OPEN HONEST FOUNDATION · RESEARCH

Language and thought

A large language model is a telescope onto language. Language is a telescope onto the intelligence that roughly 100 billion people deposited into it over deep time.

The Foundation's second line of research treats a language model as an instrument of observation rather than an instrument of generation. What a model trained on one language can and cannot do is a measurement of what that language supplies, and the measurements so far point to intelligence embedded in language.

The argument that language is intelligent

Intelligence, functionally, is intentionality coupled to agency. Intentionality means a model of the world, an aboutness. Agency means the capacity to direct action. Agent-based, predictive-processing, and evolutionary approaches in cognitive science all treat an intelligent system as a world model joined to goal-directed action. No definition of intelligence is accepted across the whole field, this is the one the argument depends on.

Language is about the world, so it has intentionality. A sign that reads DANGER. STOP. has the capacity to stop the walker who reads it. The walker supplies the muscles and the nervous system; the sign supplies the direction. Under the definition, that is agency, and language is intelligent.

The same holds for a large language model. Its weights are frozen and it keeps no memory between uses, yet it still completes the loop from representation to action in whoever or whatever acts on its output. The claim is about functional intelligence, not consciousness.

Phenomenology, theology, philosophy of mind, and law use intentionality and agency in other senses, and the position that only biological minds have intentionality is a neighboring instrument with its own limits rather than a refutation of this one.

Where the cognition is

The 4E reading in cognitive science describes cognition as embodied, embedded, enacted, and above all extended. The Foundation treats that reading as a working hypothesis rather than a settled position. Some measurements point toward it and do not force it.

In the Foundation's experiments, structural alignment tracked acquired grammatical competence rather than architecture or training duration. A French model's word space was mapped onto two English word spaces. Against GPT-2, which is grammatically competent, held-out word pairs matched 68.84 percent of the time, where chance is 3.80. Against an English model of the same architecture as the French one, trained on 71 times more text and never above chance on English grammar probes, the match fell to 25 percent. The two competent models share a structure that is generic to language rather than particular to English or French, and independent of the architecture that acquired it.

What a model might be displaying, on this reading, is the cognition deposited into language by everyone who has ever used it. The model is a telescope onto language. Language is a telescope onto that distributed intelligence. Nothing measured so far rules out a different account of why the corpora differ.

What the measurements show

How grammatical competence emerges

Many in the industry assume that competence in a language model comes only from scale. These studies change one factor at a time and hold the rest constant. Some train small models from scratch and vary only the language of the corpus. Others swap the tokenizer under a fixed model, run placebo prompts, or compare existing models under a single change to the prompt. The cross-linguistic ablations are registered on the Open Science Framework under block SJ48B, dated December 6, 2025, with separate deposits for the interleaving and Rust experiments.

Three levels, kept apart

The Foundation keeps three levels of claim apart, and is clear about which level any particular claim belongs to. A measurement is never offered as proof of a philosophy, and a philosophical position is never presented as an empirical finding.

The program needs all three levels, because none of them can answer the questions that belong to the others. The Foundation measures first, and it argues from what the measurements have shown.

Discount bias

Anthropomorphic projection is the error of seeing a mind where there is none. Discount bias is its counterpart: the dismissal of intelligence that does not resemble human conscious experience. The first error is named and guarded against everywhere. The second is not.

Cognition located in external and collective systems is what discount bias rules out before any evidence is examined. Naming the bias is what makes the 4E reading available as a hypothesis rather than a category error, and the Foundation treats that as part of the scholarly contribution.

A topology of meaning

The cross-linguistic measurements suggest that human language has a discoverable structure underneath its surface forms. Several traditions have proposed something of that kind and named it differently: Plato's forms, the Logos of the Johannine prologue, the Tao that cannot be spoken, the two truths of Madhyamaka, and the apophatic claim about what remains when every predicate fails.

Those are not evidence, and the Foundation does not present them as evidence. They are prior attempts on the same object by instruments that share no assumptions with a controlled experiment. The method below is about what agreement between instruments that far apart can support.

The convergence-of-signals method

The method reads agreement among independent instruments, at the boundaries where each one stops, as licensed inference. Instruments can be empirical experiments, philosophical inquiries, or scholarly and theological traditions. Agreement between instruments that share no assumptions supports a conclusion that none of them supports alone.

The language-contingency result is a worked case. One instrument is a controlled training comparison across typologically different languages. Another is the typological record that field linguists compiled over decades, which ranks languages by how redundantly they mark grammatical relations. Neither was built to check the other. How efficiently a language trains a model tracks where that language falls in the record.

The program cannot prove a universal, so it advances by ruling out. Each failed probe and each named limit adds an edge to what the program can claim. Formalizing the method into a peer-reviewable specification is one of the Foundation's stated contributions, and a draft with worked examples is in preparation. The Epistemology page describes the discipline behind it.

MÉTRON is a proposed family of measuring instruments for cross-linguistic research. One part of it exists. MÉTRON-FR is a trained French checkpoint, released with the BabyLM submission and used for the results above. Everything else on this page is a design.

The design is for a platform that would expose a small set of levers rather than the machine learning beneath them. A researcher could remove one feature of a language to find what it causally shapes. A researcher could align two meaning spaces to test whether a concept is shared between them or bound to one. A researcher could read the structural signature of a body of text, and could run any of this on a small corpus such as a single author, period, or tradition.

A research question would become an experiment by being fitted to one or more of those levers. The scholars it would be built for are philosophers of language, theologians, computational linguists, cognitive scientists, and people working across traditions, none of whom should have to become machine-learning engineers first.

The design also includes a channel where native speakers of any language could build grammar test sets for their own language, receiving compute in exchange for releasing those test sets openly.

Take part

The Foundation also hosts cross-traditional convenings on this work, which are described on the Epistemology page. Get in touch, or read the correctness research that shares the Foundation's pipelines and review.