in an implementation relation to particular circuits (Williams, Reference Williams2025). To describe what a system does at the algorithmic level of analysis is to specify a particular sequence of representational states and computations through which it transforms inputs into outputs.

We can formalise algorithmic-level claims about LMs within the framework of causal abstraction (Geiger et al., Reference Geiger, Ibeling, Zur, Chaudhary, Chauhan, Huang, Arora, Wu, Goodman, Potts and Icard2025). An LM is represented as a low-level causal model, while an algorithm hypothesised to compute a given syntactic feature (e.g., subject–verb agreement) is represented as a high-level causal system whose nodes are linguistic variables (subject number, attractor number, verb form, etc.). We then ask: does there exist an alignment between them such that interventions on the LM’s internal states have the same behavioural outcome as interventions on the variables of the algorithm? Concretely, this can be probed by interchange interventions. We feed the LM a base input and a counterfactual input that differ with respect to the target feature, then “patch” a candidate subspace of the LM’s hidden activations in the base run with those from the counterfactual run. If that subspace really encodes the target feature (e.g., subject number), then patching it should change the model’s output (e.g., predicted verb form) exactly as the algorithm predicts, across many sentences and many intervention sites. When a simple alignment of this kind achieves high interchange intervention accuracy, we have evidence that the LM implements the candidate algorithm.

Focusing on algorithms sharpens the relevance of LMs to linguistic theory, which offers hypotheses about algorithms that construct and constrain structured linguistic representations. Psycholinguistics proposes algorithms for how such representations are used in parsing, prediction, and retrieval. Interpretability under causal abstraction lets us ask whether an LM has in fact converged on those algorithms, or on some different procedure that nonetheless supports similar surface patterns. Either answer is informative: convergence would vindicate the algorithmic hypothesis as a robust solution to a hard problem, while divergence might lead us to reconsider what goes on in the human case. In addition, algorithmic-level explanations for the linguistic behaviour of LMs allow us to carve out a meaningful double dissociation between competence and performance in humans and machines alike (Millière & Rathkopf, Reference Millière and Rathkopf2026).

One might worry that talk of implementation is cheap: with a sufficiently baroque mapping between states, any network can be said to implement any algorithm. Causal abstraction can be supplemented to address this triviality concern in two ways (Geiger, Harding, & Icard, Reference Geiger, Harding and Icard2025). First, we can restrict the admissible mappings, for example, to reasonably simple linear transformations of activation space plus coarse-graining. Second, and more importantly, we can build in a demand for generalisation. A good candidate algorithm must predict the model’s behaviour not only on inputs that resemble its training data but also under novel interventions and out-of-distribution cases.

In sum, F&M are right that linguistic structure should be treated as a real pattern, independent of its innateness, and that LMs dramatically expand our ability to study such patterns in silico. My proposal is that, by explicitly foregrounding algorithms as a distinguished class of real patterns – understood as interventionally robust causal models under causal abstraction – we can further consolidate the bridge between LMs and theoretical linguistics. LMs let us test, in unprecedented detail, which algorithms statistical learners converge on to solve linguistic tasks.