The paper asks whether AI can serve as a co-scientist for linguistic theory, rather than being used mainly as a model of human language ability. Its proposal is to use AI to formalize existing analyses, compare competing theories, help construct revisions and assess their strengths. The proposal groups those activities into one approach to theory development.
This is a conceptual proposal illustrated by the researchers' own experience in semantics. The evidence base is a series of theoretical case studies, not a reported sample of human or animal participants. It is therefore an account of possible workflows, not a participant study.
What the AI was asked to do
One case concerns linguistic classifiers. Claude offered two equivalent restatements of an analysis in cleaner, simpler geometric terms and offered Lean 4 verification of the equivalences. Lean 4 is a proof assistant, so it can check formal proof steps. That check still depends on an earlier human decision: whether the formalized principles actually capture the intended linguistic theory.
Another example looks at presupposition projection, the way background assumptions behave in sentences involving quantifiers. Claude re-proved sample predictions and derived further facts for additional quantifiers. In a separate comparison, it checked equivalence results among five presupposition-projection theories, corrected the results, proposed new results and offered general theorems linking Limited Symmetry with earlier analyses.
The hard part is deciding what counts as better
The paper then describes a Codex-assisted revision of an anaphora theory, the way expressions refer to one another. The authors say the revision addresses cataphora, a case in which an expression is understood through material that comes later, appears no more complex than Spector's original theory and offers better empirical coverage. The case studies are ongoing and still in flux, so the account remains provisional.
The paper proposes looking at theory length, including Minimum Description Length, or MDL, as well as generality and possible judgments from language models. MDL is presented as a way to ask how much description a theory requires. Human specialists, however, remain the ultimate judges of explanatory value.
The authors are more cautious about using large language model, or LLM, proxies as stand-ins for people. They describe proxies for human behavior or psycholinguistic results as a fast but error-prone shortcut, and say their reliability should be tracked. For questions about humans and animals, the paper says direct data collection remains necessary.
A loop that remains untested
The broader idea is a recursive workflow, with each round feeding the next. AI would help construct a theory, check theorems, assess parsimony, search for new data and support another revision, while a human remains in the loop. For questions about LLM behavior, the paper says the empirical part of recursive theory improvement may be fully automated; for human and animal targets, direct data collection would still be needed.
Taken together, these examples lead the authors to argue that AI has substantial potential to increase the depth, breadth and speed of linguistic theory construction. The conclusion is prospective, however: the case studies illustrate what might be possible, rather than measuring a gain in scientific productivity.
The limits are clearest in the verification step. Lean can check proof steps, but linguists must still determine whether the principles were formalized correctly, and the Appendix results are provisional and require further checking against the original theories. The proposed system is therefore a human-in-the-loop research program, not a replacement for human conceptual judgment or direct human and animal data.
Paper data and sources
Original title: Co-Linguistics: AI-augmented Theory Construction in Linguistics
Authors: Emmanuel Chemla, Benjamin Spector, Alexandros Kalomoiros, Philippe Schlenker
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-09-29
DOI: Not available
Original paper · Full text