Source-linked AI summary
Co-Linguistics: AI-augmented Theory Construction in Linguistics
Emmanuel Chemla, Benjamin Spector, Alexandros Kalomoiros, Philippe Schlenker
TL;DR
Linguistic theories are mathematically formalizable, but the key task is often finding principles that yield empirical consequences, not proving derivations. The paper proposes AI-assisted construction and assessment of theories; it concludes that AI could accelerate theory development while human expertise remains central.
Problem
Although linguistic theories can be formalized mathematically, the key challenge is often finding axioms that derive desirable empirical statements rather than proving the derivations.
Source: §1.1 · pp. 1–2Method
The paper proposes using AI to make linguistic theories explicit, compare them, propose new ones, and identify and test crucial predictions.
Source: §1.1 · pp. 1–2Results
AI could improve the depth, breadth, and speed of linguistic theory construction.
Source: §5 · p. 6Takeaways & Limitations
Linguists remain responsible for providing scientific direction and evaluating theories conceptually.
Source: §1.1 · p. 2Human linguistic judgments and experiments may be needed to assess empirical predictions outside LLMs' reach.
Source: §1.1 · p. 2, §3.2 · p. 6
Abstract
from arXiv · showhide
LLMs have been studied in recent linguistics as potential models of humans' linguistic abilities. Here we discuss an entirely different use of AI, namely as a co-scientist, to help construct and assess linguistic theories (we refer to the result as "Co-Linguistics"). Since the 1960s, linguistics has developed theories that are in principle mathematically formalizable, often in the language of formal language theory or model theory. The AI revolution in mathematics will thus have consequences in linguistics-but with an essential twist: proving new theorems is rarely the linguist's goal. Rather, one seeks to find the best set of axioms to derive empirical statements. AI could accelerate research by making existing theories fully explicit, by comparing competing theories, and more ambitiously, by proposing new theories (in machine learning, this relates to "program induction"). It will also help assess theories by accelerating the identification and test of crucial predictions, thanks to unparalleled access to data (in machine learning, this relates to "active learning"). While the cycle from theory evaluation to theory construction may give rise to recursive and possibly autonomous improvement of linguistic theories, humans remain central: linguists provide scientific directions and evaluate theories conceptually, and experimental participants are needed to assess empirical predictions that are outside the reach of LLMs.
1. Introduction
The paper proposes AI as a co-scientist for constructing and assessing formal linguistic theories, rather than as a model of human linguistic ability. It argues that AI may help build a recursive theory-improvement process, while linguists remain responsible for scientific direction and judgment.
1.1 Goals
- Because linguistic theories seek empirical consequences from formal principles, the central challenge is often finding suitable axioms rather than proving their derivations.The paper relates this theory-construction problem to program induction.Source: §1.1 · pp. 1–2
- AI could support linguistic research by making theories explicit, comparing and proposing theories, and identifying and testing crucial predictions.The authors illustrate their proposal with examples from semantics while hoping it applies across linguistic subfields.Source: §1.1 · pp. 1–2
- AI-assisted theory evaluation could feed new data into theory construction, but linguists must provide direction and judge theories, and human judgments or experiments may be needed beyond LLMs' reach.The authors describe this as a human-in-the-loop process.Source: §1.1 · p. 2
1.2 Precedents and parallels
- The proposal builds on proof-assistant precedents, including a model coupling untrusted theory generators with trusted proof-assistants for recursive scientific formalization.The paper identifies Chojecki (2026) as proposing this approach across fields including biology, chemistry, and economics.Source: §1.2 · p. 2
- Bayesian Program Induction searches a predefined program space for simplicity and predictive accuracy, while recent AI-driven recipes leave theory format and search more open to AI heuristics and possible human intervention.The paper contrasts these recipes with Ellis et al. (2022), whose predefined rules and search procedure model child discovery; the authors call them a non-starter for that purpose.Source: §1.2 · p. 3
2. AI-assisted Theory Development
Formal linguistics develops theories that can be expressed mathematically, and AI may help make them explicit, compare them, and construct alternatives. The section illustrates these roles with work on semantics, while noting that linguists must verify formalizations against the theories they intend to represent.
2. AI-assisted Theory Development
- AI can formalize linguistic theories, compare their predictions, and construct new theories by modifying existing principles or exploring alternatives.Formal linguistics has sought mathematically explicit accounts of linguistic behavior.Source: §2 · p. 3
2.1 Formalizing existing theories
- Claude verified two equivalent geometric restatements in Lean and re-proved quantifier predictions while deriving further facts, though the Appendix formalizations remain provisional.Formalizing theories can clarify claims, uncover gaps, and help check formal results efficiently.Source: §2.1 · pp. 3–4
2.2 Comparing existing theories
- Claude checked and corrected hand-proven equivalences among five presupposition-projection theories, proposed additional results, and linked Limited Symmetry to earlier systems.The related analyses were Schlenker’s Transparency and Local Contexts systems.Source: §2.2 · p. 4
2.3 Building new theories
- Codex revised Spector’s anaphora theory to handle disjunctive cataphora while preserving positive predictions, with no greater complexity and better empirical coverage.The original theory’s left-to-right licensing handled conjunction but not disjunction.Source: §2.3 · p. 4
3. AI-assisted Theory Evaluation
The section evaluates theories through explanatory criteria and empirical adequacy, asking how theories compare and what data can test their predictions. AI can help assess parsimony, generality, and evidence, but human judgment remains necessary.
3. AI-assisted Theory Evaluation
- Theory evaluation asks about explanatory power and elegance, coverage of published data, and whether new data can adjudicate among competing theories.The section frames these as three forms of evaluation.Source: §3 · p. 5
3.1 How explanatory is the theory?
- Minimum Description Length operationalizes parsimony through theory length, while generality favors principles applicable to more data; specialists remain the ultimate judges of explanatory value.Code length is only an initial approximation and depends on the implementation language; LLMs may serve as an imperfect proxy for specialist judgment.Source: §3.1 · p. 5
3.2 How empirically adequate is the theory?
- AI can gather published linguistic results and corpus data, but crucial predictions may require comparative acceptability, inferential, truth-conditional, developmental, or processing evidence.Such evidence can motivate detailed introspection or psycholinguistic experiments.Source: §3.2 · p. 5
- LLM behavior and LLM predictions offer fast but error-prone shortcuts; tracking their area-specific reliability may provide some confidence, while human judgments remain necessary to adjudicate theories.Critical human judgments can concern marginal examples outside LLM training data; animal communication often requires new experiments, whereas LLM behavior can be studied with automatically collected data.Source: §3.2 · p. 6
4. Recursive Improvement of Theories
AI contributions can be combined into a recursive process in which new data feed theory construction. The authors expect AI to accelerate theoretical progress, while human experts and participants remain central for human and animal research.
4. Recursive Improvement of Theories
- AI may construct theories, check derivations, compare parsimony, and find adjudicating data, with new data feeding back into theory construction.The authors describe this as a potentially recursive improvement cycle.Source: §4 · p. 6
- Human experts must judge theories and guide iterations, and predictions about human or animal abilities still require data from people or animals.For LLM behavior, the empirical part of recursive theory improvement may be fully automated.Source: §4 · p. 6
- The authors expect AI to significantly accelerate theoretical progress.Source: §4 · p. 6
5. Conclusion
The conclusion presents AI as a potential tool for formalizing, verifying, constructing, and assessing linguistic theories, with possible recursive improvement. It cautions that this development does not necessarily benefit linguists or linguistics as a field.
5. Conclusion
- AI could formalize, verify, and construct linguistic theories, assess parsimony and empirical coverage, and support recursive theory improvement when all steps are present.The authors see potential to improve the depth, breadth, and speed of theory construction.Source: §5 · p. 6
- The authors caution that AI-driven development does not necessarily benefit linguists as individuals or linguistics as a field.Source: §5 · p. 7
Appendix
The authors’ ongoing case studies are available online.
Appendix
- The authors direct readers to an online link for their ongoing case studies.Source: Appendix · p. 7